Machine Learning for Healthcare Wearable Devices: The Big Picture

Farida Sabry; Tamer Eltaras; Wadha Labda; Khawla Alzoubi; Qutaibah Malluhi

doi:10.1155/2022/4653923

Machine Learning for Healthcare Wearable Devices: The Big Picture

Sabry, Farida;Eltaras, Tamer;Labda, Wadha;Alzoubi, Khawla;Malluhi, Qutaibah 2022-04-18 00:00:00 Hindawi Journal of Healthcare Engineering Volume 2022, Article ID 4653923, 25 pages https://doi.org/10.1155/2022/4653923 Review Article Machine Learning for Healthcare Wearable Devices: The Big Picture 1 1 1 2 Farida Sabry , Tamer Eltaras , Wadha Labda , Khawla Alzoubi , and Qutaibah Malluhi Computer Science and Engineering Department, Faculty of Engineering, Qatar University, Doha, Qatar Engineering Technology Department, Community College of Qatar, Doha, Qatar Correspondence should be addressed to Farida Sabry; faridasabry@qu.edu.qa Received 17 November 2021; Accepted 22 March 2022; Published 18 April 2022 Academic Editor: Yuxiang Wu Copyright © 2022 Farida Sabry et al. *is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Using artiﬁcial intelligence and machine learning techniques in healthcare applications has been actively researched over the last few years. It holds promising opportunities as it is used to track human activities and vital signs using wearable devices and assist in diseases’ diagnosis, and it can play a great role in elderly care and patient’s health monitoring and diagnostics. With the great technological advances in medical sensors and miniaturization of electronic chips in the recent ﬁve years, more applications are being researched and developed for wearable devices. Despite the remarkable growth of using smart watches and other wearable devices, a few of these massive research eﬀorts for machine learning applications have found their way to market. In this study, a review of the diﬀerent areas of the recent machine learning research for healthcare wearable devices is presented. Diﬀerent challenges facing machine learning applications on wearable devices are discussed. Potential solutions from the literature are presented, and areas open for improvement and further research are highlighted. features can be extracted for machine learning (ML) algo- 1. Introduction rithms to detect and learn useful patterns. *is can be very *e last few years have witnessed great advances in wearable useful in many healthcare and elderly care applications such technologies. Wearable devices include any device that can as activity detection for health state assessment, fall detec- be worn by humans such as wristwatches, glasses, chest tion, stress detection, ﬁtness tracking, vital signs monitoring, straps, rings, and prosthetic sockets. Wearable devices be- and diseases’ diagnosis. Using machine learning techniques long to the Internet of medical things (IoMT), together with to learn from human body signals, recorded by wearable the implantable, ambient, and stationary devices used in devices, has been an active research area in the last decade hospitals. *ese devices are typically connected to a network with a lot of published research studies. Despite this huge research eﬀort and the remarkable growth in using wearable and communicate remotely with mobile devices as shown in Figure 1. devices, especially smart watches, few machine learning Wearable devices may include diﬀerent types of sensors applications for wearable devices have found their way into to continuously monitor various human signals, e.g., tem- the market. perature sensors, accelerometers, optical sensors, and bio- Examples include irregular rhythm notiﬁcation feature metric sensors. Although the readings of some of these [3] in Apple Watch, which won U.S. Food and Drug Ad- sensors are not yet as accurate as stationary devices in ministration (FDA) approval with a long list of warnings and hospitals, they are sometimes considered acceptable [1, 2], precautions in 2018 (https://www.accessdata.fda.gov/ depending on the application. cdrh_docs/reviews/DEN180044.pdf), and Eko’s heart murmur detection algorithm, which has been recently Sensors in IoMT devices and human interaction with these devices are considered a big source of data from which published [4], which is not really for a personal wearable 2 Journal of Healthcare Engineering Smart Wearable devices phone Watch Helmet Glasses Middleware server Cloud-based infrastructure Chest strap On-device Edge/fog Cloud computing computing computing Figure 1: Wearable device application model. device but an electronic stethoscope. Additionally, some of the (2) What are the challenges facing machine learning for wearable devices that were used for monitoring, which were healthcare wearable devices? surveyed in [5], are no longer available in the market. Practical (3) What are the possible solutions for these challenges and reliable use of machine learning techniques in the domain in literature? of wearable devices is still facing many challenges. *us, the main focus of this study was to identify the Several review papers have discussed some challenges for challenges of developing machine learning applications for wearable devices. In a survey paper published in 2012 [6], the healthcare wearable devices and alternative solutions found authors focused on some features of wearable devices and in the literature. Diﬀerent categories of recent healthcare their types such as diseases that can be monitored, research machine learning research are reviewed while spotting the prototypes, and challenges such as system eﬃciency, user challenges and highlighting potential research areas and perception, cost, social inclusion, and ethical issues. In [5], applications that need further investigation. the authors provided a survey of commercially available *e rest of the review is organized as follows. In the next wearable devices at that time (2017). *ey focused on com- section, the necessary background for IoMTand the diﬀerent munication security issues, power eﬃciency, and wearable human body signals used in wearable devices research are computing. Neither of the surveys focused on the chal- presented. Moreover, applications for machine learning in lenges facing machine learning applications for healthcare IoMT are reviewed and categorized referencing some of the wearable devices speciﬁcally. recent research work published in each area. In Section 3, In this study, we review recent applied machine learning diﬀerent challenges facing machine learning research for research for wearable devices. We identify many challenges wearable devices are reviewed, relevant privacy and security facing machine learning applications on wearable devices aspects for machine learning applications in IoMT are from design to deployment, such as diﬀerent deployment discussed, and possible solutions in literature are presented. alternatives, storage, power consumption, user acceptance, In Section 4, we discuss these solutions, their applicability, reliability, communication, security, and privacy. We discuss and their shortcomings. Additionally, we highlight the main security and privacy both from the data and the model research gaps we perceived in the domain. Finally, the study perspectives listing potential solutions to keep subjects’ conclusions are provided in Section 5. personal data from wearable devices private and secure. Additionally, we review the diﬀerent privacy-preserving techniques used for machine learning training and inference 2. Wearable Devices and Machine Learning and discuss their applicability to the model of wearable *e wearable device domain is being actively researched for device usage shown in Figure 1. the sake of enhancing ease of use, comfort, and non- *e review includes the recent research papers in the invasiveness of monitoring physiological vital signs and ﬁeld of wearable devices that have been published from 2017 sometimes psychological or emotional state, which can be to December 2021 to answer the following questions: detected by analyzing data from diﬀerent sensors. Following (1) What are the healthcare machine learning tasks that the tremendous technological advances in the design of have been researched in the literature, the body system on chip (SoC), the development and use of wearable signals, and techniques used in these tasks? devices have remarkably achieved high growth rates in the Journal of Healthcare Engineering 3 last few years. *e wearable device market size was valued at are often used in a wide variety of applications to capture or USD 32.63 billion in 2019 and is expected to expand at high recognize body movement and activities, which can tell a lot rates in the next few years according to the wearable about the health and the lifestyle of the person [1, 25–39]. technology market industry report by Grand View Research Other signals that have been used sparingly in literature (https://www.grandviewresearch.com/industry-analysis/ include electrogastrogram (EGG), which records the elec- wearable-technology-market). trical activity of the stomach [40], and electrooculogram *e number of globally connected wearable devices is (EOG), which is generated by eye movements and can be about to reach 1 billion according to Statista (https://www. measured with electrodes placed around the eye [41]. statista.com/statistics/487291/global-connected-wearable- Sensory and especially olfactory inputs are challenging to devices/). Examples of wearable devices are smart watches, model, but it was observed that the human body has diﬀerent armbands, chest straps, shoes, helmets, glasses, lenses, rings, autonomic responses to diﬀerent odors, which can be an- patches, textiles, and hearing aids [6]. alyzed through the GSR and ECG signals [42]. *ese inputs Despite this grand growth, there is still a great need for can be used in applications including personalized treat- ongoing research in this area for enhancing the accuracy of ments based on odors and foods for neuropsychiatric and these devices, using diﬀerent body signals for new appli- eating disorders. Some other inputs may need visual cation areas, and dealing eﬀectively with the complexity of monitoring. Figure 3 summarizes the diﬀerent sensors used the human body. in the literature for diﬀerent machine learning healthcare tasks. Features are extracted from these signals to learn a model for either classiﬁcation or regression of a certain 2.1. Wearable Device’s Signals Used in Learning. *e human variable. Some literature studies use statistical values such as body can be seen in an abstract way as a group of systems mean, minimum, maximum, mode, variance, standard de- (circulatory system, nervous system, respiratory system, di- viation, entropy, and kurtosis. However, it is often hard to gestive system, etc.). It receives a group of inputs and releases a interpret how some of these statistical features aﬀect the set of outputs as shown in Figure 2. Inputs include the inhaled classiﬁcation or the outcome variable. Additionally, model’s air, water, food, visual input for all the scenes and objects seen accuracy is usually negatively aﬀected by adding more ir- during the day, auditory input for all sounds and voices heard, relevant features as more is not always better, and domain- sensory inputs for the things touched, and olfactory input for speciﬁc features that are expressive achieve better perfor- things smelled during the day. Outputs include the exhaled air, mance [43]. Estimating heart rate and breathing rate as excretions such as urine, feces, and sweat, skin moisture, body features from the PPG signal, change in acceleration mag- temperature, blood in case of injuries and laboratory tests, nitude, jerk of motion, and transient changes in skin re- energy released by the human in terms of body movements, or sistance for seizure detection are examples of domain- performing mental activities and voice output, which can be speciﬁc features. Some applications are concerned with normal speech, singing, or shouting. Analysis of the inputs and changes happening over a long time period, and some are outputscan, to some extent,predictthehealthstate of aperson, concerned with transient changes due to certain events such diagnose possible diseases/disorders, and assist with thera- as fall detection and emotion recognition. peutic suggestions. *ese inputs and outputs need to be monitored by wearable devices worn during the day. Wearable devices include any device mounted on the 2.2. Machine Learning for Wearable Devices. Machine body and can capture noninvasive signals from the human learning involves getting wearable devices to act/take de- body through the use of diﬀerent types of sensors. *ere are cisions without explicit programming for a speciﬁc scenario numerous well-known signals and signs that are read from through learning from past experiences. As it is well known, the human body in literature to identify the vital signs and machine learning is usually classiﬁed as either supervised, other information about the health or mental state of the unsupervised, semi-supervised, or reinforced according to subject. Examples of these sensors include skin temperature the type of the available training data. Learning from past sensor used in [7, 8] and electrodermal activity (EDA) sensor experiences is encoded in terms of data examples that can be or sometimes known as galvanic skin response (GSR) sensor either labeled or unlabeled. *e target variable for labeled used on the skin to record the skin conductance that varies data can be categorical or numerical. Among the tasks that with the sympathetic state of the subject [2]. Other examples involve machine learning are classiﬁcation in case of the include an electrocardiogram (ECG) sensor to capture categorical target output variable, regression in case of electrical changes in the skin corresponding to heartbeats numerical labels, and clustering for unlabeled data. Most used in [9–12]. To capture features of the electrical activity of machine learning research for wearable devices belongs to the brain and the health of muscles and the nerve cells, the classiﬁcation tasks, some are for clustering [44–46], and electroencephalogram (EEG) and electromyography (EMG) few can be tackled as regression problems [43]. sensors are used [13–19]. Blood volume pulse (BVP) can be Applied research to explore applying machine learning captured using an optical photoplethysmography (PPG) techniques using the body signals discussed in the last sensor to estimate heart rate and heart rate variation as in subsection for health monitoring, elderly care, and ﬁtness tracking has been growing over the last decade. Among the [1, 20–22]. PPG sensor [23] is also used to give an ap- proximation for the oxygen saturation in blood (SpO ) as in areas that got researchers’ attention are fall detection, seizure detection, vital sign monitoring and prediction [47], and [22, 24]. Accelerometers, gyroscopes, and magnetometers 4 Journal of Healthcare Engineering EEG EOG Skin Temperature Microphone Exhaled air Urine ECG Feces Sweat Air Skin moisture Water Temperatre Food tears Visual input Energy Auditory input (Body movement Sensory input + mental activities) Olfactory input Voice output Blood Accelerometer EDA signals EMG BVP Figure 2: Human body as a system and signals that can be used as a source of data for machine learning models. Eating Monitoring Fall Detection Activity Recognition Fitness Tracking accelerometer accelerometer gyroscope accelerometer accelerometer gyroscope magnetometer gyroscope gyroscope magnetometer microphone magnetometer magnetometer proximity sensor Seizure Detection Hydration Monitoring Arrhythmia Detection Sleep Monitoring EEG EDA, PPG, accelerometer ECG accelerometer accelerometer gyroscope, magnetometer PPG EDA gyroscope temperature acoustic sensor Temperature acoustic sensor PPG Stress Detection Emotion Recognition Disease Diagnosis Rehabilitation Tasks ECG EEG ECG EMG EDA EDA PPG inertial sensor Temperature PPG Temperature (accelerometer & PPG Temperature accelerometer gyroscope) Figure 3: Healthcare machine learning tasks and sensors used for each one in literature. activity recognition for ﬁtness tracking or identifying human machine learning technique(s), sensor(s), and dataset(s) daily activities. Additionally, wearable devices have been used in each study. researched for their use in stress detection and detection of heart rate arrhythmia and rehabilitation tasks. Tables 1–3 show the diﬀerent areas and a sample of the most recent 2.2.1. Fall Detection. *ree categories of fall detection re- research work done in each area. *e table also shows the search eﬀorts can be identiﬁed in the literature based on the Journal of Healthcare Engineering 5 Table 1: Machine learning research work for healthcare wearables for fall detection, activity recognition, eating monitoring, ﬁtness tracking, and stress detection. Research Task ML technique(s) Sensors/signals used Dataset(s) work J48 (96.7%), logistic regression 3D accelerometer and MobiAct (https://bmi.hmu.gr/the-mobifall- [48] (94.9%), MLP (98.2%) gyroscope in smartphone and-mobiact-datasets-2/) KNN (84.1), naive Bayes UMAFall dataset (https://ﬁgshare.com/articles/ Accelerometer, gyroscope, [49] (61.5%), SVM (68.25%), and dataset/UMA_ADL_FALL_Dataset_zip/ and magnetometer ANN (72%) 4214283) Temporal signal angle 12 features for 7 subjects performing 5 fall types Fall measurements Inertial measurement unit [30] detection (93.3%@200 Hz to 91.8% (IMU) (9 times each with 3 diﬀerent speeds) @10 Hz) KNN and RF SisFall dataset [51] Accelerometer and [50] (99.80% KNN and 96.82% for gyroscope (For falling and non-falling activities) falling activity recognition) SVM (97% F1 score and 99.7% Accelerometer and [52] Public fall detection dataset [27] recall) gyroscope CNN UCI-HAR dataset and study set Accelerometer and [25] (UCI-HAR dataset: 95.99%, gyroscope 21 participants and 6 ADLs study set: 93.77%) Locally linear embedding Accelerometer, [53] UCI-HAR dataset transfer learning magnetometer, gyroscope Tri-axis accelerometer, tri- Activity Sequence-to-sequence axis gyroscopes, Postures dataset, mini MobiAct, and UCI-HAR [26] recognition matching network magnetometer (depending dataset on the dataset) sEMG signals of the upper 6 males and 6 females for 3 motion states of [54] SVM: 90% limb by Delsys, virtual vehicle: left turn, stop, and right turn accelerometer Tri-axis accelerometer, tri- 6550 pieces of data for 4 activities: walking, [39] ATRCNN: 97% axis gyroscope sitting down, running, and climbing stairs A public dataset for performing diﬀerent [34] Proximity-based active learning 3D accelerometer activities including eating [34] One IMU and a proximity Two datasets: 12.5 hrs for 16 participants in Random forest (89.6% in the sensor on ear and one IMU semi-controlled setting with 6 labels and 3 hrs [55] laboratory and 72.2% outside on the upper back and a for each of 15 participants outside the laboratory the laboratory) microphone with chewing and non-chewing labels A public dataset for performing diﬀerent Eating [37] DBSCAN clustering 3D accelerometer activities including eating [34] monitoring A study dataset of 25 participants, 10 in a Random forest and DBSCAN Inertial sensor on the laboratory setting and 15 in the wild doing [56] clustering algorithm (average downside of the lower jaw diﬀerent activities including eating a meal of precision of 92.3%) diﬀerent food types Gyroscope and Gradient boosted decision tree [33] accelerometer in Apple 79 features for 16 subjects taking pills (80.27% accuracy) Watch Logistic regression (0.9356), random forest (0.9203), Study set of 39 participants with a total of extremely randomized trees 2 accelerometers (hip and [38] 55 days in which sport and jogging activities Fitness (ERT) (0.9177), and SVM ankle) were logged tracking (0.9328)—best accuracy reported in diﬀerent scenarios 3-Axis accelerometer and [57] L2-SVM 114 participants over 146 sessions 3-axis gyroscope Zephyr BioHarness for BN, SVM, KNN, J48, 2 participants with 324 instances [2] ECG RF and AB learning methods Shimmer3 GSR for EDA At rest and exercise sessions Neural network model (92% ECG, GSR, body Stress accuracy for metabolic [24] temperature, SpO2, glucose 312 biosignal records from 30 participants detection syndrome patients and 89% for level, and blood pressure the rest) HR and RR data for 44 children (26 with ASD LR (87% accuracy) and SVM [58] ECG sensor in a chest strap and 18 without ASD) while at rest (7 min) and (93%) while engaged in stressful tasks (9 min) 6 Journal of Healthcare Engineering Table 2: Machine learning research work for healthcare wearables for arrhythmia detection, seizure detection, rehabilitation tasks, and hydration monitoring. Research Task Techniques Sensors Dataset(s) work 14 subjects recordings for a 30-minute SVM and K-medoids clustering- [59] ECG and PPG sensors training session and a 30-minute testing based template learning session ECG sensor, PPG sensor [60] Deep learning (max 89% accuracy) Cleveland database on UCI (SpO2) ECG patch (from 91,232 single-lead ECGs from 53,549 [9] DNN (0.837 F1 score) iRhythm) patients 402 PPG recordings for 29 free-moving 50-Layer convolutional network subjects (13 with persistent AF) and the [61] PPG sensor (95% AUC) NSR dataset of 341 PPG recordings from Arrhythmia 53 healthy free-moving subjects detection PPG sensor in a ring-type 13,038 30-s PPG samples (5850 for SR [10] Deep learning (94.7%) device and 7188 for AF) Public available dataset from Computing in Cardiology Challenge (CinC) 2017 [11] SVM and bagging trees ECG (https://physionet.org/content/ challenge-2017/1.0.0/) 5878 deidentiﬁed audio recordings, totaling >rbin 34 hours from 5318 ResNet of 34 layers of 1D rectiﬁed [4] Acoustic recordings unique patients labeled by a majority linear unit vote of 3 cardiologists as heart murmur, no heart murmur, or inadequate signal SVM (97.31), RF (97.08), NB (95.08), UCI EEG sampled dataset for epileptic [62] K-nearest neighbor (90.01), and EEG seizures neural network (93.53) Accelerometer and SVM ((Sens > 92%) and bearable 135 patients with generalized tonic- [63] electrodermal activity FAR (0.2–1)) clonic seizures with 22 seizures from Empatica Embrace Accelerometer and 40 pediatric patients with generalized Seizure [64] Not mentioned electrodermal activity tonic-clonic seizures detection Two classiﬁers (the models are EDA and accelerometer 5,928 h of data of 55 convulsive [65] not mentioned) best sensitivity 95% from three wristbands Epileptic seizures from 22 patients and< 1 false alarm rate Temperature, 69 patients with epilepsy accelerometer [8] LSTM and 1DConv (total duration > 2311 hours, 452 Blood volume and EDA seizures) sEMG acquisition Muscle signals sEMG for 3 users doing 9 [66] SVM, RF module hand gestures 12 times IMU sensor module and 81654 samples for 10 people data, each K-means clustering, SVM, and [67] plantar pressure sample has 10 features calculated from Rehabilitation artiﬁcial neural network (ANN) measuring foot insoles 64 sensing nodes in the foot insole tasks Inertial features and anthropometric [68] Support vector regression (SVR) IMU in SportSole characteristics of 14 healthy subjects Multiple regression, inference tree, Two-sensor (fore and aft) Kinematic and pressure features for 30 [69] and RF insole (LoadsolTM) participants, each doing 120 steps SVM for drinking detection Acoustic sensor Frequency and cepstral domain [70] Gradient boosting decision tree for and inertial sensor Features are extracted from the signals activity recognition LDA, quadratic discriminant 51 hydrated samples and 17 dehydrated analysis, logistic regression, SVM, [21] EDA and PPG for 17 subjects with features from EDA Hydration Gaussian kernel, KNN, decision and PPG monitoring trees, ensemble of KNN ECG (not wearable (RR SVM (60%) and K-means clustering 10 minutes ECG for 16 athletes at rest, [71] interval, RMSSD, and (42%) post-exercise, and post-hydration SDRR recorded)) Shimmer (IMU, GSR, 3386 min for 11 subjects under fasting [43] DNN, RF, extra trees PPG, etc.) and non-fasting conditions Journal of Healthcare Engineering 7 Table 3: Machine learning research work for healthcare wearables for emotion recognition, sleep monitoring, and disease diagnosis. Research Task Techniques Sensors Dataset(s) work Liquid state machine (LSM)—above [16] 94% accuracy for valence, arousal, and EEG sensor DEAP dataset [72] liking recognition MUSE headband (EEG) and Emotion KNN (accuracy ranges from 53.6% [73] Shimmer GSR + device 54 subjects watching 24 pictures recognition to 69.9%) (SC and HR) Random forest, SVM, and logistic Respiratory belt (RB), PPG, and [74] regression—73.08% for arousal and DEAP dataset [72] ﬁngertip temperature sensor 72.18% for valence Auto-correlated wave detection with an adaptive threshold (ACAT), UCI-HAR dataset and study set of [1] Accelerometer and gyroscope accuracy for UCI-HAR dataset: 21 participants and 6 ADLs Sleep 95.99%, study set: 93.77% monitoring Accelerometer data during one night for 134 participants (70 with [32] Random forest (F1 score: 73.93%) Accelerometer in wristband sleep disorder and 64 good healthy sleepers) (MIMIC III) waveform database for ICU patients and a database of ResNet with LSTM for hypertension ECG, PPG, and invasive BP in patients with cardiac arrhythmias [75] detection ICU collected from Fuwai Hospital, Chinese Academy of Medical Sciences Everion wearable Disease (skin temperature, respiratory 200–1000 asymptomatic subjects diagnosis Machine learning techniques for early [76] rate, blood pressure, pulse rate, with close COVID-19 contact detection of COVID-19 blood oxygen saturation, and under quarantine in Hong Kong daily activities) Heart rate, heart rate variability, 34 patients with PCR-conﬁrmed Multivariate regression for case respiration rate, oxygen COVID-19 were admitted to the [22] deterioration saturation, blood pulse wave, isolation wards of Queen Mary skin temperature, and actigraphy Hospital used technology: (1) wearable devices that use accelerom- techniques is still questionable as the experiments were done in eters and magnetometers, (2) ambient devices such as ﬂoor a controlled environment with a limited number of participants and have the limitation of a high false alarm rate [77]. Another sensors and pressure sensors, and (3) vision-based devices that use monitoring cameras [77]. study to simulate fall data [31] was done to generate forward and syncope accelerometer data to form a larger dataset for fall In [50], the authors reported an accuracy of 99.80% using KNN classiﬁer and 96.82% for falling activity recognition using detection training. the random forest classiﬁer. Using tri-axial accelerometer de- vices [27], KNN, SVM, ANN, and RF classiﬁers were tested to get a mean average accuracy ranging from 48% to 98% 2.2.2. Activity Recognition. Activity recognition enables depending on the classiﬁer’s task. Some tasks involved dis- health professionals to get information about a patient’s tinguishing fall samples from daily activities’ samples. Other ability (or inability) to perform activities of daily living tasks were to distinguish between diﬀerent fall samples or (ADLs) as a measurement of their health status. Human diﬀerent daily activities. *e results showed that the classiﬁ- activity recognition has been researched using convolutional neural networks by the authors in [25], and an accuracy of cation results on raw data are better than depending solely on the magnitude as feature vector. On the contrary, the magnitude approximately 96% and 94% was achieved for the UCI-HAR performs better than raw data in the case of subject-indepen- dataset and their study dataset. However, the accuracy of dent evaluation. It was easier to distinguish between falls and no machine learning algorithms for activity recognition for falls and subject-independent evaluation testing showed that the human subjects greatly drops whenever a context of diﬀerent classiﬁer performance strongly depends on the subject data. *e data distribution compared with that of the training data is authors in [78] show the eﬀect of using an optimization confronted [53]. Personalized exercises may be inadequate technique to increase the accuracy of an SVM classiﬁcation to be directly used as training data for another subject so the model. authors in [53] applied a cross-subject transfer learning While the reported accuracy in most of the research done algorithm that can link source and target signals through the for fall detection is above 90% [28–30], the practicality of these construction of manifolds at the feature level. Another way 8 Journal of Healthcare Engineering with autism spectrum disorder (ASD) was investigated in to approach this problem is to build a personalized model for each subject, and this approach was investigated by the [58]. *e authors in [24] studied stress detection using a neural network for metabolic syndrome patients as the authors in [26] as they see that people perform activities in diﬀerent ways and that general models may average out increase in stress may result in chronic symptoms. important individual characteristics, besides that personal- Other mental disorders such as depression, anxiety, and ized models can learn from much fewer data and guarantee bipolar disorder [84] have also been studied in the literature better privacy for data collected from accelerometers and [85] using features from biosignals, eye sensors, micro- gyroscopes in wearable devices. Earable (ear-worn) devices phone, camera, or interactions with smartphone to assess can also be used for human activity recognition. *ey were social behaviors. found to have a superior signal-to-noise ratio under the inﬂuence of motion artifacts, and they are less susceptible to 2.2.4. Arrhythmia Detection. Heart rate tracking could be acoustic environment noise [79]. noticeably seen in some commercial wristband and smart Eating activity monitoring, sometimes also referred to as watches. *e detection of irregular heartbeats (arrhythmia) automated dietary monitoring (ADM), is essential for pa- is a relatively recent goal for commercial wearables. Fast tients’ diet assessment and following up with taking med- heartbeats (> rbin100 bpm) are called tachycardia, while ication [33] for elderly people by monitoring taking a pill slow heartbeats (< 60 bpm) are called bradycardia. Atrial activity. *is is considered an activity recognition task, but it ﬁbrillation is one type of arrhythmia that involves the rapid is added to a separate category “Eating Monitoring” in and irregular beating of the atrial chambers of the heart. Table 1. In [37], the authors proposed a proximity-based Apple conducted a clinical study to detect atrial ﬁbrillation active learning on accelerometer data obtained from a [3] in 419,297 participants using PPG sensors in Apple wrist wristband wearable device, which is a novel proximity-based watch patches, but they used non-machine learning algo- model to recognize eating gestures. In [36], the author rithm based on a proprietary threshold analyzed from data assessed using an EMG sensor and contact microphone for the degree of dispersion of inter-peak intervals to de- behind the ear near the jaw to record chewing sounds and termine irregularity. After a monitoring period and ana- detect eating activities. *ey used 8 features extracted for a 3- lyzing the results, participants with detected irregularities second window size for eating detection of crunchy and soft were notiﬁed to do ambulatory ECG monitoring using ECG food. A study for eating episode recognition [55] used two patches, of which only 34% responded (450 participants). IMUs, with one put on ear and the other one on the upper Similar to the clinical study done by Apple [3] to detect atrial back, and they trained a random forest with the sensors’ data ﬁbrillation, Huawei and Fitbit have recently launched their and labels. Another study [56] used features from inertial atrial ﬁbrillation study in mid-2020 (https:// sensor data placed on the downside of the lower jaw to detect cardiacrhythmnews.com/wearables-devices-the-new- eating episodes. A review of the research done until 2019 in frontier-in-arrhythmia-management/). the ﬁeld of eating detection comparing diﬀerent studies in *e authors in [59] used the SVM model to identify the terms of the used sensors, methods for collecting the data, raw heartbeats. *en, with an unsupervised dynamic time and evaluation metrics was discussed in [80]. *e authors warping (DTW)-based learning approach using the pointed out that most of the studies included accelerometer K-medoids clustering method, the distorted heartbeats are data from a wrist-worn device for accessibility and ease of identiﬁed and puriﬁed. SVM and bagging trees have been use, and they mentioned that the implementation of novel used in [11] to detect atrial ﬁbrillation from features from methods for naturally acquiring ground truth remains a ECG signals. challenge. A similar approach can be used for drinking In [10], the PPG signal was alternatively used. It was episode detection [81] and smoking cigarette detection [82]. recorded for patients with atrial ﬁbrillation using both a Fitness tracking is another application that can also be conventional oximeter and a cardiotracker ring, which considered as an activity recognition task. In [38], the au- generated comparable results. A convolutional neural net- thors were able to identify jogging periods using acceler- work achieved better results when compared to diﬀerent ometers and they concluded that there is no signiﬁcant SVM variants. A worst case accuracy of 94.7% was achieved beneﬁt from using accelerometers on both hip and ankle for 10-second recording periods. Although PPG signals have locations over using only one accelerometer. Segmentation limitations such as noise introduced by motion artifacts, the of exercise and non-exercise periods and recognizing which authors concluded that the ring PPG-based wearable has exercise is being performed were investigated in [57]. good diagnostic performance for atrial ﬁbrillation and can replace ECG-based detection. *ey also mentioned that considering longer periods for PPG signals may aﬀect the 2.2.3. Stress Detection. A survey for stress detection using performance due to false positives with atrial tachyar- diﬀerent signals such as heart rate (HR), blood volume rhythmia episodes. A deep learning model has also been pressure (BVP), inter-beat interval (IBI), electrodermal used in [60] but with the best accuracy of 89% achieved activity (EDA), temperature data, and behavioral features learning from both ECG and PPG sensor data. was conducted in [20]. *e authors found that the most distinctive features for detecting stress are EDA and HR. Remote monitoring of child safety through stress patterns 2.2.5. Seizure Detection. Epilepsy is a neurological disorder was tackled in [83]. Detecting stress and anxiety in children that aﬀects the central nervous system, causing seizures or Journal of Healthcare Engineering 9 [21, 70, 71]. In [71], the authors used heart rate variability periods of unusual behavior such as twitching in legs and arms and sometimes loss of consciousness. Detecting sei- (HRV) parameters: RR interval of the ECG signal, standard deviation of RR interval (SDRR), and root mean sum of zures is important to help the patient get help when needed by alarming a caregiver. Diﬀerent types of seizures and squares of diﬀerences between adjacent RR intervals wearable devices for detecting seizures are reviewed in [86]. (RMSSD), from the ECG signal as features with labels at rest, *e study in [63] included the highest number of partici- post-exercise, and post-hydration to detect dehydration. *e pants (135) and used Embrace Empatica Watch [64] with an authors in [21] used EDA and other heart rate variability accelerometer and EDA sensors. parameters extracted from PPG signal for mild dehydration Forecasting seizures can also be useful to alarm the identiﬁcation. *e authors in [43] used multimodal sources patient to rest and take protective measures. Seizure fore- of features from diﬀerent sensors to predict the last drinking casting has been investigated in [8] using deep learning on time of the user, which would ease collection of data and provide ways for personalization on-device. multimodal wristband sensor data from 69 patients with epilepsy (total duration >rbin 2,311 hours, 452 seizures). In [62], the authors investigated the use of support vector 2.2.8. Emotion Recognition. Emotional state monitoring for machine (SVM), random forest (RF), naive Bayes (NB), construction workers in a real worksite using a wearable K-nearest neighbor (KNN), and neural network (NN) to EEG sensor [15] was classiﬁed as positive (e.g., excitement, diagnose an epileptic seizure based on EEG sampled dataset happiness, contentment, or satisfaction) or negative (e.g., available at the UCI machine learning repository. Similarly, fear, anger, frustration, or depression) based on measuring the authors in [64, 65] have used both EDA and acceler- the EEG valence level and cortisol biochemical response as a ometer data for detecting seizures but with diﬀerent datasets reliable marker tested from saliva samples after each task. and techniques. *is can better be performed using machine learning techniques to replace the cumbersome cortisol testing. 2.2.6. Rehabilitation Tasks. Rehabilitation tasks involve Considering fear emotional state, fear level classiﬁcation tasks to improve abilities needed for daily life, which may be using diﬀerent machine learning techniques (KNN, RF, physical or mental abilities that have been lost or impaired LDA, SVM, and deep learning) has been researched in [89] depending on features extracted from EEG, GSR, HR, and due to injury, underlying disease, genetic disorders, or birth defect. One example for a rehabilitation task is foot strike subjective unit of distress (SUD) values in a virtual reality therapy setting. Using EEG raw data, the authors in [16] angle prediction, which was studied in [69], which can help in the coaching of running movements and consumer-based introduced using a liquid state model (LSM) for training to predict valence, arousal, and liking levels at diﬀerent du- shoe prescription. Diﬀerent machine learning techniques were compared, and random forest achieved the best ac- rations of the EEG input signal. curacy of 94.1%. Linear discriminant analysis (LDA) was used in [87] to 2.2.9. Sleep Monitoring. For sleep monitoring and sleep classify each subject as a normal or abnormal gait pattern. quality assessment, detecting diﬀerent sleep states (awake, *e authors used real-time acoustic feedback (RTAF) to rapid eye movement (REM) sleep stage, and non-REM support the subjects when they are performing the tasks stages) is a requirement. Classiﬁcation of sleep episodes has during the rehabilitation session, so that they are able to been studied in [32], where a random forest model was used adjust their motion pattern to the acoustic feedback. Support to detect diﬀerent sleep-wake states with an F1 score of vector regression (SVR) models yielded excellent intraclass 73.93% after being trained with the data from accelerometers correlation coeﬃcients (ICC) in the gait parameters (stride on the wrists of 134 subjects. length, stride velocity, and foot clearance) analyzed in [68] Sleep monitoring applications such as detection of sleep for both walking and running exercises. Similarly, the au- apnea episodes have been studied in [1]. Sleep apnea is a thors in [67] investigated the same problem on a diﬀerent problem accompanied by increased cardiovascular risk and dataset using the K-means clustering, SVM, and artiﬁcial decrease in the quality of life. *e authors compared auto- neural network (ANN). correlated wave detection with an adaptive threshold (ACAT) for both electrocardiogram (ECG) data and PPG 2.2.7. Hydration Monitoring. Hydration monitoring to sensor data to detect the cyclic variation of heart rate detect dehydration is another problem being researched for (CVHR). *e classiﬁer was able to discriminate sleep apnea its importance, especially for athletes, battleﬁeld soldiers, episodes from non-apnea episodes with 82% sensitivity, 89% workers in hot conditions, and elderly people who are not speciﬁcity, and 85% accuracy depending on PPG signals. able to communicate their need for water. *ere is an on- Electrodermal activity, accelerometer data, heart rate going progress in the development of biochemical sensors variability, and blood volume pressure during sleep have been used in [7] using wearable Empatica E4 smart watch for that can measure the concentration of diﬀerent electrolytes in sweat and hence determine the hydration state [88]. Side early detection of migraine from the quality of sleep to enable early alarms to take preventive medication. *ey by side, machine learning research studies try to learn from other diﬀerent body signals to detect dehydration based on achieved a balanced accuracy of over 84% for detecting the eﬀect of cognitive stress triggered by dehydration on the migraine attacks using quadratic discriminant analysis as a autonomic reactions of the body as the work done in classiﬁer. Another dataset is published in [90] that can be 10 Journal of Healthcare Engineering they use diﬀerent datasets, rely on diﬀerent features, and solve used for sleep stage prediction for which accelerometer and heart rate data were collected from Apple Watch, while the diﬀerent problems with multiple experiments. In Figure 4, a box plot is shown for the range of values reported for accuracy subjects underwent polysomnography during night sleep and it was used for sleep-wake classiﬁcation in [91] using of classiﬁcation models in research studies cited in Tables 1–3 ANN with accuracy 90%. under 5 group models (KNN, SVM, logistic regression, tree- based models (random forest, decision trees, extremely ran- (i) Disease diagnosis artiﬁcial intelligence research has domized trees), and deep learning (DNN, MLP, LSTM, been in health care and medical diagnosis of diseases CNN)). It can be seen that the best median average accuracy for a long time ago. Starting from expert systems in achieved is for using deep learning. Logistic regression, SVM, the 50 s of the last century, continuous eﬀorts have and tree-based models follow deep learning with very close been going on in this ﬁeld until recently applying values. All models are away from perfect classiﬁcation, but deep learning techniques for improved diagnosis of some are useful and there is a room for improvement over all diseases [92]. Examples include lung cancer diagnosis tasks using larger datasets, extracting more meaningful fea- based on CT scans and diagnosis of skin conditions tures, and modeling for personalization as body signals vary through scanning skin images, which has recently according to each person lifestyle, weight, height, and activity been announced by Google [93]. *e use of artiﬁcial level. intelligence and machine learning techniques through wearable devices for initial assisting diag- nosis and detection of symptoms is foreseen to be the 2.3. Datasets. PhysioNet (https://www.physionet.org/about/ upcoming future, especially in the COVID-19 pan- database/) is a big database that oﬀers large collections of demic circumstances we are passing through and the physiological and clinical data and related open-source quarantine protective requirements imposed in most software for research purposes in many areas such as sleep of the countries. apnea detection, arrhythmia recognition, stress detection, and human activity recognition. It was established by the With the onset of the COVID-19 pandemic, the authors National Institutes of Health (NIH) and is maintained by in [76] proposed a protocol for using a mobile health MIT Laboratory for Computational Physiology. platform to analyze the biosignals recorded by Everion wearable (skin temperature, respiratory rate, blood pressure, *ere are many other datasets available for human ac- tivity recognition and fall detection that are mentioned in pulse rate, blood oxygen saturation, and daily activities), together with a recording for the cough for early detection of [27]. MobiAct [48] (57 subjects/9 ADLs and 4 fall types) is the largest of them in terms of the number of subjects and is COVID patients. For the limited places for patients to re- ceive hospital care, which was observed by the spread of suitable for both fall detection and human activity recog- nition tasks. It is available upon request for research pur- COVID-19 in some countries, the researchers in [22] used poses. UCI-HAR dataset [96] provided by the University of machine learning-based analytic systems to detect early California Irvine is the most famous and cited dataset in the signs of clinical deterioration to schedule and guarantee domain of human activity recognition. *ey recorded the 3- resources’ optimization. *ey also used Everion biosensor to axis accelerometer data and 3-axis gyroscope angular ve- record many physiology parameters such as heart rate, heart locity time series at 50 Hz for 30 subjects while doing 6 rate variability, respiration rate, oxygen saturation, blood pulse, skin temperature, and actigraphy to monitor mild activities (WALKING, WALKING_UPSTAIRS, WAL- KING_DOWNSTAIRS, SITTING, STANDING, and LAY- COVID-19 cases and predict clinical deterioration accord- ingly. Hypertension diagnosis has also been studied in [75] ING) using smartphones banded on the waist. Friedrich-Alexander-University oﬀers many movement using deep learning for continuous monitoring of blood analysis datasets (https://www.mad.tf.fau.de/research/ pressure depending on one-channel ECG and PPG signal activitynet/) such as daily life activities, step activities, cy- that can be obtained from a wearable device. clic activities, gait analysis datasets, and energy expenditure *e use of electronic monitoring devices for asthma has estimation. A dataset of 3D accelerometer data speciﬁcally been reviewed in [94]. *e authors suggested that clinicians for eating activity recognition for 20 participants in labo- should evaluate asthma management applications to ensure high-quality and evidenced-based information before pa- ratory setting and 7 participants in free-living conditions is made available for research purposes by the authors of [56]. tients use them as current studies only analyzed asthma patients according to their sleep quality and physical activity A recent real-life human activity dataset was published by the University of A Coruna [97]. *ey recorded about measures. In [95], multiple features of motion and dexterity and 189 hours of measurements from the accelerometer, gyro- scope, magnetometer, and GPS of smartphones for 19 dif- sleep measures were collected using IMU sensors on the ferent subjects with no restriction for mobile position. *e chest, wrist, and ankle to correlate these measures with data have four labels that deﬁne diﬀerent activities (inactive measures of neurological disability in multiple sclerosis. for not carrying the mobile phone, active for carrying the As can be seen from Tables 1–3 and the review in this mobile phone and moving (making dinner, being in a subsection, no single model can be chosen for every problem as concert, etc.), walking for moving to a speciﬁc place whether it depends on the dataset size, the features extracted, and the problem being learned. It is very diﬃcult to compare diﬀerent jogging or running, and driving for moving in a car, bus, truck, etc.). techniques and the results reported in the research studies as Journal of Healthcare Engineering 11 k-NN SVM LR Tree-based Deep learning Machine learning models Figure 4: Box plot of accuracy for the machine learning techniques used in diﬀerent classiﬁcation problems for papers cited in Tables 1–3 with accuracy as the evaluation metric. On each box, the central mark is the median and the edges of the box are the 25th and 75th percentiles. *e small circles represent outliers. Sleep Data (https://sleepdata.org/datasets) have large 3. Challenges for ML Applications on collections of de-identiﬁed physiological signals and clinical Wearable Devices data elements that are oﬀered by the National Sleep Research Resource (NSRR) to help in sleep monitoring research. Some Developing machine learning applications in general follows of these signals can be obtained through wearable devices the Cross-Industry Standard Process for Data Mining Cycle using diﬀerent sensors’ types. (CRISP-DM 1999) [100]. *e development-to-deployment RecoFit [57] contains accelerometer and gyroscope re- process involves many challenges in collecting the data, cordings from over 200 participants performing various gym selecting the best features, selecting the libraries and exercises. framework [100], evaluating the trained model(s), selecting *e seizure gauge dataset (https://www.epilepsyeco the best model, and relying on the ML model decision since system.org/my-seizure-gauge-1) records long-term physio- practically no ML model is guaranteed to be 100% accurate. logical signals such as EMG, PPG, EEG, ECG, accelerometer ML learning models for health care are to be designed to signals, BVP, EDA, and temperature from diﬀerent people generalize well and deal with unseen examples while taking with epilepsy using three diﬀerent wearable devices. into account personal features, providing interpretation for iRhythm arrhythmia detection public test dataset (https:// the result, and communicating the results cautiously. Some irhythm.github.io/cardiol_test_set/) is a dataset used in [9] for issues can be handled through clinical and preclinical testing a model used for arrhythmia classiﬁcation for 336 studies, to provide a suitable user interface and a note for records of 30 s strips single-lead ECGs captured at 200 Hz from conﬁdence or reliance on the results to be provided as per 328 patients who used a single-lead ambulatory ECG moni- the regulatory requirements. *e model needs to be toring patch. Each record is annotated by a consensus label implemented and used in both retrospective and prospective obtained by a committee of three cardiologists. studies, and the clinical impact measured [101]. A database for emotion analysis using EEG signals (https:// In addition to the typical challenges facing any machine www.eecs.qmul.ac.uk/mmv/datasets/deap/) and peripheral learning application concerning the used data and model, physiological signals was collected, while the participants there are many challenges that developers of a machine watched 40 one-minute music videos [72]. A dataset for learning application for a wearable device should take care studying social stress using blood volume pulse (BVP) and of, all challenges are shown in Figure 5 and are presented in electrodermal activity (EDA) signals has been recently pub- the next subsections, and how they aﬀect the choices lished [98]. Cognitive load, eﬀect, and stress recognition have available for developers. been studied in [99] through recording the biosignals (ECG, PPG, EDA, and accelerometer data) of 62 healthy volunteers whileansweringmathproblems,logicproblems,andtheStroop 3.1. Data Availability and Reliability. Machine learning test. *e Stanford wearable dataset (http://ipop-data.stanford. approaches, in general, and especially in certain healthcare applications, require the availability of enough data for edu/wearable_data/Stanford_Wearables_data.tar) was used in [9] for arrhythmia detection and classiﬁcation. training to generalize well for unseen data. As presented in Accuracy 12 Journal of Healthcare Engineering To ensure data reliability, conducting a wide range of clinical experiments while reporting the results transpar- Data ently is critical for evaluating diﬀerent techniques [103] and Availability & ﬁnding promising research directions. Medicolegal aspects Reliability Model Storage need to be well-deﬁned and regulated [104]. As an example, Selection & Limitations Reliability the authors in [105] provide guidelines for future study data collection and design for heart rate data, data cleaning and processing, analysis, and reporting that may help alleviate the data reliability challenge. Power Challenges to ML Consumption Application Deployment Limitations on Wearable alternatives 3.2. Model Selection and Reliability. For reliably calculating the accuracy of machine learning models, the use of cross- validation techniques is considered one approach to achieve this by testing the model on unseen data that have not been used in training. In [106], the authors reviewed the research Security Communication work of using either record-wise or subject-wise cross- Utility Privacy & user validation. *ey experimented with a publicly available acceptance dataset for activity recognition and simulation data to ﬁnd that using record-wise cross-validation overestimates the prediction accuracy of machine learning algorithms. *is Figure 5: Challenges to healthcare ML applications on wearable result agrees with the research ﬁndings in [27]. Diﬀerently, devices. some of the authors in [107] criticized the work done in [106] with arguments that no within-subject dependence between observations can be detected so record-wise cross-validation Tables 1–3 in the last section, the study with the maximum can be used. *e authors also suggest avoiding leave-one-out number of subjects [9] included 53,877 patients in a retro- cross-validation and recommend keeping the number of spective study that was funded by a commercial company for folds large enough while following strategies such as re- manufacturing ECG patches. *e rest of the studies depend peated test-train split, shuﬄe-split, repeated K-fold, or on far fewer data since the health data collection is expensive, Monte Carlo cross-validation to avoid overﬁtting and ensure which makes their results questionable for reliability. generalization. For wearable devices, since the models *e data obtained from wearable technologies have to be usually represent. reliable as well with deﬁnite conﬁdence and clear warnings to For model selection, there are many criteria that aﬀect ask for medical staﬀ help for any concern as human health is the the decision when it comes to wearable devices [43]. One of ultimate goal. *e authors in [102] investigated the sources of them is to maximize the evaluation metric used to report inaccuracy in diﬀerent wearable optical heart rate sensors. *ey accuracy for classiﬁcation or minimize the error metric used explored heart rate and PPG data from consumer and research- in regression problems. Usually training an ensemble of grade wearables while doing diﬀerent activities for diﬀerent skin diﬀerent models achieves the best performance. Interpret- tone subjects. According to their ﬁndings, there was statistically ability or explainability of the model [108] is another cri- no signiﬁcant diﬀerence in accuracy across skin tones, but terion as most of the applications for wearable devices target signiﬁcant diﬀerences between devices and between activity healthcare application for which the result of classiﬁcation/ types were remarkable with an average absolute error of 30% regression or clustering is to be explainable and makes sense more than during rest. *e reliability of data in health care is so for the user. Tree-based models are seen as more inter- important for the patient and physician to rely on the device pretable than neural network-based models [43]. *e size of readings to take the most appropriate medical decision, which the model to ﬁt on the wearable device with limited memory may in some cases threaten the life of a human. *is what led is among the criteria. Additionally, the computational Verily Life Sciences (formerly Google) to discontinue their complexity for inference and for online training on the glucose-sensing lens project (https://www.business-standard. device for personalization is a concern due to the limited com/article/news-ians/google-halts-project-to-build- computation power for wearable devices until now. On- glucose-sensing-contact-lens-118111800398_1.html) when device deep learning and transfer learning for personali- their ﬁndings reveal that there is insuﬃcient consistency in the zation have been researched in [43, 109–111]. correlation between tear glucose and blood glucose concen- trations. *eranos is another example that went dreadful after being charged for wire fraud (https://www.fda.gov/inspections- 3.3. Deployment Alternatives. *ere are three deployment compliance-enforcement-and-criminal-investig alternatives for the machine learning model for the wearable ations/press-releases/june-15-2018-theranos-founder-and- device scenario, either to deploy the model on the wearable device, or on an edge device or on the cloud as shown in former-chief-operating-oﬃcer-charged-alleged-wire-fraud- schemes) when investigations found they advertised rapid Figure 1. Each deployment alternative has some advantages and disadvantages that might make it impractical in some blood test devices that they knew were likely to contain inac- curate and unreliable results for diﬀerent blood tests. cases. Deploying the machine learning model on the device Journal of Healthcare Engineering 13 alternatives and maintaining the continuous update of the has the advantages of keeping the data private and de- creasing the latency for the prediction/classiﬁcation as there ML cycle. A survey for the diﬀerent automated machine learning approaches to automate feature extraction and is no need to transmit large amounts of data from the device to the cloud. In particular, for healthcare applications, selection, hyperparameter optimization, pipeline optimizers, having the patient’s data and the machine learning model on and neural architecture search for healthcare systems can be the device is more safe from privacy-preserving perspective. found in [115]. Low-latency and real-time feedback is also required for many healthcare applications that require immediate alert for the users or their caregivers such as fall detection and 3.4. Power Consumption. Power consumption is the main limitation of wearable devices due to their limited battery stress detection. On the other side, the main disadvantages of on-device computing include the limited device computing lifetime in general. For machine learning applications on wearable devices, the power consumption is greatly aﬀected by power, storage, and battery life. With these limitations, oﬄoading computations to be the need to send physiological data measured by the device’s sensors to the cloud to perform computations on the cloud. At done on one or more edge devices (e.g., smartphones or locally on a hospital/house/oﬃce terminal/gateway) is one the time of writing this study, the best commercial smart watch solution [112]. Edge/fog computing has also many advan- battery lifetime is just a few weeks, which monitors walking and tages over the cloud computing alternative [112] in terms of running activities and give an approximate measure of the pulse security, latency, power consumption, real-time processing, rate and oxygen saturation. *is could be far less in practice and and bandwidth load [113]. Edge computing can reduce data could be as low as a few hours for wearables that monitor transmission to the cloud and consequently reduce power multiple vital signs continuously for alerting users to abnormal situations (e.g., alerting for abnormal heart rhythm or detecting consumption and improve privacy by analyzing sensitive private data on a local gateway, ﬁltering it, and compressing fall). *e elements that aﬀect the power consumption in it, instead of doing it on a cloud away from the user’s control. However, this depends on the size of the machine learning wearable devices include the board, its components of diﬀerent biosensors and their sampling rate, the operating model and the data streams to be used in training and testing, the need for online training and real-time prediction, system and other software running on the board, the and the computational power needed for training and wearable display, the rate of logging data on the device, and testing. the amount of data transmitted over the communication *e main advantage of adopting cloud computing is the channel (e.g., Bluetooth or Wi-Fi) to be sent to the edge/ ﬂexibility of storage and computational resource on-demand cloud. scalability. *is comes as a trade-oﬀ for higher costs, power Transmission and reception of data are thought to consume more energy than sensing and logging data. consumption, latency, and challenges for preserving the privacy of both the data and the machine learning model as Research in the area of reduction in the power con- sumption can be seen to go in diﬀerent directions, de- will be discussed later. Data drift (how data distribution could change over veloping special embedded hardware for running time) and continuous integration and delivery are other machine learning algorithms [116, 117], reducing data to aspects that determine the decision of which deployment be transferred [118–120], compression [121] or sched- alternative to employ in machine learning applications for uling of the data to be transferred [122], computational wearable devices. oﬄoading [123, 124], and developing self-powered *e development process for wearable machine learn- wearable devices [125, 126]. ing-based software requires the same operations for any One approach suggested by the authors in [127] to software with some speciﬁc operations related to machine save the consumed power by the data transfer over the wireless connection is to perform embedded machine learning applications such as data collection, cleaning and preprocessing, continuous (re)training of the ML model, learning on the device, i.e., following the tinyML ap- proach. According to the analysis in their work, this can and continuous (re)-deployment of the updated model to the device or to the edge nodes or to the cloud service [114]. increase the battery lifetime by more than 70%. Re- Figure 6 shows the typical machine learning operations searchers in [128] proposed a hybrid approach of using (MLOps) in the development process of wearable software less battery, low sampling rate, and wearable RFID tags, for ML-based applications with the diﬀerent deployment which can be powered intermittently by a reader with alternatives (device, edge node, cloud service). *e feedback additional passive RF tags that capture the presence and arrows from the deployment process are orchestrated based use of speciﬁc objects for daily activities’ recognition. on the performance of the model on edge nodes or cloud As previously mentioned, another way to reduce service after getting feedback from users or updated models power consumption is to reduce the data stored and transmitted to the cloud, and the authors in [129] pro- received in case of federated learning scenario, which will be discussed in Section 3.8.3 to ensure continuous integration posed a variant of symbolic aggregation approximation (SAX) tested for compressing heart rate data, which (CI) and continuous deployment (CD) requirements. Tools such as Apache Airﬂow, Kubeﬂow, and Google Cloud proved to achieve the best trade-oﬀ between diﬀerent AutoML support the software lifecycle operations of ML performance metrics for systems that require short components by orchestrating the diﬀerent deployment latency. 14 Journal of Healthcare Engineering Orchestration Containerization Scope Collect Train Deploy and and Extract machine model in design clean features learning production the project the data model Cloud Edge Device Figure 6: MLOps for wearable device application. 3.5. Storage and Memory. Typically, existing wearable de- *us, research goes on in many directions to overcome vices have limited memories (e.g., Apple Watch Series 6 these factors. From the data perspective, data selection and released in September 2020 has only 1 GB RAM) due to dimensionality reduction techniques are employed. From the small device size and weight requirements. Wearable and model perspective, designing new models with acceptable IoT devices use nonvolatile memory (e.g., ﬂash, EEPROM, prediction accuracy while minimizing model size and pre- diction costs such as Bonsai [132] is another approach. Com- MRAM, and F-RAM) to ensure resilient system recovery on sudden shutdown with the limited battery lifetime and to pressionofmodelscantakeplacebypruning(usinglessnumber of weights), quantization (using less bits per weight) [133], and ensure short boot time. While ﬂash-based storage is con- sidered a de facto storage standard for IoT devices for its encoding. *e authors in [134] reviewed model compression speed and stability [130], F-RAM is commonly used for techniques. Some of these techniques are implemented in medical wearables for its low power operation and high- TensorFlow Lite (https://www.tensorﬂow.org/lite). write cycle endurance, which allow it to reliably and eﬃ- ciently store more data logs from sensors [129]. EEPROM is sometimes also used since it is more reliable and smaller 3.6. Utility and User Acceptance. Users of wearable devices have been growing over the past few years, especially ﬁtness than ﬂash memories for applications that do not require frequent write operations and requires less power. In [131], trackers. Nevertheless, there is still a lack of user acceptance to adopt other wearable devices incorporating AI solutions the authors proposed using battery-backed RAM on for healthcare tasks. wearable devices and eﬃciently oﬄoad energy-intensive According to [135], 35% of 1,183 adult patients in France tasks to the smartphone/edge device to perform small and would refuse using wearable monitoring devices and AI- energy-eﬃcient tasks locally using battery-backed RAM. In addition to the development of memory archi- based tools in their care. Another study in the United States [136] examined the response of 307 consumers to the per- tectures (in-memory computing) and hardware (appli- cation-speciﬁc integrated circuits (ASICs)) that are ceived beneﬁts and risks of AI medical devices with clinical decision support (CDS) features. *e results of the study show capable of running machine learning applications on battery-operated devices, tinyML Foundation (https:// that performance/accuracy and communication, besides the ethical and regulatory concerns to keep the data private and www.tinyml.org/), which started in 2019, has also fo- secure, signiﬁcantly contribute to the perceived risks of using cused on signiﬁcant progress on algorithms, networks, AI applications in health care. Regulatory agencies should and models down to 100 kB and below to perform on- establish a standard and evaluation guidelines for the device analytics at extremely low power, thus minimizing implementation and use of AI in health care. Privacy and bandwidth and latency concerns while providing higher security concerns are among the major concerns raised for the privacy. *e practicality of deploying a machine learning application use of wearables. For example, there are security concerns raised for using Google Glass for recording people data on a wearable device or an edge device depends on many factors: the size of the device, the data size (features and time without their permission. It has been proven to be a serious issue since it can be used (like any recording device) to steal span of physiological data used for prediction), the complexity of the model (no. of parameters and layers), and use of batch or passwords by recording and analyzing the shadows of ﬁnger movements on a screen while typing a password [137]. *us, real-time processing. A model with high accuracy often requires the ﬁrst version of Google Glass failed to gain social accep- more memory for the number of parameters and layers in the tance [138] before releasing its second version and funding model than lower accuracy models. Depending on the machine some research studies about its usability, for example, its learning application, some machine learning models can reach desirability for a sample of school children with autism [139]. up to an order of 100 megabytes or even gigabyte (speciﬁcally Another important factor for user acceptance is how those including image inputs), which cannot ﬁt on the best wearable device along with the memory needed for doing the comfortable the device is for daily use. Design guidelines for wearable devices are identiﬁed in [140]. For example, computations. Journal of Healthcare Engineering 15 [144]. For example, accelerometer and gyroscope data on a designing a wearable should follow the anatomical structure of the body, take into consideration diﬀerent gender re- smart watch can be analyzed to reveal passwords and credit card information (https://securelist.com/trojan-watch/ quirements, and choose materials that are comfortable for the body and do not cause irritation to the skin. Addi- 85376/). Other attacks on IoMT devices can be life-threat- tionally, it is preferable to be used in a free-moving envi- ening such as attacks disrupting the medical service, e.g., ronment and it is required to be as easy as possible to use denial-of-service attacks (DoS) and ransomware attacks. without the need for many setup and conﬁguration steps. Whether wearable devices are used for health monitoring or *us, a wearable device should be compact and simple to for ﬁtness tracking, sensors’ data and other personal data are operate and maintain while providing secure and private being exchanged and analyzed by machine learning services experience for both the wearer and the people around him. to detect patterns and do classiﬁcation/prediction based on More awareness endeavors of the wearable technology to the the data. While it seems to be “a no problem to share” for some users, most end users are skeptical about how their public need to take place and the advertisers should abide by honest marketing about the product’s actual impact. personal data exchanged with such services is being used and how secure they are against diﬀerent types of attacks. *e issue of security and privacy of personally identiﬁable in- 3.7. Communication. In case of edge computing model, the formation and medical data in wearable and other IoMT intra-communication between the wearable device and the devices’ applications is critical and is regulated by diﬀerent edge device can be done over one of the standards such as data protection standards across the globe. Bluetooth, Zigbee, RFID, NFC, and UWB. Usually, light- In the case of wearables, the connection is usually done weight Bluetooth is employed for its low power consump- over lightweight Bluetooth as mentioned earlier and as tion [141]. However, according to Bluetooth 5 speciﬁcation, shown in Figure 1. Security guidelines for Bluetooth pro- the Bluetooth protocol allows up to 7 devices’ simultaneous vided in [145] consider wearable sensor devices as “Class 1.5 connections to a device and practically performance de- Low Energy” devices with a maximum output power of grades and pairing problems arise when there are multiple 10 mW that can operate for up to 30 meters distance but are connections to a smartphone. Other factors that aﬀect the typically used within 5 meters. *e guidelines show that for choice of communication technology are the maximum this class, each service request can have its own security distance between the wearable and the edge device, the requirements. It recommends the use of Security Mode 1 required data rate for the wearable-to-edge device, and the Level 4 for medical devices, which requires low energy secure required latency [142]. *e intercommunications in the connections authenticated pairing and encryption using wearable model over the Internet run between the edge AES-CMAC and P-256 elliptic curve to the edge device. device and the remote service or directly between the *e main challenge for edge computing is to incorporate wearable device and the remote service are two-way data security into the design of wearable devices through using communication channels over transmission control proto- encryption and providing solutions to manage, update, and col (TCP) or user datagram protocol (UDP) at the transport secure the wearable devices. Security risks include but are layer with the Internet protocol (IP) at the network level. not limited to malicious hardware or software injections, TCP/IP is mostly adopted for lossless transmission of health denial-of-service attacks, and diﬀerent routing and physical data or machine learning model parameters over wide area attacks. Some of these attacks can be defended using ap- network (WAN). propriate administrative policy settings and incorporating At the application layer, hypertext transfer protocol diﬀerent ML-based solutions for detecting diﬀerent attacks (HTTP) is commonly used as the request-response model that may compromise the communication network, com- from the edge to the cloud services. TLS is often employed to putations, battery consumption, or storage [146]. secure HTTP communication over TCP; however, HTTP is Additionally, securing the data stored on the cloud, resource-intensive and is more suitable to be used for edge which is fed to the machine learning inference model, and or fog devices with high power and storage capabilities. securing the model itself represent a big challenge [147]. Not Other less-weight application layer protocols include con- only the medical data itself and the machine learning model strained application protocol (CoAP), message queuing are considered prone to privacy attacks, but also the social telemetry transfer (MQTT), and advanced message queuing dynamics and interactions with other users can be analyzed protocol (AMQP) [143]. MQTT is a well-known publish- as done in [148]. subscribe model standard used for IoT and wearable devices Potential solutions for privacy-preserving ML are dis- for being a lightweight protocol. It can facilitate one-to- cussed in detail in [149, 150]. *ese include techniques for many communications between wearable device(s) with low achieving diﬀerential privacy, cryptographic techniques, and power and storage and the edge device on the other side. client-based federated learning techniques. *e following *e two communication channels with their running provides a brief discussion of these methods. protocols at diﬀerent network layers are susceptible to the diﬀerent well-known network security attacks. 3.8.1. Diﬀerential Privacy. *e diﬀerential privacy concept 3.8. Security and Privacy. User data captured on wearable was ﬁrst introduced in [151] and refers to the process of devices and sent to machine learning cloud services as shown protecting private data by adding noise based on Laplace, in Figure 1 are subject to many security and privacy threats exponential, or Gaussian distributions. *e noise is added in 16 Journal of Healthcare Engineering technique to use a clustering model over encrypted such a way that enables data analytics while providing privacy guarantees of the perturbed data. Diﬀerential privacy data. *ey employed the mean-shift algorithm and homomorphic encryption for the arithmetic of ap- can be useful for applications such as health care due to its useful properties such as group privacy, composition, and proximate numbers. To overcome the computational robustness to auxiliary information. With diﬀerential pri- load of the mean-shift algorithm, they performed vacy, healthcare applications that employ machine learning each iteration on a sample of the data instead of the algorithms can still learn from the distribution of data whole dataset. without revealing the actual data of the patients. However, (2) Trusted Execution Environments (TEEs): TEE is a researchers in [152] concluded that privacy compromises secure area located inside the main processor in must be made to preserve utility, especially in the chal- particular architectures. It ensures the conﬁdentiality lenging multi-class classiﬁcation tasks based on their ex- and integrity of the data and code within the TEE. periments on two datasets with membership attack and Examples of TEEs are Software Guard Extensions attribute inference attack. *is utility-privacy trade-oﬀ has (SGX) from Intel and TrustZone from Arm. Intel’s also been discussed in [153], where the authors found that as SGX provides a trusted execution environment, the privacy level increases, the machine learning algo- called an enclave, which trusts only the CPU and the rithm—diﬀerentially private stochastic gradient descent in on-chip cache [166]. A user program (code and data) their case—targets the body of the distribution but loses must be partitioned into an untrusted portion and a important information about minority classes such as dying trusted portion that will run inside the enclave. SGX patients and minority ethnicity that are usually represented protects the conﬁdentiality and integrity of code and in the tail of distributions. data during execution within the enclave from Another challenge for practically using diﬀerential pri- malicious programs that may be running alongside vacy in healthcare wearable applications is that it is best used it, including privileged programs, such as the OS and for high-dimensional balanced big datasets. *is is not the hypervisor. Hunt et al. [167] employed the SGX to case in some personalized healthcare wearable applications build their system for privacy-preserving outsourced such as a fall detector, which only learns from accelerometer machine learning called Chiron to protect the signals where falls are considered of low frequency. training algorithm and the user data. Segarra et al. [168] employed SGX to present a secure streaming processing system speciﬁcally ﬁtted for medical data. 3.8.2. Cryptography-Based Methods. Traditional cryptogra- phy is valuable and eﬃcient to achieve conﬁdentiality when (3) Secure Multiparty Computation (SMPC): SMPC used in secure communication between parties and out- oﬀers cryptographic protocols in which the com- putation is distributed across multiple parties where sourcing the data for storage, but it is not valid when we need to perform the computation on conﬁdential data as it needs no individual party can see the other parties’ data preliminary data decryption. Here, we introduce some [169]. Two common approaches to achieve SMPC methods employed to perform computations on sensitive are garbled circuits and secret sharing. data without violating privacy. (i) Garbled Circuits: in this technique, two (or (1) Homomorphic Encryption (HE): the idea behind HE more) parties can jointly evaluate a function over is to use special encryption functions that enable the their private inputs [170]. *e main idea behind computation of encrypted data [154]. HE ensures this technique is to use a Boolean circuit to that the result from performing operations on represent the function that needs to be evaluated. *e gates of the function are garbled by one encrypted data, when it gets decrypted, is equivalent to the result of performing the same operations party, and the private inputs are garbled and without any encryption. HE has the drawback of exchanged using an oblivious transfer protocol. being impractically slow. However, it has been A garbled circuit can provide a solution for getting more practical and standardized over the last privacy-preserving computations [171]. For ex- few years. HE can play a very useful role in healthcare ample, consider a patient who wants to use a applications where privacy is crucial, and using the diagnosis service without revealing his data and a data is subject to regulations. Many works in liter- service provider also wants to hide his algorithm ature have demonstrated the idea of using homo- parameters, which are considered trade secrets. morphic encryption for privacy-preserving machine In this case, a service provider can convert his learning in medical applications [155]. Research algorithm into a Boolean circuit, garble the circuit, and send it to the patient to be evaluated studies in [156–161] have presented diﬀerent tech- niques to train a logistic regression model over without loss of privacy. encrypted data using homomorphic encryption. In (ii) Secret Sharing: in this technique, an entity can [162–164], techniques of using the naive Bayes preserve the privacy of its sensitive data by classiﬁer model without leaking privacy information breaking it up into multiple shares and distrib- by applying homomorphic encryption have been uting the shares to a set of non-colluding parties presented. Cheon et al. [165] have presented a where each party computes a partial result Journal of Healthcare Engineering 17 depending on the shares it received [172]. Fi- artiﬁcial intelligence and machine learning still face some nally, one of the parties can receive these partial challenges in medical wearable devices as presented in this results and combine them to get the ﬁnal result. review. In this section, we will discuss brieﬂy a summary for the main perceived pitfalls or diﬃculties facing applying SMPC protocols are widely used to provide privacy- machine learning research for wearable devices and high- preserving in machine learning applications. However, light the related machine learning research directions that SMPC fails to protect against exploratory attacks. Explor- need further development. atory attacks act by performing several queries on a fully *e training data input to a machine learning model is trained model to leak some information about the model considered the most crucial element in the machine learning parameters and its training data, such as if a speciﬁc example process as garbage in mean garbage out. *e ﬁrst step is to was used in the training set or not. With this information, choose well-calibrated sensors that are better validated the attacker can gradually train a substitute model that against benchmarked devices used in hospitals that have reproduces the same prediction of the target model [147]. To undergone plenty of clinical experiments or other gold address these kinds of attacks, Kesarwani et al. [173] pro- standard devices [180]. Care should be taken as some of the posed a monitoring scheme called extraction monitor to research wearable devices provide raw data [102] that re- track the queries issued by the user, evaluate the information quire clean-up of the signals for removing noise and motion that a user might leak from these queries, and give a warning artifacts as the work done in [181]. Identifying the inac- when the user exceeds the average number of queries needed curacies in the data collected and considering that most of to reconstruct the model. the sensors are only accurate during rest [102] have to be taken into account as this has implications on the drawn conclusions and health-related decisions using wearable 3.8.3. Federated Learning Methods. Federated learning was devices in research. Most of the research works that have ﬁrst introduced in [174]. Federated learning is a machine been cited in Section 2 use research-grade wearable devices learning setting in which many devices collaborate in and do not mention the preprocessing and cleaning up steps training a model in a centralized manner while keeping the of data. Signal processing techniques are better to be training data private and decentralized [175]. In the cross- employed to cure these signals and remove motion artifacts device federated learning setting, the server sends out an [182]. Moreover, clinical experiments are to be done to help initial model to the devices, and the clients then train the in deﬁning the reference signal/ground truth and obtaining model on-device with their data locally and send the updated clinical evidence. device model to the server. Updated models are combined at For some applications, the ground truth signal is not the server using federated averaging to update the initial known due to the complexity of the human body’s response model. *is process goes on by sending the updated com- and the diﬀerent responses for each individual. For this bined model until the metrics are satisfactory. reason, the collection of data from as many subjects as possible *is approach was tested in [176] by applying federated is recommended to develop algorithms to clean the data and learning to heart activity data collected from multiple smart build more accurate models that generalize well. However, the bands in a stress-level monitoring scenario. *e authors process of data collection is an expensive and time-taking achieved comparable accuracy while preserving the privacy process that most academic research work cited in Section 4, of the data and reducing the communication burden by only which was not funded by companies, depending on data from communicating the models’ parameters. Additional privacy- a relatively small set of subjects. A transparent and repro- preserving protections such as secure multiparty compu- ducible process for collection of data from wearables, training, tation or diﬀerential privacy may further be included in the and evaluation of models is recommended for gaining trust in federated learning setting to keep data and model statistics the research results and eﬀectively building over accumu- private from malicious clients [177]. In the research done by lating research eﬀorts. *e authors in [183] pointed out [178], the authors utilized both SMC and diﬀerential privacy recommendations for reporting machine learning results in to balance the trade-oﬀ between vulnerability to inference clinical research and similar guidelines are to be followed for and low accuracy in a federated learning setting. machine learning research for wearable devices, especially For most of the healthcare applications, the machine those used in healthcare as they aﬀect human life. learning model is better to be personalized as per the bio- One cheap approach for big data collection is crowd- signals for each patient. Model aggregation with federated sourcing data collection such as the one initiated by a re- averaging as mentioned above does not provide this per- search group at Stanford University (https://innovations. sonalization. *e authors in [179] applied transfer learning stanford.edu/wearables), which collects data from wear- in the federated learning setting so that each device can train able devices remotely through a mobile application. How- a personalized model tailored to the user’s data by utilizing ever, this approach is susceptible to many privacy issues that the cloud model and data and the local data. existing commercial smart watch entities do not handle and the data collection process will be only protected by the 4. Discussion privacy policy of the application. Privacy-aware sharing of *e use of artiﬁcial intelligence research has clearly been data and learning from it without revealing the actual users’ data employing some privacy-preserving techniques men- rapidly growing in healthcare applications. However, for healthcare wearable devices, it can be seen that practical tioned in the last Section 3 is an active research area. An 18 Journal of Healthcare Engineering wearable healthcare domain to give insight into the conﬁ- example for that is the work by the authors in [184] for privacy-preserving data collection using a local diﬀerential dence in the ML model, which would be helpful, especially in tasks such as seizure detection and diseases’ diagnosis de- privacy technique over salient data to protect users’ data. Research work in diﬀerential privacy has also open issues to vices. Besides uncertainty modeling in wearable ML appli- investigate new learning solutions, which can learn from cations, joining upregulation among key stakeholders in the data distribution tails (data that represent minority class) ﬁeld of healthcare wearables is a key to make sure that ML is eﬀectively while maintaining an acceptable privacy loss as introduced in wearables with more transparency from tech suggested by [153]. Federated learning is also a relatively new companies and for gaining better users’ trust and accept- privacy-preserving method that needs further attention and ability. As it is pretty obvious, ML applications for healthcare wearables are a multidisciplinary ﬁeld that requires stan- exploration in the ﬁeld of machine learning for wearable devices, which can also promote the personalization of the dards about naming conventions, evaluation metrics, ethical reporting of research results, and clinical impact as sug- local model. Another approach for augmenting the data input to the gested in [101] as these devices may directly aﬀect human life. Mentioning evaluation of machine learning models, it ML model is generating training data using generative adversarial network (GAN) variants that may help train was found that the use of subject-based cross-validation is good quality models without exposing users’ wearable data recommended since subject data represent a clinically more and signals used in training or without even using any real relevant scenario for disease diagnosis application than data but only simulated data as in [185]. using records from a subject in training while using some For guaranteeing some level of privacy for users of other records for the same subject in testing the machine wearable devices, the machine learning application, which learning algorithm. *is can let the machine learning al- gorithm learn an association between unique features of the usually holds identiﬁable information at the edge device (e.g., smartphone running the application), should follow a same subject and accordingly may fall into overﬁtting. *is raises a question about whether building personalized set of regulations to gain users’ trust. Besides compliance with HIPAA (Health Insurance Portability and Account- models can be more eﬀective as it learns from signals of the same individual to avoid averaging out important individual ability Act: https://www.hhs.gov/hipaa/for-professionals/ index.html), GDPR (General Data Protection Regulation: characteristics such as age, sex, weight, height, eating style, https://gdpr.eu/), HITECH (Health Information Technology and way of doing an exercise [26]. Furthermore, person- for Economic and Clinical Health: https://www.hhs.gov/ alized models can learn from much less data and guarantee hipaa/for-professionals/special-topics/hitech-act- better privacy for data. enforcement-interim-ﬁnal-rule/index.html), and act regu- Deciding upon the time window of the signal to learn lations, machine learning applications for wearable devices from is yet another challenging decision as more data do not necessarily mean better results. It faces memory limitations should follow the OWASP security standards (Open Web Application Security Project (OWASP): https://owasp.org/ on the wearable device and power limitations for sending this amount of sampled data from the device to the edge or www-project-top-ten/). Among other data challenges in machine learning for to the cloud. Consequently, the sampling rate for diﬀerent wearables is identifying which data to be collected from these physiological signals needs to be optimized as per the ap- devices. Diﬀerent modalities can increase the accuracy of the plication to optimize the use of resources and decrease the model as usually many signals can be used for a single task. power consumption. Nevertheless, exploring tinyML em- Another big challenge is how to model the uncertainty bedded solutions and models’ optimization techniques in arising from the complex input received by the human body IoT is a recent research area that is open to some applications as shown in Figure 2, which may aﬀect the accuracy of any in healthcare wearables as well. However, for computa- model. For example, considering the stress detection task is a tionally intensive applications, full computation oﬄoading can be eﬀective while for data-intensive applications oﬀ- complex task that involves many inputs and can aﬀect many body systems. Using multimodal sources is believed to be a loading techniques that oﬄoad the processing of some of the data will be more suitable as suggested by [187] while very rich source of information that can help in health monitoring by identifying the emotional state, stress level, preserving the privacy of users’ data. and diagnosis of some diseases. As an example, monitoring Most of the research works for applying machine audio signals from the user to detect laughing, crying, learning for healthcare wearable devices tasks that we shouting, or coughing can help in these applications on reviewed in Section 3 are experiments for learning from data wearables, but it faces other challenges [186]. EEG signal, obtained from one or more sensors for detection or rec- despite the diﬃculty of capturing it, can also hold a lot of ognition of some pattern. Complete analysis of the proposed models in terms of memory requirements and amount of information, which can help in achieving higher accuracy in many applications such as dehydration detection, emotion communicated data in case of edge or cloud deployment, which greatly aﬀects power consumption, is better to be recognition, and mental disorders detection. Moreover, depending on learning from only few body provided. Overcoming these diﬃculties with AI solutions, together signals, which may also be noisy, would lead to uncertain decision from the ML model. As handling and estimating with the ongoing research and development in the ﬁeld of uncertainty in ML modeling remain an active area of re- medical sensors, storage, SoCs, and power-eﬃcient man- search in ML, we recommend applying its techniques in the agement and generation for wearable devices, would ensure Journal of Healthcare Engineering 19 digital stethoscope platform,” Journal of American Heart having AI-enabled healthcare wearable devices that can help Association, vol. 10, no. 9, Article ID e019905, 2021. reliably with remote patient monitoring, detect problems [5] S. Seneviratne, Y. Hu, T. Nguyen et al., “A survey of wearable with the human body earlier and assist in diagnosis, perform devices and challenges,” IEEE Communications Surveys & elderly people care and monitoring, and act as a lifestyle Tutorials, vol. 19, no. 4, pp. 2573–2620, 2017. guide and much more. A lot of many other AI applications [6] M. Chan, D. Esteve, ` J.-Y. Fourniols, C. Escriba, and can receive user acceptance by solving problems with the E. Campo, “Smart wearable systems: current status and reliability, comfort, and privacy of the wearable devices. future challenges,” Artiﬁcial Intelligence in Medicine, vol. 56, no. 3, pp. 137–156, 2012. [7] P. Siirtola, H. Koskimaki, ¨ H. Monttinen, ¨ and J. Roning, ¨ 5. Conclusion “Using sleep time data from wearable sensors for early de- *e growth of using wearable devices over the recent years is tection of migraine attacks,” Sensors, vol. 18, no. 5, 2018. clearly noticeable. With the huge amount of research eﬀorts [8] C. Meisel, R. El Atrache, M. Jackson, S. Schubach, C. Ufongene, and T. Loddenkemper, “Machine learning to employ artiﬁcial intelligence solutions in healthcare-re- from wristband sensor data for wearable, noninvasive seizure lated tasks for wearable devices, their use is expected to grow forecasting,” Epilepsia, vol. 61, no. 12, pp. 2653–2666, 2020. even more and change from a “nice to have” devices with fun [9] A. Y. Hannun, P. Rajpurkar, and M. Haghpanahi, “-level applications to necessary devices for remote patient moni- arrhythmia detection and classiﬁcation in ambulatory toring and detection of any irregularities with the human electrocardiograms using a deep neural network,” Nature body. In this review, we presented ML tasks that have been Medicine, vol. 25, pp. 65–69, 2019. researched in the healthcare wearable devices ﬁeld, the [10] S. Kwon, J. Hong, E. Choi et al., “Detection of atrial ﬁ- machine learning techniques used, the diﬀerent modalities brillation using a ring-type wearable device (CardioTracker) used, and the available datasets in the ﬁeld. *e diﬀerent and deep learning analysis of photoplethysmography signals: challenges facing machine learning applications on wearable prospective observational proof-of-concept,” Study J Med devices (deployment alternatives, power consumption, Internet Res, vol. 22, no. 5, p. 2020. [11] Z. Mei, X. Gu, H. Chen, and W. Chen, “Automatic atrial storage and memory, utility and user acceptance, data ﬁbrillation detection based on heart rate variability and availability and reliability, communication, security and spectral features,” IEEE Access, vol. 6, Article ID 53566, 2018. privacy) were discussed while identifying possible solutions [12] N. Rashid and M. A. Al Faruque, “Energy-eﬃcient real-time found in the literature. Finally, the study highlights issues myocardial infarction detection on wearable devices,” Annu that require further research concerning data availability, Int Conf IEEE Eng Med Biol Soc, vol. 2020, pp. 4648–4651, reliability, and privacy to enable eﬀective and eﬃcient learning from data generated by wearable devices. [13] R. Buettner, J. Frick, and T. Rieg, “High-performance De- Contents of the research are solely the responsibility of tection of Epilepsy in Seizure-free EEG Recordings,” in the authors and do not necessarily represent the oﬃcial Proceedings of the International Conference on Information views of the Qatar National Research Fund (QNRF). systems, p. 12, Munich, Germany, December, 2019. [14] C. Ieracitano, N. Mammone, A. Hussain, C. Francesco, and Morabito, “A novel multi-modal machine learning based Conflicts of Interest approach for automatic classiﬁcation of EEG recordings in dementia,” Neural Networks, vol. 123, pp. 176–190, 2020. *e authors declare that there are no conﬂicts of interest in [15] S. Hwang, H. Jebelli, B. Choi, M. Choi, and S.H. Lee, this study. “Measuring workers’ emotional state during construction tasks using wearable EEG,” Journal of Construction Engi- Acknowledgments neering and Management, vol. 144, p. 01, 2018. [16] O. Al Zoubi, M. Awad, K. Nikola, and Kasabov, “Anytime *is publication was made possible by a grant from the multipurpose emotion recognition from EEG data using a Qatar National Research Fund (QNRF), project number Liquid State Machine based framework,” Artiﬁcial Intelli- ECRA 01-006-1-001. gence in Medicine, vol. 86, pp. 1–8, 2018. [17] A. Jaramillo-Yanez, ´ M. E. Benalcazar, ´ and E. Mena-Mal- donado, “Real-time hand gesture recognition using surface References electromyography and machine learning: a systematic lit- erature review,” Sensors, vol. 20, no. 9, 2020. [1] J. Hayano, H. Yamamoto, I. Nonaka et al., “Quantitative [18] C. Fang, Y. Wang, and S. Gao, “Utilizing wearable GRF and detection of sleep apnea with wearable watch device,” PLoS EMG sensing system and machine learning algorithms to ONE, vol. 15, Article ID e0237279, 2020. enable locomotion mode recognition for in-home rehabili- [2] F. Delmastro, F. D. Martino, and C. Dolciotti, “Cognitive training and stress detection in MCI frail older people tation,” in Proceedings of the 2020 IEEE International Conference on Flexible and Printable Sensors and Systems, through wearable sensors and machine learning,” IEEE Access, vol. 8, Article ID 65573, 2020. vol. 1–4, Manchester, UK, August, 2020. [19] B. Milosevic, S. Benatti, and E. Farella, “Design challenges for [3] M. V. Perez, K. W. Mahaﬀey, H. Hedlin et al., “Large-Scale Assessment of a smartwatch to identify atrial ﬁbrillation,” wearable EMG applications,” in Proceedings of the Design, Automation Test in Europe Conference Exhibition, New England Journal of Medicine, vol. 381, no. 20, pp. 1909–1917, 2019. pp. 1432–1437, Lausanne, Switzerland, March 2017. [20] Y. S. Can, B. Arnrich, and C. Ersoy, “Stress detection in daily [4] J. S. Chorba, A. M. Shapiro, L Le et al., “Deep learning al- gorithm for automated cardiac murmur detection via a life scenarios using smart phones and wearable sensors: a 20 Journal of Healthcare Engineering survey,” Journal of Biomedical Informatics, vol. 92, Article ID International Symposium on Wearable Computers, ISWC ’15, 103139, 2019. pp. 91-92, New York, NY, USA, September, 2015. [21] H. F. Posada-Quintero, N. Reljin, A. Moutran et al., “Mild [36] S. Bi, T. Wang, E. Davenport et al., “Toward a wearable dehydration identiﬁcation using machine learning to assess sensor for eating detection,” in Proceedings of the 2017 autonomic responses to cognitive stress,” Nutrients, vol. 23, Workshop on Wearable Systems and Applications, WearSys no. 12, p. 42, 2019. ’17, pp. 17–22, Association for Computing Machinery, New [22] K. C. Un, C. K. Wong, and Y. M. Lau, “Observational study York, NY, USA, June, 2017. on wearable biosensors and machine learning-based remote [37] M. Nourollahi, S. A. Rokni, P. Alinia, and G. Hassan, monitoring of COVID-19 patients,” Scientiﬁc Reports, “Proximity-based active learning for eating moment rec- vol. 11, p. 4388, 2021. ognition in wearable systems,” in Proceedings of the 6th ACM [23] D. Castaneda, A. Esparza, M. Ghamari, C. Soltanpur, and Workshop on Wearable Systems and Applications, WearSys H. Nazeran, “A review on wearable photoplethysmography ’20, pp. 7–12, Association for Computing Machinery, New sensors and their potential future applications in healthcare,” York, NY, USA, June, 2020. International journal of biosensors & bioelectronics, vol. 4, [38] E. Zdravevski, B. R. Stojkoska, M. Standl, and H. Schulz, pp. 195–202, 2018. “Automatic machine-learning based identiﬁcation of jogging [24] F. P. Akbulut, B. Ikitimur, and A. Akan, “Wearable sensor- periods from accelerometer measurements of adolescents based evaluation of psychosocial stress in patients with under ﬁeld conditions,” PLOS ONE, vol. 12, no. 9, pp. 1–28, metabolic syndrome,” Artiﬁcial Intelligence in Medicine, vol. 104, Article ID 101824, 2020. [39] D. Zhang, H. Zhang, and C. Zhao, “Prediction of behavior [25] C. T. Yen, J. X. Liao, and Y. K. Huang, “Human daily activity and activity of patients through wearable devices based on recognition performed using wearable inertial sensors deep learning,” Journal of Sensors, vol. 2022, Article ID combined with deep learning algorithms,” IEEE Access, 3067840, 12 pages, 2022. vol. 8, Article ID 174105, 2020. [40] A. A. Gharibans, B. L. Smarr, and D. C. Kunkel, “Artifact [26] C. Paul, G. Lefebvre, S. Duﬀner, and C. Garcia, “Learning rejection methodology enables continuous, noninvasive personalized ADL recognition models from few raw data,” measurement of gastric myoelectric activity in ambulatory Artiﬁcial Intelligence in Medicine, vol. 107, Article ID 101916, subjects,” Scientiﬁc Reports, vol. 8, p. 5019, 2018. [41] J. Heo, H. Yoon, and K. S. Park, “A novel wearable forehead [27] E. Casilari, J. A. S. Ramon, ´ and J. M. C. Garc´ıa, “Analysis of EOG measurement system for human computer interfaces,” public datasets for wearable fall detection systems,” Sensors, Sensors, vol. 17, no. 7, 2017. vol. 17, no. 7, 2017. [42] A. Tonacci, L. Billeci, I. Mambro, R. Marangoni, [28] A. Harris, H. True, Z. Hu, J. Cho, N. Fell, and M. Sartipi, “Fall C. Sanmartin, and F. Venturi, “Wearable sensors for recognition using wearable technologies and machine learning assessing the role of olfactory training on the autonomic algorithms,” in Proceedings of the 2016 IEEE International response to olfactory stimulation,” Sensors, vol. 21, no. 3, Conference on Big Data, pp. 3974–3976, (Big Data), Washington, p. 770, 2021. DC, USA, December, 2016. [43] F. Sabry, T. Eltaras, W. Labda, F. Hamza, K. Alzoubi, and [29] K. Liu, C. Hsieh, S. J. Hsu, and C. Chan, “Impact of sampling Q. Malluhi, “Towards on-device dehydration monitoring rate on wearable-based fall detection systems based on using machine learning from wearable device’s data,” Sen- machine learning models,” IEEE Sensors Journal, vol. 18, sors, vol. 22, no. 5, 2022. no. 23, pp. 9882–9890, 2018. [44] S. Park, S. W. Lee, S. Han, and M. Cha, “Clustering insomnia [30] S. B. Kwon, J. Park, C. Kwon, H. J. Kong, J. Y. Hwang, and patterns by data from wearable devices: algorithm devel- H. C. Kim, “An energy-eﬃcient algorithm for classiﬁcation opment and validation study,” JMIR mHealth and uHealth, of fall types using a wearable sensor,” IEEE Access, vol. 7, vol. 7, no. 12, 2019. Article ID 31321, 2019. [45] A. Abedin, F. Motlagh, Q. Shi, R. Hamid, and D. Ranasinghe, [31] A. Collado-Villaverde, M. Cobos, P. Muñoz, and “Towards deep clustering of human activities from wear- D. F. Barrero, “A simulator to support machine learning- ables,” in Proceedings of the 2020 International Symposium based wearable fall detection systems,” Electronics, vol. 9, on Wearable Computers, ISWC ’20, pp. 1–6, Association for no. 11, 2020. Computing Machinery, New York, NY, USA, September, [32] K. Sundararajan, S. Georgievska, B. Te Lindert et al., “Sleep classiﬁcation from wrist-worn accelerometer data using [46] Po-T. Lee, W.-C. Chiu, Y.-H. Ho, Y.-C. Tai, C.-C. K. Lin, and random forests,” Scientiﬁc Reports, vol. 11, no. 24, 2021. C.-L. Lin, “Development of wearable device and clustering [33] A. Cheon, S. Y. Jung, C. Prather, M. Sarmiento, K. Wong, based method for detecting falls in the elderly,” in Pro- and D. M. k. Woodbridge, “A machine learning approach to ceedings of the 2021 IEEE 10th Global Conference on Con- detecting low medication state with wearable technologies,” sumer Electronics (GCCE), pp. 231-232, Kyoto, Japan, in Proceedings of the 42nd Annual International Conference Octobar, 2021. of the IEEE Engineering in Medicine Biology Society, [47] J. Dunn, L. Kidzinski, and R. Runge, “Wearable sensors pp. 4252–4255, 2020. enable personalized predictions of clinical laboratory mea- [34] E. *omaz, I. Essa, and D. Gregory, “Abowd. A practical surements,” Nature Medicine, vol. 27, 2021. approach for recognizing eating moments with wrist- [48] C. Chatzaki, M. Pediaditis, V. George, and M. Tsiknakis, mounted inertial sensing,” in Proceedings of the 2015 ACM “Human daily activity and fall recognition using a smart- International Joint Conference on Pervasive and Ubiquitous Computing, pp. 1029–1040, Osaka, Japan, 2015. phone’s acceleration sensor,” in Proceedings of the Infor- mation and Communication Technologies for Ageing Well [35] A. Bedri, A. Verlekar, E. *omaz, V. Avva, and T. Starner, “A wearable system for detecting eating activities with proximity and e-Health, pp. 100–118, Springer International Publish- sensors in the outer ear,” in Proceedings of the 2015 ACM ing, Cham, Switcher land, June, 2017. Journal of Healthcare Engineering 21 [49] A. Ramachandran, R. Adarsh, P. Pahwa, and K. R. Anupama, [62] P. Resque, A. Barros, D. Rosario, ´ and E. Cerqueira, “An “Machine learning-based techniques for fall detection in investigation of diﬀerent machine learning approaches for epileptic seizure detection,” in Proceedings of the 15th In- geriatric healthcare systems,” in Proceedings of the 2018 9th ternational Wireless Communications Mobile Computing International Conference on Information Technology in Conference (IWCMC), pp. 301–306, Tangier, Morocco, June, Medicine and Education (ITME), pp. 232–237, Hangzhou, China, Octobar, 2018. [63] G. Regalia, F. Onorati, M. Lai, C. Caborni, W. Rosalind, and [50] F. Hussain, F. Hussain, M. Ehatisham-ul-Haq, and Picard, “Multimodal wrist-worn devices for seizure detection M. A. Azam, “Activity-aware fall detection and recognition and advancing research: focus on the Empatica wristbands,” based on wearable sensors,” IEEE Sensors Journal, vol. 19, Epilepsy Research, vol. 153, pp. 79–82, 2019. no. 12, pp. 4528–4536, 2019. [64] C. Caborni, F. Onorati, G. Regalia et al., “Prospective [51] J. D. Lopez, ´ A. Sucerquia, and J. F. V.-B. SisFall, “A fall and evaluation of seizure detection performance of the Embrace movement dataset,” Sensors, vol. 17, no. 1, p. 198, 2017. wristband on pediatric patients in the epilepsy monitoring [52] D. Giuﬀrida, B. Guido, D. De Martini, and T. Facchinetti, unit,” in Proceedings of the 2nd International Congress on “Fall detection with supervised machine learning using Mobile Devices & Seizure Detection in Epilepsy, Lausanne, wearable sensors,” in Proceedings of the 2019 IEEE 17th Switzer land, September, 2019. International Conference on Industrial Informatics (INDIN), [65] F. Onorati, G. Regalia, C. Caborni et al., “Multicenter Clinical vol. 1, pp. 253–259, Helsinki, Finland, July, 2019. Assessment of Improved Wearable Multimodal Convulsive [53] R. Saeedi, K. Sasani, S. Norgaard, and A. H. Gebremedhin, Seizure Detectors,” Epilepsia, vol. 58, 2017. “Personalized human activity recognition using wearables: a [66] G. Yang, J. Deng, G. Pang et al., “An IoT-enabled stroke manifold learning-based knowledge transfer,” in Proceedings rehabilitation system based on smart wearable armband and of the 2018 40th Annual International Conference of the IEEE machine learning,” IEEE Journal of Translational Engineering Engineering in Medicine and Biology Society (EMBC), in Health and Medicine, vol. 6, no. 1–10, 2018. pp. 1193–1196, Honolulu, HI, USA, July, 2018. [67] S. Potluri, A. B. Chandran, C. Diedrich, and L. Schega, [54] X. Li, Z. Zhou, J. Wu, and Y. Xiong, “Human posture de- “Machine learning based human gait segmentation with tection method based on wearable devices,” Journal of wearable sensor platform,” in Proceedings of the 2019 41st Healthcare Engineering, vol. 2021, Article ID 8879061, Annual International Conference of the IEEE Engineering in 8 pages, 2021. Medicine and Biology Society (EMBC), pp. 588–594, Berlin, [55] A. Bedri, R. Li, M. Haynes et al., “EarBit: using wearable Germany, July, 2019. sensors to detect eating episodes in unconstrained envi- [68] H. Zhang, Y. Guo, and D. Zanotto, “Accurate ambulatory ronments,” Proceedings of the ACM on interactive, mobile, gait analysis in walking and running using machine learning wearable and ubiquitous technologies, vol. 1, no. 3, p. 37, models,” IEEE Transactions on Neural Systems and Reha- bilitation Engineering, vol. 28, no. 1, pp. 191–202, 2020. [56] K. San Chun, H. Jeong, R. Adaimi, and E. *omaz, “Eating [69] S. R. Moore, C. Kranzinger, J. Fritz, T. Stggl, J. Krll, and episode detection with jawbone-mounted inertial sensing,” H. Schwameder, “Foot strike angle prediction and pattern in Proceedings of the 2020 42nd Annual International Con- classiﬁcation using LoadsolTM wearable sensors: a com- ference of the IEEE Engineering in Medicine Biology Society parison of machine learning techniques,” Sensors, vol. 20, (EMBC), pp. 4361–4364, Montreal, QC, Canada, July, 2020. no. 23, 2020. [57] D. Morris, S. Scott, A. Guillory, and I. Kelner, “RecoFit: using [70] Y. Mengistu, M. Pham, M. Do, and W. Sheng, “Autohydrate: a wearable sensor to ﬁnd, recognize, and count repetitive A Wearable Hydration Monitoring System,” in Proceedings exercises,” in Proceedings of the SIGCHI Conference on of the 2016 IEEE/RSJ International Conference on Intelligent Human Factors in Computing Systems, pp. 3225–3234, ACM, Robots and Systems (IROS), pp. 1857–1862, Daejeon, Korea Toranto, Canada, April, 2014. (South), Octobar, 2016. [58] A. J. Masino, D. Forsyth, H. Nuske et al., “m-Health and [71] A. Alvarez, E. Severeyn, J. Velasquez, ´ S. Wong, G. Perpiñan, Autism: recognizing stress and anxiety with machine and M. . Huerta, “Machine learning methods in the classi- learning and wearables data,” in Proceedings of the 2019 IEEE ﬁcation of the athletes dehydration,” in Proceedings of the 32nd International Symposium on Computer-Based Medical 2019 IEEE Fourth Ecuador Technical Chapters Meeting Systems (CBMS), pp. 714–719, Cordoba, Spain, June, 2019. (ETCM), pp. 1–5, Guayaquil, Ecuador, November, 2019. [59] Q. Zhang, X. Zeng, W. Hu, and D. Zhou, “A machine [72] S. Koelstra, C. Muehl, M. Soleymani et al., Deap, A database learning-empowered system for long-term motion-tolerant for emotion analysis using physiological signals,” IEEE wearable monitoring of blood pressure and heart rate with Transactions on Aﬀective Computing, vol. 3, no. 1, pp. 18–31, ear-ECG/PPG,” IEEE Access, vol. 5, Article ID 10547, 2017. [60] S. Tuli, N. Basumatary, S. S. Gill et al., “HealthFog: an en- [73] R. Laureanti, M. Bilucaglia, M. Zito et al., “Emotion as- semble deep learning based smart healthcare system for sessment using Machine Learning and low-cost wearable automatic diagnosis of heart diseases in integrated IoT and devices,” in Proceedings of the 42nd Annual International fog computing environments,” Future Generation Computer Conference of the IEEE Engineering in Medicine Biology Systems, vol. 104, pp. 187–200, 2020. Society (EMBC), pp. 576–579, Montreal, QC, Canada, July, [61] Y. Shen, M. Voisin, A. Aliamiri, A. Anand, A. Hannun, and A. Ng, “Ambulatory atrial ﬁbrillation monitoring using [74] D. Ayata, Y. Yaslan, and M. E. Kamasak, “Emotion recog- wearable photoplethysmography with deep learning,” in nition from multimodal physiological signals for emotion Proceedings of the 25th ACM SIGKDD International Con- aware healthcare systems,” Journal of Medical and Biological ference on Knowledge Discovery Data Mining, KDD ’19, Engineering, vol. 40, pp. 149–157, 2020. pp. 1909–1916, Association for Computing Machinery, New [75] F. Miao, Bo Wen, Z. Hu et al., “Continuous blood pressure York, NY, USA, July, 2019. measurement from one-channel electrocardiogram signal 22 Journal of Healthcare Engineering using deep-learning techniques,” Artiﬁcial Intelligence in [90] O. Walch, “Motion and Heart Rate from a Wrist-Worn Medicine, vol. 108, Article ID 101919, 2020. Wearable and Labeled Sleep from Polysomnography,” [76] C. K. Wong, D. T. Y. Ho, A. R. Tam, M. Zhou, Y. M. Lau, and PhysioNet, vol. 101, 2019. [91] O. Walch, Y. Huang, D. Forger, and C. Goldstein, “Sleep M. O. Y. Tang, “Artiﬁcial intelligence mobile health platform stage prediction with raw acceleration and photo- for early detection of COVID-19 in quarantine subjects using plethysmography heart rate data derived from a consumer a wearable biosensor: protocol for a randomised controlled wearable device,” Sleep, vol. 42, no. 12, p. 08, 2019. trial,” BMJ Open, vol. 10, 2020. [92] D. Houfani, S. Slatnia, O. Kazar, H. Saouli, and A. Merizig, [77] S. Rastogi and J. Singh, “A Systematic Review on Machine “Artiﬁcial intelligence in healthcare: a review on predicting Learning for Fall Detection System,” Computational Intel- clinical needs,” International Journal of Healthcare Man- ligence, vol. 37, 2021. agement, vol. 0, no. 0, pp. 1–9, 2021. [78] M. Kilany, A. E. Hassanien, and A. Badr, “Accelerometer- [93] A. Jain, D. Way, V. Gupta, Yi Gao, G. O. Marinho, and based human activity classiﬁcation using Water Wave Op- J. Hartford, “Development and assessment of an artiﬁcial timization approach,” in Proceedings of the 2015 11th In- intelligence–based tool for skin condition diagnosis by ternational Computer Engineering Conference (ICENCO), primary care physicians and nurse practitioners in tele- pp. 175–180, Cario, Egypt, December, 2015. dermatology practices,” JAMA Network Open, vol. 4, no. 4, [79] C. Min, A. Mathur, and F. Kawsar, “Exploring audio and Article ID e217249, 2021. kinetic sensing on earable devices,” in Proceedings of the 4th [94] J. Greiwe and S. M. Nyenhuis, “Wearable technology and ACM Workshop on Wearable Systems and Applications, how this can Be implemented into clinical practice,” Current WearSys ’18, pp. 5–10, Association for Computing Ma- Allergy and Asthma Reports, vol. 20, no. 8, p. 36, 2020. chinery, New York, NY, USA, June, 2018. [95] T. Chitnis, B. I. Glanz, and C. Gonzalez, “Quantifying [80] B. M. Bell, R. Alam, and N. Alshurafa, “Automatic, wearable- neurologic disease using biosensor measurements in-clinic based, in-ﬁeld eating detection approaches for public health and in free-living settings in multiple sclerosis,” Npj Digital research: a scoping review,” Npj Digital Medicine, vol. 3, Medicine, vol. 2, no. 123, 2019. no. 38, 2020. [96] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes- [81] S. Bae, D. Ferreira, B. Suﬀoletto et al., “Detecting drinking Ortiz, “A Public Domain Dataset for Human Activity episodes in young adults using smartphone-based sensors,” Recognition Using Smartphones,” in Proceedings of the 21st Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.vol.1, European Symposium on Artiﬁcial Neural Networks, Com- no. 2, 2017. putational Intelligence And Machine Learning, Bruges, [82] M. H. Imtiaz, R. I. Ramos-Garcia, S. Wattal, T. Stephen, and Belgium, April 2013. E. Sazonov, “Wearable sensors for monitoring of cigarette [97] D. Garcia-Gonzalez, D. Rivero, E. Fernandez-Blanco, and smoking in free-living: a systematic review,” Sensors, vol. 19, M. R. Luaces, “A public domain dataset for real-life human no. 21, 2019. activity recognition using smartphone sensors,” Sensors, [83] Y. Choi, Yu-Mi Jeon, L. Wang, and K. Kim, “A biological vol. 20, no. 8, 2020. signal-based stress monitoring framework for children using [98] R. Meziatisabour, Y. Benezeth, P. De Oliveira, Julien Chappe, wearable devices,” Sensors, vol. 17, no. 9, 2017. and Y. Fan, “UBFC-phys: a multimodal database for psy- [84] A. Greco, G. Valenza, A. Lanata, G. Rota, and E. P. Scilingo, chophysiological studies of social stress,” IEEE Transactions “Electrodermal activity in bipolar patients during aﬀective on Aﬀective Computing, p. 1, 2021. elicitation,” IEEE Journal of Biomedical and Health Infor- [99] V. Markova, T. Ganchev, and K. Kalinkov, “A database for matics, vol. 18, no. 6, pp. 1865–1873, 2014. cognitive load, aﬀect and stress recognition,” in Pro- [85] E. Garcia-Ceja, M. Riegler, T. Nordgreen, P. Jakobsen, ceedings of the First International Conference on Bio- K. J. Oedegaard, and Jim Tørresen, “Mental health moni- medical Innovations and Applications, Article ID 8967457, toring with multimodal sensing and machine learning: a Varna, Bulgaria, 2019. survey,” Pervasive and Mobile Computing, vol. 51, pp. 1–26, [100] G. Nguyen, S. Dlugolinsky, and M. Bobak, “Machine learning and deep learning frameworks and libraries for [86] T. Rukasha, S. I. Woolley, T. Kyriacou, and T. Collins, large-scale data mining: a survey,” Artiﬁcial Intelligence “Evaluation of wearable electronics for epilepsy: a systematic Review, vol. 52, 2019. review,” Electronics, vol. 9, no. 6, 2020. [101] P. C. Chen, Y. Liu, and L. Peng, “How to develop machine [87] J. C. Alcaraz, S. Moghaddamnia, N. Poschadel, and J. Peissig, learning models for healthcare,” Nature Materials, vol. 18, “Machine learning as digital therapy assessment for mobile no. 5, pp. 410–414, 2019. gait rehabilitation,” in Proceedings of the 2018 IEEE 28th [102] B. Bent, B. Goldstein, K. Warren, and J. Dunn, “Investigating International Workshop on Machine Learning for Signal sources of inaccuracy in wearable optical heart rate sensors,” Processing (MLSP), pp. 1–6, Aalborg, Denmark, September, Npj Digital Medicine, vol. 3, no. 18, 2020. [103] B. A. Mateen, J. Liley, and A. K. Denniston, “Improving the [88] R. Ghaﬀari, J. A. Rogers, and T. R. Ray, “Recent progress, quality of machine learning in health applications and challenges, and opportunities for wearable biochemical clinical research,” Nature Machine Intelligence, vol. 2, sensors for sweat analysis,” Sensors and Actuators B: pp. 554–556, 2020. Chemical, vol. 332, Article ID 129447, 2021. [104] A. S. Keerthy and S. Manju Priya, Artiﬁcial Intelligence in [89] O. Blan, G. Moise, A. Moldoveanu, M. Leordeanu, and Healthcare Databases, Springer International Publishing, F. Moldoveanu, “An investigation of various machine and Cham, Switcher land, pp. 19–34, 2021. deep learning techniques applied in automatic fear level [105] B. Nelson, C. Low, N. Jacobson, P. Arean, J. Torous, and detection and acrophobia virtual therapy,” Sensors, vol. 20, N. Allen, “Guidelines for wrist-worn wearable assessment of no. 2, 2020. heart rate in biobehavioral research,” Nature, vol. 3, 2020. Journal of Healthcare Engineering 23 [106] S. Saeb, A. Jayaraman, D. C. Mohr, K. P. Kording, and [122] R. Simon Sherratt, B. Janko, T. Hui et al., “Task scheduling to L. Lonini, “*e need to approximatee the use-case in clinical constrain peak current consumption in wearable healthcare machine learning,” GigaScience, vol. 6, no. 5, pp. 1–9, 2017. sensors,” Electronics, vol. 8, no. 7, 2019. [107] M. A. Little, G. Varoquaux, S. Saeb et al., “Using and un- [123] Alma Secerbegovic, A. Gogic, N. Suljanovic, M. Zajc, and derstanding cross-validation strategies. Perspectives on Saeb A. Mujcic, “Computational balancing between wearable etal,” GigaScience, vol. 6, no. 5, p. 03, 2017. sensor and smartphone towards energy-eﬃcient remote [108] A. B. Arrieta, N. D´ıaz-Rodr´ıguez, A. Bennetot et al., “Ex- healthcare monitoring,” Advances in Electrical and Computer plainable Artiﬁcial Intelligence (XAI): concepts, taxonomies, Engineering, vol. 18, no. 3–10, p. 11, 2018. opportunities and challenges toward responsible AI,” In- [124] M. Antonio Scrugli, D. Loi, L. Raﬀo, and P. Meloni, “A formation Fusion, vol. 58, pp. 82–115, 2020. runtime-adaptive cognitive IoT node for healthcare moni- [109] N. Mairittha, T. Mairittha, and S. Inoue, “On-device deep toring,” in Proceedings of the 16th ACM International personalization for robust activity data collection,” Sensors, Conference on Computing Frontiers, CF ’19, pp. 350–357, vol. 21, 2020. Association for Computing Machinery, New York, NY, USA, [110] Z. Fu, X. He, E. Wang, J. Huo, J. Huang, and D. Wu, May, 2019. “Personalized human activity recognition based on inte- [125] T. Takao, “Design consideration of sustainable self-suﬃcient grated wearable sensor and transfer learning,” Sensors, energy system using ferroelectric material for wearable vol. 21, no. 3, 2021. healthcare,” International Journal of Energy Research, vol. 40, [111] A. Ferrari, D. Micucci, M. Mobilio, and P. Napoletano, no. 15, pp. 2176–2186, 2016. “Deep learning and model personalization in sensor-based [126] J. Yuan and R. Zhu, “A fully self-powered wearable moni- human activity recognition,” Journal of Reliable Intelligent toring system with systematically optimized ﬂexible ther- Environments, vol. 54, 2022. moelectric generator,” Applied Energy, vol. 271, Article ID [112] L. Sun, X. Jiang, H. Ren, and Y. Guo, “Edge-cloud computing 115250, 2020. and artiﬁcial intelligence in internet of medical things: ar- [127] X. Fafoutis, L. Marchegiani, A. Elsts, J. Pope, R. Piechocki, chitecture, technology and application,” IEEE Access, vol. 8, and I. Craddock, “Extending the Battery Lifetime of Article ID 101079, 2020. Wearable Sensors with Embedded Machine Learning,” in [113] F. A. Kraemer, A. E. Braten, N. Tamkittikhun, and D. Palma, Proceedings of the IEEE 4th World Forum on Internet of “Fog computing in healthcare – a review and discussion,” Gings (WF-IoT), pp. 269–274, Singapore, Febrauary, 2018. IEEE Access, vol. 5, pp. 9206–9222, 2017. [128] D. Jaiswal, A. Gigie, T. Chakravarty, A. Ghose, and A. Misra, [114] D. A. Tamburri, “Sustainable MLOps: trends and chal- “Table of interest: activity recognition and behaviour analysis lenges,” in Proceedings of the 2020 22nd International using a battery less wearable sensor,” in Proceedings of the 4th Symposium on Symbolic and Numeric Algorithms for Sci- ACM Workshop on Wearable Systems and Applications, entiﬁc Computing (SYNASC), pp. 17–23, Timisoara, Roma- WearSys ’18, pp. 39–44, Association for Computing Ma- nia, September, 2020. chinery, New York, NY, USA, June, 2018. [115] J. Waring, C. Lindvall, and R. Umeton, “Automated machine [129] L. Klus, E. S. Lohan, C. Granell, and J. Nurmi, “Lossy Com- learning: review of the state-of-the-art and opportunities for pression Methods for Performance-Restricted Wearable healthcare,” Artiﬁcial Intelligence in Medicine, vol. 104, Devices,” in Proceedings of the 2020 International Conference Article ID 101822, 2020. on Localization and GNSS (ICL-GNSS), Tampere, Finland, [116] J. K. Kim, H. Jung, G. B. Hwang, O. S. Gwon, and S. E. Lee, June, 2020, https://doi.org/10.5281/zenodo.3908174. “Design of low-power SoC for wearable healthcare device,” [130] S. H. Baek and K. W. Park, “A Durable Hybrid RAM Disk Journal of Circuits, Systems, and Computers, vol. 29, no. 06, with a Rapid Resilience for Sustainable IoT Devices,” Sensors, Article ID 2050085, 2020. vol. 20, 2020. [117] E. Torti, C. D’Amato, Giovanni Danese, and F. Leporati, “An [131] J. Huang, A. Badam, and R. Chandra, Nightingale, Wear- hardware recurrent neural network for wearable devices,” in Drive: fast and energy-eﬃcient storage for wearables,” in Proceedings of the 2020 23rd Euromicro Conference on Digital Proceedings of the Annual Technical Conference (ATC’15). System Design (DSD), pp. 293–300, Kranj, Slovenia, August, USENIX - Advanced Computing Systems Association, Best 2020. Paper Award, Berkeley, CA, USA, July, 2015. [118] M. Faustino, J. Calado, J. Sarraipa, and R. Jardim-Gonçalves, [132] A. Kumar, S. Goyal, and M. Varma, “Resource-eﬃcient “Adaptable power consumption proﬁles for wearable lo- machine learning in 2 KB RAM for the internet of things,” in calization devices,” in Proceedings of the SEIA’2019 Con- Proceedings of the 34th International Conference on Machine ference Proceedings, p. 200, Tenerife, Canary, September, Learning - Volumevol. 70, pp. 1935–1944, JMLR.org, 2017. 2022. [133] M. Sek, “Chai. Quantization-guided training for compact [119] L. Zhang, J. Hu, C. Guo, and H. Xu, “Dynamic power op- TinyML models,” in Proceedings of the Research Symposium timization for secondary wearable biosensors in e-healthcare on Tiny Machine Learning, San Francisco, USA, March, 2021. leveraging cognitive WBSNs with imperfect spectrum [134] M. Merenda, C. Porcaro, and D. Iero, “Edge machine sensing,” Future Generation Computer Systems, vol. 112, learning for AI-enabled IoT devices: a review,” Sensors, pp. 67–92, 2020. vol. 20, no. 9, 2020. [120] M. Lewandowski, B. O. Płaczek, and M. Bernas, “Classiﬁer- [135] V. T. Tran, C. Riveros, and P. Ravaud, “Patients’ views of Based Data Transmission Reduction in Wearable Sensor wearable devices and AI in healthcare: ﬁndings from the Network for Human Activity Monitoring,” Sensors, vol. 21, ComPaRe e-cohort,” Npj Digital Medicine, vol. 2, 2019. no. 1, 2021. [136] P. Esmaeilzadeh, “Use of AI-based tools for healthcare [121] J. Zhou and C. Wang, “An ultra-low power turning angle purposes: a survey study from consumers’ perspectives,” based biomedical signal compression engine with adaptive BMC Medical Informatics and Decision Making, vol. 20, threshold tuning,” Sensors, vol. 17, p. 1809, 2017. p. 170, 2020. 24 Journal of Healthcare Engineering [137] Q. Yue, Z. Ling, X. Fu, B. Liu, W. Yu, and W. Zhao, “My [153] V. M. Suriyakumar, N. Papernot, A. Goldenberg, and Google Glass Sees Your Passwords! Black Hat USA,” Sensors, M. Ghassemi, “Challenges of diﬀerentially private prediction in healthcare settings,” in Proceedings of the IJCAI 2021 [138] D. Weiz, G. Anand, and C. P. H. Ernst, Ge Inﬂuence of Workshop on AI for Social Good, Montreal, Canada, 2021. [154] R. L. Rivest, L. Adleman, and M. L. Dertouzos, “On data Subjective Norm on the Usage of Smartglasses, Springer, banks and privacy homomorphisms,” Foundations of secure Cham, Switzerland. [139] N. T. Sahin, N. U. Keshav, J. P. Salisbury, and computation, vol. 4, no. 11, pp. 169–180, 1978. [155] A. Wood, K. Najarian, and D. Kahrobaei, “Homomorphic A. Vahabzadeh, “Second version of google glass as a wearable encryption for machine learning in medicine and bio- socio-aﬀective aid: positive school desirability, high usability, informatics,” ACM Computing Surveys, vol. 53, no. 4, 2020. and theoretical framework in a sample of children with [156] M. Kim, Y. Song, S. Wang, Y. Xia, and X. Jiang, “Secure autism,” JMIR human factors, vol. 5, no. 1, p. 1, 2018. logistic regression based on homomorphic encryption: de- [140] K. L. LaBat and K. S. Ryan, Human Body A Wearable Product sign and evaluation,” JMIR medical informatics, vol. 6, no. 2, Designer’s Guide, CRC Press, London, UK, 1 edition, 2019. p. e19, 2018. [141] T. Zhang, J. Lu, F. Hu, and H. Qi, “Bluetooth low energy for [157] Y. Aono, T. Hayashi, L. T. Phong, and L. Wang, “Privacy- wearable sensor-based healthcare systems,” in Proceedings of preserving logistic regression with distributed data sources the2014 IEEE Healthcare Innovation Conference (HIC), via homomorphic encryption,” IEICE - Transactions on Info pp. 251–254, Seattle, USA, September 2014. and Systems, vol. 99, no. 8, pp. 2079–2089, 2016. [142] C. Balakrishna, E. Rendon-Morales, R. Aviles-Espinosa, [158] H. Chen, R. Gilad-Bachrach, K. Han et al., “Logistic re- D. Henry, and Z. Luo, “Challenges of wearable health gression over encrypted data from fully homomorphic en- monitors: a case study of foetal ECG monitor,” in Pro- cryption,” BMC Medical Genomics, vol. 11, no. 4, pp. 3–12, ceedings of the 2019 global IoT summit (GIoTS), pp. 1–6, Aarhus, USA, June 2019. [159] Y. Aono, T. Hayashi, L. T. Phong, and L. Wang, “Scalable and [143] E. Adi, A. Anwar, and Z. Baig, “Machine learning and data secure logistic regression via homomorphic encryption,” in analytics for the IoT,” Neural Computing & Applications, Proceedings of the Sixth ACM Conference on Data and Ap- vol. 32, Article ID 16205, 2020. plication Security and Privacy, pp. 142–144, Washington, [144] F. Alsubaei, A. Abuhussein, and S. Shiva, “Security and USA, April 2016. privacy in the internet of medical things: taxonomy and risk [160] J. L. H. Crawford, G. Craig, S. Halevi, D. Platt, and V. Shoup, assessment,” in Proceedings of the 2017 IEEE 42nd Conference “Doing real work with FHE: the case of logistic regression,” on Local Computer Networks Workshops (LCN Workshops), in Proceedings of the 6th Workshop on Encrypted Computing pp. 112–120, Singapore, Octobar, 2017. & Applied Homomorphic Cryptography, pp. 1–12, Toranto, [145] J. Padgette, J. Bahr, M. Batra et al., “NIST Special Publication Canada, October 2018. 800-121 Revision 2,” Guide to BluetoothSecurity. [161] S. Carpov, N. Gama, M. Georgieva, and J. R. T. Pastoriza, [146] L. Xiao, X. Wan, X. Lu, Y. Zhang, and Di Wu, “IoT security “Privacy-preserving semi-parallel logistic regression training techniques based on machine learning: how do IoT devices with fully homomorphic encryption,” BMC Medical Geno- use AI to enhance security?” IEEE Signal Processing Maga- mics, vol. 13, no. 7, pp. 1–10, 2020. zine, vol. 35, no. 5, pp. 41–49, 2018. [162] S. Kim, M. Omori, T. Hayashi, T. Omori, L. Wang, and [147] R. N. Reith, T. Schneider, and O. Tkachenko, “Eﬃciently S. Ozawa, “Privacy-preserving naive bayes classiﬁcation stealing your machine learning models,” in Proceedings of the using fully homomorphic encryption,” in Proceedings of the 18th ACM Workshop on Privacy in the Electronic Society, International Conference on Neural Information Processing, WPES’19, pp. 198–210, Association for Computing Ma- pp. 349–358, Springer, Siem Reap, Cambodia, December chinery, New York, NY, USA, November 2019. [148] K. Hansel, ¨ K. Katevas, O. Guido, D. C. Richardson, [163] Y. Yasumura, I. Yu, and H. Yamana, “Secure Na¨ıve Bayes A. Alomainy, and H. Haddadi, “*e potential of wearable classiﬁcation protocol over encrypted data using fully ho- technology for monitoring social interactions based on in- momorphic encryption,” in Proceedings of the 21st Inter- terpersonal synchrony,” in Proceedings of the 4th ACM national Conference on Information Integration and Web- Workshop on Wearable Systems and Applications, WearSys based, pp. 45–54, Applications & Services, Munich, Ger- ’18, pp. 45–47, Association for Computing Machinery, New many, December 2019. York, NY, USA, November 2018. [164] Li Jing, X. Kuang, S. Lin, Ma Xu, and Yi Tang, “Privacy [149] B. Amine, A. Derhab, and Y. Challal, “A review of privacy- preservation for machine learning training and classiﬁcation preserving techniques for deep learning,” Neurocomputing, based on homomorphic encryption schemes,” Information vol. 384, pp. 21–45, 2020. Sciences, vol. 526, pp. 166–179, 2020. [150] A. Qayyum, J. Qadir, M. Bilal, and A. Al-Fuqaha, “Secure and [165] H. Jung, D. Kim, and J. H. Park, “Towards a practical cluster robust machine learning for healthcare: a survey,” IEEE analysis over encrypted data,” in Proceedings of the Inter- Reviews in Biomedical Engineering, vol. 14, pp. 156–180, 2021. national Conference on Selected Areas in Cryptography, [151] C. Dwork, “Diﬀerential privacy,” in Proceedings of the 33rd pp. 227–249, Springer, Waterloo, Canada, August 2019. International Colloquium on Automata, Languages and [166] V. Costan and S. Devadas, “Intel SGX explained,” IACR Programming, part II (ICALP 2006), pp. 1–12, Springer- Cryptol. ePrint Arch.vol. 2016, no. 86, pp. 1–118, 2016. Verlag, Venice, Italy, July 2006. [167] T. Hunt, C. Song, R. Shokri, V. Shmatikov, and E. Witchel, [152] B. Jayaraman and D. Evans, “Evaluating diﬀerentially private “Chiron: privacy-preserving machine learning as a service,” machine learning in practice,” in Proceedings of the 28th 2018. USENIX Security Symposium (USENIX Security 19), [168] C. Segarra, R. Delgado-Gonzalo, M. Lemay, P.-L. Aublin, pp. 1895–1912, USENIX Association, Santa Clara, CA, USA, P. Peter, and V. Schiavoni, “Using trusted execution envi- August 2019. ronments for secure stream processing of medical data,” in Journal of Healthcare Engineering 25 Proceedings of the IFIP International Conference on Dis- ’19, pp. 41–46, Association for Computing Machinery, New York, NY, USA, June, 2019. tributed Applications and Interoperable Systems, pp. 91–107, Springer, Kongens Lyngby, Denmark, June 2019. [186] D. Liaqat, R. Wu, A. Gershon, H. Alshaer, R. Frank, and E. de Lara, “Challenges with real-world smartwatch based [169] M. Nassar, A. Erradi, F. Sabri, and Q. M. Malluhi, “Secure Outsourcing of Matrix Operations as a Service,” in Pro- audio monitoring,” in Proceedings of the 4th ACM Workshop on Wearable Systems and Applications, WearSys ’18, ceedings of the 2013 IEEE Sixth International Conference on pp. 54–59, Association for Computing Machinery, New Cloud Computing, pp. 918–925, Santa Clara, CA, USA, June York, NY, USA, November, 2018. [187] M. R. Nakhkash, T. N. Gia, I. Azimi, A. Anzanpour, [170] R. Lazzeretti and M. Barni, “Private computing with garbled A. M. Rahmani, and P. Liljeberg, “Analysis of performance circuits [applications corner],” IEEE Signal Processing and energy consumption of wearable devices and mobile Magazine, vol. 30, no. 2, pp. 123–127, 2013. gateways in IoT applications,” in Proceedings of the Inter- [171] Y. Wang, Q. M. Malluhi, and K. M. D. Khan, “Garbled national Conference on Omni-Layer Intelligent Systems, computation in cloud,” Future Generation Computer Sys- COINS ’19, pp. 68–73, Association for Computing Ma- tems, vol. 62, pp. 54–65, 2016. chinery, New York, NY, USA, May, 2019. [172] Adi Shamir, “How to share a secret,” Communications of the ACM, vol. 22, no. 11, pp. 612-613, 1979. [173] M. Kesarwani, B. Mukhoty, V. Arya, and S. Mehta, “Model extraction warning in MLaaS paradigm,” in Proceedings of the 34th Annual Computer Security Applications Conference, pp. 371–380, San Juan, PR, USA, December 2018. [174] H. Brendan McMahan, E. Moore, D. Ramage, and y A. Blaise Aguera, ¨ “Federated Learning of Deep Networks Using Model Averaging,” CoRR, abs/1602, Article ID 05629, 2016. [175] K. Peter, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, and A. N. Bhagoji, Advances and Open Problems in Federated Learning, 2019, Article ID 04977. [176] Y. S. Can and C. Ersoy, Privacy-Preserving Federated Deep Learning for Wearable IoT-Based Biomedical Monitoring, vol. 21, no. 1, 2021. [177] J. Xu, B. S. Glicksberg, and C. Su, “Federated learning for healthcare informatics,” J Healthc Inform Res, vol. 5, no. 1–19, 2021. [178] S. Truex, N. Baracaldo, A. Ali et al., “A hybrid approach to privacy-preserving federated learning,” in Proceedings of the 12th ACM Workshop on Artiﬁcial Intelligence and Security, pp. 1–11, Association for Computing Machinery, New York, NY, USA, November, 2019. [179] Y. Chen, X. Qin, J. Wang, C. Yu, and W. Gao, “FedHealth: a federated transfer learning framework for wearable health- care,” IEEE Intelligent Systems, vol. 35, no. 4, pp. 83–93, 2020. [180] D. Batista, P. Hugo, A. Fred, C. Moreira, M. Reis, and H. A. Ferreira, “Benchmarking of the BITalino biomedical toolkit against an established gold standard,” Healthcare Technology Letters, vol. 6, no. 32–36, 2019. [181] M. Huang, D. Chen, and X. Fan, “An eﬀective adaptive ﬁlter to reduce motion artifacts from ECG signals using accel- erometer,” in Proceedings of the 2019 9th International Conference on Biomedical Engineering and Technology, pp. 83–88, Association for Computing Machinery, New York, NY, USA, March 2019. [182] D. Pollreisz and N. Taherinejad, “Detection and Removal of Motion Artifacts in PPG Signals,” Mobile Networks and Applications, vol. 24, 2019. [183] L. M. Stevens, B. J. Mortazavi, R. C. Deo, L. Curtis, and D. P. Kao, “Recommendations for reporting machine learning analyses in clinical research,” Circulation: Cardio- vascular Quality and Outcomes, vol. 13, no. 10, Article ID e006556, 2020. [184] J. W. K. S. Moon, S.U. Kang, and B. Jang, “Eﬀective privacy- preserving collection of health data from a user’s wearable device,” Applied Sciences, vol. 10, no. 18, 2020. [185] C. Kang, H. Jung, and Y. Lee, “Towards machine learning with zero real-world data,” in Proceedings of the Ge 5th ACM Workshop on Wearable Systems and Applications, WearSys http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Healthcare Engineering Hindawi Publishing Corporation http://www.deepdyve.com/lp/hindawi-publishing-corporation/machine-learning-for-healthcare-wearable-devices-the-big-picture-3ZYixxb1Lr

Loading next page...

References (197)

Max Little, G. Varoquaux, Sohrab Saeb, L. Lonini, A. Jayaraman, D. Mohr, Konrad Kording (2017)
Using and understanding cross-validation strategies. Perspectives on Saeb et al.
GigaScience, 6
Stacey Truex, N. Baracaldo, Ali Anwar, T. Steinke, Heiko Ludwig, Rui Zhang (2018)
A Hybrid Approach to Privacy-Preserving Federated Learning
Informatik Spektrum, 42
Adaptable power consumption profiles for wearable localization devices
K. Chun, Hyoyoung Jeong, Rebecca Adaimi, Edison Thomaz (2020)
Eating Episode Detection with Jawbone-Mounted Inertial Sensing
2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)
Carlos Segarra, R. Delgado-Gonzalo, M. Lemay, Pierre-Louis Aublin, P. Pietzuch, V. Schiavoni (2019)
Using Trusted Execution Environments for Secure Stream Processing of Medical Data - (Case Study Paper)
ArXiv, abs/1906.07072
J. Hong S. Kwon
Detection of atrial fibrillation using a ring-type wearable device (CardioTracker) and deep learning analysis of photoplethysmography signals: prospective observational proof-of-concept,
Study J Med Internet Res, 22
Ashish Kumar, Saurabh Goyal, M. Varma (2017)
Resource-efficient Machine Learning in 2 KB RAM for the Internet of Things
M. Perez, K. Mahaffey, H. Hedlin, J. Rumsfeld, Ariadna Garcia, Todd Ferris, V. Balasubramanian, A. Russo, Amol Rajmane, L. Cheung, Grace Hung, Justin Lee, P. Kowey, N. Talati, Divya Nag, Santosh Gummidipundi, A. Beatty, M. Hills, S. Desai, C. Granger, M. Desai, M. Turakhia (2019)
Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation.
The New England journal of medicine, 381 20
A. Shamir (1979)
How to share a secret
Commun. ACM, 22
Jack Crawford, Craig Gentry, S. Halevi, Daniel Platt, V. Shoup (2018)
Doing Real Work with FHE: The Case of Logistic Regression
Proceedings of the 6th Workshop on Encrypted Computing & Applied Homomorphic Cryptography
M. Kesarwani, B. Mukhoty, V. Arya, S. Mehta (2017)
Model Extraction Warning in MLaaS Paradigm
Proceedings of the 34th Annual Computer Security Applications Conference
J. Dunn, L. Kidzinski, Ryan Runge, Daniel Witt, J. Hicks, S. Rose, Xiao Li, Amir Bahmani, S. Delp, T. Hastie, M. Snyder (2021)
Wearable sensors enable personalized predictions of clinical laboratory measurements
Nature Medicine, 27
Sohrab Saeb, L. Lonini, A. Jayaraman, D. Mohr, Konrad Kording (2017)
The need to approximate the use-case in clinical machine learning
GigaScience, 6
Denisse Castaneda, Aibhlin Esparza, M. Ghamari, C. Soltanpur, H. Nazeran (2018)
A review on wearable photoplethysmography sensors and their potential future applications in health care
International journal of biosensors & bioelectronics, 4
Jie Xu, B. Glicksberg, Chang Su, Peter Walker, J. Bian, Fei Wang (2019)
Federated Learning for Healthcare Informatics
Journal of Healthcare Informatics Research, 5
(2019)
Body AWearable Product Designer’s Guide, CRC Press, London, UK, 1 edition, 2019
Awni Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, G. Tison, Codie Bourn, M. Turakhia, A. Ng (2019)
Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network
Nature Medicine, 25
Ned Sahin, Neha Keshav, Joseph Salisbury, Arshya Vahabzadeh (2018)
Second Version of Google Glass as a Wearable Socio-Affective Aid: Positive School Desirability, High Usability, and Theoretical Framework in a Sample of Children with Autism
JMIR Human Factors, 5
Paulo Resque, Alex Barros, D. Rosário, E. Cerqueira (2019)
An Investigation of Different Machine Learning Approaches for Epileptic Seizure Detection
2019 15th International Wireless Communications & Mobile Computing Conference (IWCMC)
Jonathan Waring, C. Lindvall, R. Umeton (2020)
Automated machine learning: Review of the state-of-the-art and opportunities for healthcare
Artificial intelligence in medicine, 104
V. Markova, T. Ganchev, Kalin Kalinkov (2019)
CLAS: A Database for Cognitive Load, Affect and Stress Recognition
2019 International Conference on Biomedical Innovations and Applications (BIA)
J. Alcaraz, S. Moghaddamnia, Nils Poschadel, J. Peissig (2018)
Machine Learning as Digital Therapy Assessment for Mobile Gait Rehabilitation
2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP)
Prospective evaluation of seizure detection performance of the Embrace wristband on pediatric patients in the epilepsy monitoring unit
Brinnae Bent, Benjamin Goldstein, W. Kibbe, J. Dunn (2020)
Investigating sources of inaccuracy in wearable optical heart rate sensors
NPJ Digital Medicine, 3
Yoshinori Aono, Takuya Hayashi, L. Phong, Lihua Wang (2016)
Privacy-Preserving Logistic Regression with Distributed Data Sources via Homomorphic Encryption
IEICE Trans. Inf. Syst., 99-D
Mohammad Nakhkash, Tuan Gia, I. Azimi, A. Anzanpour, A. Rahmani, P. Liljeberg (2019)
Analysis of Performance and Energy Consumption of Wearable Devices and Mobile Gateways in IoT Applications
Proceedings of the International Conference on Omni-Layer Intelligent Systems
Yoshiko Yasumura, Yu Ishimaki, H. Yamana (2019)
Secure Naïve Bayes Classification Protocol over Encrypted Data Using Fully Homomorphic Encryption
Proceedings of the 21st International Conference on Information Integration and Web-based Applications & Services
Sedigh Ghamari, Koray Ozcan, Thu Dinh, A. Melnikov, Juan Carvajal, Jan Ernst, S. Chai (2021)
Quantization-Guided Training for Compact TinyML Models
ArXiv, abs/2103.06231
J. Hayano, Hiroaki Yamamoto, Izumi Nonaka, Makoto Komazawa, K. Itao, Norihiro Udea, Haruhito Tanaka, E. Yuda (2020)
Quantitative detection of sleep apnea with wearable watch device
PLoS ONE, 15
M. Imtiaz, Raul Ramos-Garcia, S. Wattal, S. Tiffany, E. Sazonov (2019)
Wearable Sensors for Monitoring of Cigarette Smoking in Free-Living: A Systematic Review
Sensors (Basel, Switzerland), 19
N. Baracaldo S. Truex (2019)
A hybrid approach to privacy-preserving federated learning,
Proceedings of the 12th ACM Workshop on Artificial Intelligence and Security
V. Suriyakumar, Nicolas Papernot, A. Goldenberg, M. Ghassemi (2020)
Challenges of Differentially Private Prediction in Healthcare Settings
Sungkyu Park, Sang Lee, Sungwon Han, M. Cha (2019)
Clustering Insomnia Patterns by Data From Wearable Devices: Algorithm Development and Validation Study
JMIR mHealth and uHealth, 7
J. Hu L. Zhang (2020)
Dynamic power optimization for secondary wearable biosensors in e-healthcare leveraging cognitive WBSNs with imperfect spectrum sensing,
Future Generation Computer Systems, 112
T. Chitnis, B. Glanz, Cindy Gonzalez, B. Healy, Taylor Saraceno, N. Sattarnezhad, C. Díaz-Cruz, Mariann Polgar-Turcsanyi, Subhash Tummala, R. Bakshi, Vikram Bajaj, David Ben-Shimol, Nikhil Bikhchandani, A. Blocker, Joshua Burkart, R. Cendrillon, Michael Cusack, Emre Demiralp, Sarel Jooste, Alaa Kharbouch, A. Lee, J. Lehár, Manway Liu, S. Mahadevan, Mark Murphy, Linda Norton, Tushar Parlikar, Anupam Pathak, Ali Shoeb, E. Soderberg, Philip Stephens, Aaron Stoertz, Florence Thng, Kashyap Tumkur, Hongsheng Wang, Jane Rhodes, R. Rudick, R. Ransohoff, G. Phillips, Effie Bruzik, W. Marks, H. Weiner, T. Snyder (2019)
Quantifying neurologic disease using biosensor measurements in-clinic and in free-living settings in multiple sclerosis
NPJ Digital Medicine, 2
R. Reith, T. Schneider, Oleksandr Tkachenko (2019)
Efficiently Stealing your Machine Learning Models
Proceedings of the 18th ACM Workshop on Privacy in the Electronic Society
Lanfang Sun, Xin Jiang, Huixia Ren, Yi Guo (2020)
Edge-Cloud Computing and Artificial Intelligence in Internet of Medical Things: Architecture, Technology and Application
IEEE Access, 8
H. Jung C. Kang (2019)
Towards machine learning with zero real-world data,
Proceedings of the The 5th ACM Workshop on Wearable Systems and Applications, WearSys ’19
B. Bell, Ridwan Alam, N. Alshurafa, Edison Thomaz, A. Mondol, K. Haye, J. Stankovic, J. Lach, D. Spruijt-Metz (2020)
Automatic, wearable-based, in-field eating detection approaches for public health research: a scoping review
NPJ Digital Medicine, 3
C. Meisel, Rima Atrache, M. Jackson, S. Schubach, Claire Ufongene, T. Loddenkemper (2020)
Machine learning from wristband sensor data for wearable, noninvasive seizure forecasting
Epilepsia, 61
Po-Ting Lee, Wen-Ching Chiu, Y. Ho, You-Cheng Tai, Chou-Ching Lin, Chih-Lung Lin (2021)
Development of Wearable Device and Clustering Based Method for Detecting Falls in the Elderly
2021 IEEE 10th Global Conference on Consumer Electronics (GCCE)
Hugo Posada-Quintero, N. Reljin, Aurelie Moutran, Dimitrios Georgopalis, E. Lee, Gabrielle Giersch, D. Casa, K. Chon (2019)
Mild Dehydration Identification Using Machine Learning to Assess Autonomic Responses to Cognitive Stress
Nutrients, 12
A. Masino, Daniel Forsyth, H. Nuske, J. Herrington, Jeffrey Pennington, Yelena Kushleyeva, Christopher Bonafide (2019)
m-Health and Autism: Recognizing Stress and Anxiety with Machine Learning and Wearables Data
2019 IEEE 32nd International Symposium on Computer-Based Medical Systems (CBMS)
M. Scrugli, Daniela Loi, L. Raffo, P. Meloni (2019)
A runtime-adaptive cognitive IoT node for healthcare monitoring
Proceedings of the 16th ACM International Conference on Computing Frontiers
Dan Morris, T. Saponas, Andrew Guillory, Ilya Kelner (2014)
RecoFit: using a wearable sensor to find, recognize, and count repetitive exercises
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems
Olivia Walch, Yitong Huang, Daniel Forger, Cathy Goldstein (2019)
Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device
Sleep, 42
A. M. Shapiro J. S. Chorba
Deep learning algorithm for automated cardiac murmur detection via a digital stethoscope platform,
Journal of American Heart Association, 10
David Pollreisz, N. Taherinejad (2019)
Detection and Removal of Motion Artifacts in PPG Signals
Mobile Networks and Applications, 27
Anna Ferrari, D. Micucci, M. Mobilio, Paolo Napoletano (2022)
Deep learning and model personalization in sensor-based human activity recognition
Journal of Reliable Intelligent Environments, 9
B. Milosevic, S. Benatti, Elisabetta Farella (2017)
Design challenges for wearable EMG applications
Design, Automation & Test in Europe Conference & Exhibition (DATE), 2017
J. Greiwe, S. Nyenhuis (2020)
Wearable Technology and How This Can Be Implemented into Clinical Practice
Current Allergy and Asthma Reports, 20
Jian Huang, Anirudh Badam, Ranveer Chandra, Edmund Nightingale (2015)
WearDrive: Fast and Energy-Efficient Storage for Wearables
R. Ghaffari, J. Rogers, Tyler Ray (2021)
Recent progress, challenges, and opportunities for wearable biochemical sensors for sweat analysis.
Sensors and actuators. B, Chemical, 332
Kai-Chun Liu, Chia-Yeh Hsieh, S. Hsu, Chia-Tai Chan (2018)
Impact of Sampling Rate on Wearable-Based Fall Detection Systems Based on Machine Learning Models
IEEE Sensors Journal, 18
Daniel Garcia-Gonzalez, D. Rivero, Enrique Fernández-Blanco, M. Luaces (2020)
A Public Domain Dataset for Real-Life Human Activity Recognition Using Smartphone Sensors
Sensors (Basel, Switzerland), 20
A. Qayyum, Junaid Qadir, M. Bilal, A. Al-Fuqaha (2020)
Secure and Robust Machine Learning for Healthcare: A Survey
IEEE Reviews in Biomedical Engineering, 14
Nattaya Mairittha, Tittaya Mairittha, Sozo Inoue (2020)
On-Device Deep Personalization for Robust Activity Data Collection †
Sensors (Basel, Switzerland), 21
Antonio Alvarez, E. Severeyn, J. Velásquez, Sara Wong, Gilberto Perpiñan, M. Huerta (2019)
Machine Learning Methods in the Classification of the Athletes Dehydration
2019 IEEE Fourth Ecuador Technical Chapters Meeting (ETCM)
Sasanka Potluri, A. Chandran, C. Diedrich, L. Schega (2019)
Machine Learning based Human Gait Segmentation with Wearable Sensor Platform
2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
E. Torti, Claudia d’Amato, G. Danese, F. Leporati (2020)
An Hardware Recurrent Neural Network for Wearable Devices
2020 23rd Euromicro Conference on Digital System Design (DSD)
A. Gharibans, B. Smarr, D. Kunkel, L. Kriegsfeld, H. Mousa, Todd Coleman (2018)
Artifact Rejection Methodology Enables Continuous, Noninvasive Measurement of Gastric Myoelectric Activity in Ambulatory Subjects
Scientific Reports, 8
T. Hunt, Congzheng Song, R. Shokri, Vitaly Shmatikov, E. Witchel (2018)
Chiron: Privacy-preserving Machine Learning as a Service
ArXiv, abs/1803.05961
Shengjie Bi, Tao Wang, E. Davenport, Ronald Peterson, R. Halter, Jacob Sorber, D. Kotz (2017)
Toward a Wearable Sensor for Eating Detection
Proceedings of the 2017 Workshop on Wearable Systems and Applications
H. True A. Harris (2016)
Fall recognition using wearable technologies and machine learning algorithms,
Proceedings of the 2016 IEEE International Conference on Big Data
Farida Sabry, T. Eltaras, Wadha Labda, Fatima Hamza, K. Alzoubi, Q. Malluhi (2022)
Towards On-Device Dehydration Monitoring Using Machine Learning from Wearable Device’s Data
Sensors (Basel, Switzerland), 22
Yichen Shen, M. Voisin, Alireza Aliamiri, A. Avati, Awni Hannun, A. Ng (2018)
Ambulatory Atrial Fibrillation Monitoring Using Wearable Photoplethysmography with Deep Learning
Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining
Qinggang Yue, Zhen Ling, Xinwen Fu, Benyuan Liu, Wei Yu, Wei Zhao (2014)
My Google Glass Sees Your Passwords!
R. Lazzeretti, M. Barni (2013)
Private Computing with Garbled Circuits [Applications Corner]
IEEE Signal Processing Magazine, 30
Soon Kwon, Jeong-Ho Park, Chiheon Kwon, Hyung Kong, J. Hwang, Hee Kim (2019)
An Energy-Efficient Algorithm for Classification of Fall Types Using a Wearable Sensor
IEEE Access, 7
Andrés Jaramillo-Yánez, Marco Benalcázar, Elisa Mena-Maldonado (2020)
Real-Time Hand Gesture Recognition Using Surface Electromyography and Machine Learning: A Systematic Literature Review
Sensors (Basel, Switzerland), 20
Yiqiang Chen, Jindong Wang, Chaohui Yu, Wen Gao, Xin Qin (2019)
FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare
IEEE Intelligent Systems, 35
F. Akbulut, B. Ikitimur, A. Akan (2020)
Wearable sensor-based evaluation of psychosocial stress in patients with metabolic syndrome
Artificial intelligence in medicine, 104
Charikleia Chatzaki, M. Pediaditis, George Vavoulas, M. Tsiknakis (2016)
Human Daily Activity and Fall Recognition Using a Smartphone's Acceleration Sensor
Zhongzheng Fu, Xinrun He, E. Wang, Jun Huo, Jian Huang, Dongrui Wu (2021)
Personalized Human Activity Recognition Based on Integrated Wearable Sensor and Transfer Learning
Sensors (Basel, Switzerland), 21
Huanghe Zhang, Yi Guo, D. Zanotto (2019)
Accurate Ambulatory Gait Analysis in Walking and Running Using Machine Learning Models
IEEE Transactions on Neural Systems and Rehabilitation Engineering, 28
Ramyar Saeedi, K. Sasani, S. Norgaard, A. Gebremedhin (2018)
Personalized Human Activity Recognition using Wearables: A Manifold Learning-based Knowledge Transfer
2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
F. Kraemer, Anders Braten, N. Tamkittikhun, David Palma (2017)
Fog Computing in Healthcare–A Review and Discussion
IEEE Access, 5
Edison Thomaz, Irfan Essa, G. Abowd (2015)
A practical approach for recognizing eating moments with wrist-mounted inertial sensing
Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing
S. Moore, C. Kranzinger, Julian Fritz, T. Stöggl, Josef Krll, H. Schwameder (2020)
Foot Strike Angle Prediction and Pattern Classification Using LoadsolTM Wearable Sensors: A Comparison of Machine Learning Techniques
Sensors (Basel, Switzerland), 20
Marjan Nourollahi, Seyed Rokni, Parastoo Alinia, Hassan Ghasemzadeh (2020)
Proximity-based active learning for eating moment recognition in wearable systems
Proceedings of the 6th ACM Workshop on Wearable Systems and Applications
(2020)
Guidelines for wrist-worn wearable assessment of heart rate in biobehavioral research
Rita Meziatisabour, Y. Benezeth, Pierre Oliveira, J. Chappé, Fan Yang (2021)
UBFC-Phys: A Multimodal Database For Psychophysiological Studies of Social Stress
IEEE Transactions on Affective Computing, 14
Bargav Jayaraman, David Evans (2019)
Evaluating Differentially Private Machine Learning in Practice
Kalaivani Sundararajan, S. Georgievska, Bart Lindert, P. Gehrman, J. Ramautar, D. Mazzotti, S. Sabia, M. Weedon, E. Someren, Lars Ridder, Jian Wang, V. Hees (2021)
Sleep classification from wrist-worn accelerometer data using random forests
Scientific Reports, 11
Enrique Garcia-Ceja, M. Riegler, T. Nordgreen, P. Jakobsen, K. Oedegaard, J. Tørresen (2018)
Mental health monitoring with multimodal sensing and machine learning: A survey
Pervasive Mob. Comput., 51
A. Tonacci, L. Billeci, Irene Mambro, R. Marangoni, C. Sanmartin, F. Venturi (2021)
Wearable Sensors for Assessing the Role of Olfactory Training on the Autonomic Response to Olfactory Stimulation
Sensors (Basel, Switzerland), 21
Sungjoo Hwang, Houtan Jebelli, Byungjoo Choi, M. Choi, SangHyun Lee (2018)
Measuring Workers’ Emotional State during Construction Tasks Using Wearable EEG
Journal of Construction Engineering and Management
S. Rastogi, Jaspreet Singh (2021)
A systematic review on machine learning for fall detection system
Computational Intelligence, 37
Alejandro Arrieta, Natalia Rodríguez, J. Ser, Adrien Bennetot, S. Tabik, A. Barbado, S. García, S. Gil-Lopez, D. Molina, Richard Benjamins, Raja Chatila, Francisco Herrera (2019)
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI
Inf. Fusion, 58
NIST Special Publication 800-121 Revision 2
Miao Huang, Dongyi Chen, Fan Xiong (2019)
An Effective Adaptive Filter to Reduce Motion Artifacts from ECG Signals Using Accelerometer
Proceedings of the 2019 9th International Conference on Biomedical Engineering and Technology
Jinfeng Yuan, R. Zhu (2020)
A fully self-powered wearable monitoring system with systematically optimized flexible thermoelectric generator
Applied Energy, 271
M. Chan, D. Estève, J. Fourniols, C. Escriba, E. Campo (2012)
Smart wearable systems: Current status and future challenges
Artificial intelligence in medicine, 56 3
Armando Collado‐Villaverde, Mario Cobos, Pablo Muñoz, David Barrero (2020)
A Simulator to Support Machine Learning-Based Wearable Fall Detection Systems
Electronics
Sangwon Bae, Denzil Ferreira, B. Suffoletto, J. Puyana, Ryan Kurtz, T. Chung, A. Dey (2017)
Detecting Drinking Episodes in Young Adults Using Smartphone-based Sensors
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 1
R. Sherratt, Balazs Janko, Terence Hui, W. Harwin, N. Dey, D. Díaz-Sánchez, Jin Wang, Fuqian Shi (2019)
Task Scheduling to Constrain Peak Current Consumption in Wearable Healthcare Sensors
Electronics
Victor Costan, S. Devadas (2016)
Intel SGX Explained
IACR Cryptol. ePrint Arch., 2016
Pouyan Esmaeilzadeh (2020)
Use of AI-based tools for healthcare purposes: a survey study from consumers’ perspectives
BMC Medical Informatics and Decision Making, 20
Suranga Seneviratne, Yining Hu, Tham Nguyen, Guohao Lan, Sara Khalifa, Kanchana Thilakarathna, Mahbub Hassan, A. Seneviratne (2017)
A Survey of Wearable Devices and Challenges
IEEE Communications Surveys & Tutorials, 19
Geng Yang, Jia Deng, Gaoyang Pang, Hao Zhang, Jiayi Li, Bin Deng, Zhibo Pang, Juan Xu, Mingzhe Jiang, P. Liljeberg, Haibo Xie, Huayong Yang (2018)
An IoT-Enabled Stroke Rehabilitation System Based on Smart Wearable Armband and Machine Learning
IEEE Journal of Translational Engineering in Health and Medicine, 6
Pekka Siirtola, Heli Koskimäki, Henna Mönttinen, J. Röning (2018)
Using Sleep Time Data from Wearable Sensors for Early Detection of Migraine Attacks
Sensors (Basel, Switzerland), 18
(2019)
Motion and Heart Rate from a Wrist-Worn Wearable and Labeled Sleep from Polysomnography
Z. Ling Q. Yue
My Google Glass Sees Your Passwords! Black Hat USA,
Sensors
R. Delgado-Gonzalo C. Segarra
Using trusted execution environments for secure stream processing of medical data,
Proceedings of the IFIP International Conference on Distributed Applications and Interoperable Systems
C. Balakrishna, E. Rendon-Morales, R. Aviles-Espinosa, Henry Dore, Zhenhua Luo (2019)
Challenges of Wearable Health Monitors : A Case study of Foetal ECG Monitor
2019 Global IoT Summit (GIoTS)
Sung Baek, Ki-Woong Park (2020)
A Durable Hybrid RAM Disk with a Rapid Resilience for Sustainable IoT Devices
Sensors (Basel, Switzerland), 20
K. Labat, K. Ryan (2019)
Human Body
Chun-Ka Wong, D. Ho, A. Tam, Mi Zhou, Yuk-Ming Lau, M. Tang, R. Tong, K. Rajput, Gengbo Chen, Soon-Chee Chan, C. Siu, I. Hung (2020)
Artificial intelligence mobile health platform for early detection of COVID-19 in quarantine subjects using a wearable biosensor: protocol for a randomised controlled trial
BMJ Open, 10
C. Dwork (2006)
Differential Privacy
E. Rendon-Morales C. Balakrishna
Challenges of wearable health monitors: a case study of foetal ECG monitor,
Proceedings of the 2019 global IoT summit (GIoTS)
Anupama R., A. R, P. Pahwa, Anita Ramachandran (2018)
Machine Learning-Based Techniques for Fall Detection in Geriatric Healthcare Systems
2018 9th International Conference on Information Technology in Medicine and Education (ITME)
Xenofon Fafoutis, Letizia Marchegiani, Atis Elsts, James Pope, R. Piechocki, I. Craddock (2018)
Extending the battery lifetime of wearable sensors with embedded machine learning
2018 IEEE 4th World Forum on Internet of Things (WF-IoT)
Faisal Hussain, F. Hussain, M. Ehatisham-ul-Haq, M. Azam (2019)
Activity-Aware Fall Detection and Recognition Based on Wearable Sensors
IEEE Sensors Journal, 19
Amine Boulemtafes, A. Derhab, Y. Challal (2020)
A review of privacy-preserving techniques for deep learning
Neurocomputing, 384
Alireza Abedin, F. Motlagh, Javen Shi, S. Rezatofighi, D. Ranasinghe (2020)
Towards deep clustering of human activities from wearables
Proceedings of the 2020 ACM International Symposium on Wearable Computers
Alexander Wood, K. Najarian, Delaram Kahrobaei (2020)
Homomorphic Encryption for Machine Learning in Medicine and Bioinformatics
ACM Computing Surveys (CSUR), 53
Davide Giuffrida, G. Benetti, D. Martini, T. Facchinetti (2019)
Fall Detection with Supervised Machine Learning using Wearable Sensors
2019 IEEE 17th International Conference on Industrial Informatics (INDIN), 1
Eftim Zdravevski, Biljana Stojkoska, M. Standl, H. Schulz (2017)
Automatic machine-learning based identification of jogging periods from accelerometer measurements of adolescents under field conditions
PLoS ONE, 12
Chih-Ta Yen, Jia-Xian Liao, Yi-Kai Huang (2020)
Human Daily Activity Recognition Performed Using Wearable Inertial Sensors Combined With Deep Learning Algorithms
IEEE Access, 8
Yongge Wang, Q. Malluhi, K. Khan (2016)
Garbled computation in cloud
Future Gener. Comput. Syst., 62
Y. Choi, Y. Jeon, L. Wang, Kwanho Kim (2017)
A Biological Signal-Based Stress Monitoring Framework for Children Using Wearable Devices
Sensors (Basel, Switzerland), 17
M. Nassar, A. Erradi, Farida Sabry, Q. Malluhi (2013)
Secure Outsourcing of Matrix Operations as a Service
2013 IEEE Sixth International Conference on Cloud Computing
Ji Kim, J. Oh, Gwan Hwang, Oh Gwon, Seung Lee (2020)
Design of Low-Power SoC for Wearable Healthcare Device
J. Circuits Syst. Comput., 29
Djihane Houfani, S. Slatnia, O. Kazar, Hamza Saouli, Abdelhak Merizig (2021)
Artificial intelligence in healthcare: a review on predicting clinical needs
International Journal of Healthcare Management, 15
Shreshth Tuli, Nipam Basumatary, S. Gill, M. Kahani, R. Arya, G. Wander, R. Buyya (2019)
HealthFog: An Ensemble Deep Learning based Smart Healthcare System for Automatic Diagnosis of Heart Diseases in Integrated IoT and Fog Computing Environments
Future Gener. Comput. Syst., 104
J. Cheon, Duhyeong Kim, J. Park (2019)
Towards a Practical Cluster Analysis over Encrypted Data
Sangwook Kim, Masahiro Omori, Takuya Hayashi, T. Omori, Lihua Wang, S. Ozawa (2018)
Privacy-Preserving Naive Bayes Classification Using Fully Homomorphic Encryption
Y. Can, Cem Ersoy (2021)
Privacy-preserving Federated Deep Learning for Wearable IoT-based Biomedical Monitoring
ACM Transactions on Internet Technology (TOIT), 21
S. Kwon, Joonki Hong, E. Choi, Byunghwan Lee, Chang-Yeoul Baik, Euijae Lee, Eui-Rim Jeong, B. Koo, Seil Oh, Yung Yi (2020)
Detection of Atrial Fibrillation Using a Ring-Type Wearable Device (CardioTracker) and Deep Learning Analysis of Photoplethysmography Signals: Prospective Observational Proof-of-Concept Study
Journal of Medical Internet Research, 22
Y. Can, B. Arnrich, Cem Ersoy (2019)
Stress detection in daily life scenarios using smart phones and wearable sensors: A survey
Journal of biomedical informatics, 92
Katrin Hänsel, Kleomenis Katevas, Guido Orgs, Daniel Richardson, A. Alomainy, H. Haddadi (2018)
The potential of wearable technology for monitoring social interactions based on interpersonal synchrony
Proceedings of the 4th ACM Workshop on Wearable Systems and Applications
T. Tomono (2016)
Design consideration of sustainable self‐sufficient energy system using ferroelectric material for wearable healthcare
International Journal of Energy Research, 40
A. Keerthy, S. Priya (2021)
Artificial Intelligence in Healthcare Databases
Marcin Lewandowski, B. Płaczek, M. Bernaś (2020)
Classifier-Based Data Transmission Reduction in Wearable Sensor Network for Human Activity Monitoring
Sensors (Basel, Switzerland), 21
Austin Harris, Hanna True, Zhen Hu, Jin Cho, N. Fell, Mina Sartipi
2016 Ieee International Conference on Big Data (big Data) Fall Recognition Using Wearable Technologies and Machine Learning Algorithms
Liang Xiao, Xiaoyue Wan, Xiaozhen Lu, Yanyong Zhang, Di Wu (2018)
IoT Security Techniques Based on Machine Learning: How Do IoT Devices Use AI to Enhance Security?
IEEE Signal Processing Magazine, 35
Jeong Heo, H. Yoon, K. Park (2017)
A Novel Wearable Forehead EOG Measurement System for Human Computer Interfaces
Sensors (Basel, Switzerland), 17
Andy Cheon, Stephanie Jung, Collin Prather, Matthew Sarmiento, Kevin Wong, D. Woodbridge (2020)
A Machine Learning Approach to Detecting Low Medication State with Wearable Technologies
2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)
M. Merenda, Carlo Porcaro, D. Iero (2020)
Edge Machine Learning for AI-Enabled IoT Devices: A Review
Sensors (Basel, Switzerland), 20
P. Compagnon, G. Lefebvre, S. Duffner, Christophe Garcia (2020)
Learning personalized ADL recognition models from few raw data
Artificial intelligence in medicine, 107
Chulhong Min, Akhil Mathur, F. Kawsar (2018)
Exploring audio and kinetic sensing on earable devices
Proceedings of the 4th ACM Workshop on Wearable Systems and Applications
C. Ieracitano, N. Mammone, A. Hussain, F. Morabito (2019)
A novel multi-modal machine learning based approach for automatic classification of EEG recordings in dementia
Neural networks : the official journal of the International Neural Network Society, 123
Sander Koelstra, C. Mühl, M. Soleymani, Jong-Seok Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, I. Patras (2012)
DEAP: A Database for Emotion Analysis ;Using Physiological Signals
IEEE Transactions on Affective Computing, 3
Qingxue Zhang, Xuan Zeng, W. Hu, Dian Zhou (2017)
A Machine Learning-Empowered System for Long-Term Motion-Tolerant Wearable Monitoring of Blood Pressure and Heart Rate With Ear-ECG/PPG
IEEE Access, 5
E. Casilari-Pérez, J. Santoyo-Ramón, J. Cano-García (2017)
Analysis of Public Datasets for Wearable Fall Detection Systems
Sensors (Basel, Switzerland), 17
Zhenning Mei, Xiao Gu, Hongyu Chen, Wei Chen (2018)
Automatic Atrial Fibrillation Detection Based on Heart Rate Variability and Spectral Features
IEEE Access, 6
Ruth Davies (2019)
Human Body: A Wearable Product Designer’s Guide
Ergonomics, 63
Erwin Adi, A. Anwar, Z. Baig, S. Zeadally (2020)
Machine learning and data analytics for the IoT
Neural Computing and Applications, 32
A. Greco, G. Valenza, A. Lanatà, G. Rota, E. Scilingo (2014)
Electrodermal Activity in Bipolar Patients during Affective Elicitation
IEEE Journal of Biomedical and Health Informatics, 18
M. Kilany, A. Hassanien, A. Badr (2015)
Accelerometer-based human activity classification using Water Wave Optimization approach
2015 11th International Computer Engineering Conference (ICENCO)
Decheng Zhang, Hengjing Zhang, Chuanxin Zhao (2022)
Prediction of Behavior and Activity of Patients through Wearable Devices Based on Deep Learning
J. Sensors, 2022
Jing Li, Xiaohui Kuang, Shujie Lin, Xu Ma, Yi Tang (2020)
Privacy preservation for machine learning training and classification based on homomorphic encryption schemes
Inf. Sci., 526
Jun Zhou, Chao Wang (2017)
An Ultra-Low Power Turning Angle Based Biomedical Signal Compression Engine with Adaptive Threshold Tuning
Sensors (Basel, Switzerland), 17
D. Jaiswal, Andrew Gigie, T. Chakravarty, Avik Ghose, Archan Misra (2018)
Table of interest: activity recognition and behaviour analysis using a battery less wearable sensor
Proceedings of the 4th ACM Workshop on Wearable Systems and Applications
Faisal Alsubaei, Abdullah Abuhussein, S. Shiva (2017)
Security and Privacy in the Internet of Medical Things: Taxonomy and Risk Assessment
2017 IEEE 42nd Conference on Local Computer Networks Workshops (LCN Workshops)
F. Miao, Bo Wen, Zhejing Hu, G. Fortino, Xi-Ping Wang, Zengding Liu, Min Tang, Ye Li (2020)
Continuous blood pressure measurement from one-channel electrocardiogram signal using deep-learning techniques
Artificial intelligence in medicine, 108
Yehenew Mengistu, Minh Pham, H. Do, W. Sheng (2016)
AutoHydrate: A wearable hydration monitoring system
2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Ayush Jain, David Way, Vishakha Gupta, Y. Gao, Guilherme Marinho, Jay Hartford, R. Sayres, K. Kanada, C. Eng, K. Nagpal, K. Desalvo, G. Corrado, L. Peng, D. Webster, R. Dunn, David Coz, S. Huang, Yun Liu, P. Bui, Yuan Liu (2021)
Development and Assessment of an Artificial Intelligence–Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices
JAMA Network Open, 4
F. Onorati, Giulia Regalia, C. Caborni, Matteo Migliorini, Daniel Bender, M. Poh, Cherise Frazier, Eliana Thropp, Elizabeth Mynatt, Jonathan Bidwell, R. Mai, W. LaFrance, A. Blum, D. Friedman, T. Loddenkemper, Fatemeh Touserkani, C. Reinsberger, Simone Tognetti, Rosalind Picard (2017)
Multicenter clinical assessment of improved wearable multimodal convulsive seizure detectors
Epilepsia, 58
Xiaoou Li, Zhiyong Zhou, Jiajia Wu, Yichao Xiong (2021)
Human Posture Detection Method Based on Wearable Devices
Journal of Healthcare Engineering, 2021
Nafiul Rashid, M. Faruque (2020)
Energy-efficient Real-time Myocardial Infarction Detection on Wearable Devices
2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)
Laura Stevens, B. Mortazavi, R. Deo, L. Curtis, D. Kao (2020)
Recommendations for Reporting Machine Learning Analyses in Clinical Research
Circulation: Cardiovascular Quality and Outcomes, 13
Giulia Regalia, F. Onorati, Matteo Lai, C. Caborni, Rosalind Picard (2019)
Multimodal wrist-worn devices for seizure detection and advancing research: Focus on the Empatica wristbands
Epilepsy Research, 153
D. Liaqat, R. Wu, A. Gershon, Hisham Alshaer, Frank Rudzicz, E. Lara (2018)
Challenges with real-world smartwatch based audio monitoring
Proceedings of the 4th ACM Workshop on Wearable Systems and Applications
Giang Nguyen, Stefan Dlugolinsky, Martin Bobák, V. Tran, Á. García, Ignacio Heredia, Peter Malík, L. Hluchý (2019)
Machine Learning and Deep Learning frameworks and libraries for large-scale data mining: a survey
Artificial Intelligence Review, 52
Angela Sucerquia, J. López, J. Vargas-Bonilla (2017)
SisFall: A Fall and Movement Dataset
Sensors (Basel, Switzerland), 17
Ronald Rivest, M. Dertouzos (1978)
ON DATA BANKS AND PRIVACY HOMOMORPHISMS
Chaoming Fang, Yixuan Wang, Shuo Gao (2020)
Utilizing Wearable GRF and EMG Sensing System and Machine Learning Algorithms to Enable Locomotion Mode Recognition for In-home Rehabilitation
2020 IEEE International Conference on Flexible and Printable Sensors and Systems (FLEPS)
Miran Kim, Yongsoo Song, Shuang Wang, Yuhou Xia, Xiaoqian Jiang (2018)
Secure Logistic Regression Based on Homomorphic Encryption: Design and Evaluation
JMIR Medical Informatics, 6
Yoshinori Aono, Takuya Hayashi, L. Phong, Lihua Wang (2016)
Scalable and Secure Logistic Regression via Homomorphic Encryption
Proceedings of the Sixth ACM Conference on Data and Application Security and Privacy
Rita Laureanti, Marco Bilucaglia, M. Zito, Riccardo Circi, Alessandro Fici, Fiamma Rivetti, Riccardo Valesi, Carlo Oldrini, L. Mainardi, V. Russo (2020)
Emotion assessment using Machine Learning and low-cost wearable devices
2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)
B. Mateen, J. Liley, A. Denniston, Chris Holmes, Sebastian Vollmer (2020)
Improving the quality of machine learning in health applications and clinical research
Nature Machine Intelligence, 2
Hao Chen, Ran Gilad-Bachrach, Kyoohyung Han, Zhicong Huang, A. Jalali, Kim Laine, K. Lauter (2018)
Logistic regression over encrypted data from fully homomorphic encryption
BMC Medical Genomics, 11
H. McMahan, Eider Moore, Daniel Ramage, B. Arcas (2016)
Federated Learning of Deep Networks using Model Averaging
ArXiv, abs/1602.05629
Ricardo Buettner, Janek Frick, Thilo Rieg (2019)
High-performance detection of epilepsy in seizure-free EEG recordings: A novel machine learning approach using very specific epileptic EEG sub-bands
J. Frick R. Buettner (2019)
High-performance Detection of Epilepsy in Seizure-free EEG Recordings,
Proceedings of the International Conference on Information systems
Lucie Klus, E. Lohan, C. Granell, J. Nurmi (2020)
Lossy Compression Methods for Performance-Restricted Wearable Devices
O. Zoubi, M. Awad, N. Kasabov (2018)
Anytime multipurpose emotion recognition from EEG data using a Liquid State Machine based framework
Artificial intelligence in medicine, 86
Alma Secerbegovic, A. Gogic, N. Suljanovic, M. Zajc, A. Mujčić (2018)
Computational Balancing between Wearable Sensor and Smartphone towards Energy-Efficient Remote Healthcare Monitoring
Advances in Electrical and Computer Engineering, 18
D. Tamburri (2020)
Sustainable MLOps: Trends and Challenges
2020 22nd International Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC)
Cholmin Kang, Hyunwoo Jung, Youngki Lee (2019)
Towards Machine Learning with Zero Real-World Data
The 5th ACM Workshop on Wearable Systems and Applications
Ting Zhang, Jiang Lu, Fei Hu, Qi Hao (2014)
Bluetooth low energy for wearable sensor-based healthcare systems
2014 IEEE Healthcare Innovation Conference (HIC)
Tendai Rukasha, Sandra Woolley, T. Kyriacou, Tim Collins (2020)
Evaluation of Wearable Electronics for Epilepsy: A Systematic Review
Electronics
Jong Kim, S. Moon, Sang-ug Kang, Beakcheol Jang (2020)
Effective Privacy-Preserving Collection of Health Data from a User’s Wearable Device
Applied Sciences
Long Zhang, Jinhuang Hu, Chao Guo, Haitao Xu (2020)
Dynamic power optimization for secondary wearable biosensors in e-healthcare leveraging cognitive WBSNs with imperfect spectrum sensing
Future Gener. Comput. Syst., 112
Viet-thi Tran, C. Riveros, P. Ravaud (2019)
Patients’ views of wearable devices and AI in healthcare: findings from the ComPaRe e-cohort
NPJ Digital Medicine, 2
Po-Hsuan Chen, Yun Liu, L. Peng (2019)
How to develop machine learning models for healthcare
Nature Materials, 18
Ka-chun Un, Chun-Ka Wong, Yuk-Ming Lau, J. Lee, F. Tam, W. Lai, Y. Lau, Hao Chen, Sandi Wibowo, Xiaozhu Zhang, Minghao Yan, Esther Wu, Soon-Chee Chan, Sze-Ming Lee, Augustine Chow, R. Tong, M. Majmudar, K. Rajput, I. Hung, C. Siu (2021)
Observational study on wearable biosensors and machine learning-based remote monitoring of COVID-19 patients
Scientific Reports, 11
Abdelkareem Bedri, Apoorva Verlekar, Edison Thomaz, Valerie Avva, Thad Starner (2015)
A wearable system for detecting eating activities with proximity sensors in the outer ear
Proceedings of the 2015 ACM International Symposium on Wearable Computers
Diana Batista, Hugo Silva, A. Fred, Carlos Moreira, M. Reis, H. Ferreira (2019)
Benchmarking of the BITalino biomedical toolkit against an established gold standard
Healthcare Technology Letters, 6
Sergiu Carpov, Nicolas Gama, Mariya Georgieva, J. Troncoso-Pastoriza (2020)
Privacy-preserving semi-parallel logistic regression training with fully homomorphic encryption
BMC Medical Genomics, 13
O. Balan, G. Moise, A. Moldoveanu, Marius Leordeanu, F. Moldoveanu (2020)
An Investigation of Various Machine and Deep Learning Techniques Applied in Automatic Fear Level Detection and Acrophobia Virtual Therapy
Sensors (Basel, Switzerland), 20
Abdelkareem Bedri, Richard Li, Malcolm Haynes, Raj Kosaraju, Ishaan Grover, T. Prioleau, Min Beh, Mayank Goel, Thad Starner, G. Abowd (2017)
EarBit
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 1
Değer Ayata, Y. Yaslan, M. Kamasak (2020)
Emotion Recognition from Multimodal Physiological Signals for Emotion Aware Healthcare Systems
Journal of Medical and Biological Engineering, 40
D. Anguita, A. Ghio, L. Oneto, X. Parra, Jorge Reyes-Ortiz (2013)
A Public Domain Dataset for Human Activity Recognition using Smartphones
John Chorba, Avi Shapiro, L. Le, J. Maidens, John Prince, Steve Pham, Mia Kanzawa, Daniel Barbosa, Caroline Currie, Catherine Brooks, Brent White, Anna Huskin, Jason Paek, Jack Geocaris, Dinatu Elnathan, Ria Ronquillo, Roy Kim, Zenith Alam, V. Mahadevan, Sophie Fuller, Grant Stalker, Sara Bravo, Dina Jean, John Lee, Medeona Gjergjindreaj, Christos Mihos, S. Forman, Subramaniam Venkatraman, P. McCarthy, James Thomas (2020)
Deep Learning Algorithm for Automated Cardiac Murmur Detection via a Digital Stethoscope Platform
Journal of the American Heart Association: Cardiovascular and Cerebrovascular Disease, 10
Franca Delmastro, Flavio Martino, C. Dolciotti (2020)
Cognitive Training and Stress Detection in MCI Frail Older People Through Wearable Sensors and Machine Learning
IEEE Access, 8

Publisher: Hindawi Publishing Corporation
Copyright: Copyright © 2022 Farida Sabry et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN: 2040-2295
eISSN: 2040-2309
DOI: 10.1155/2022/4653923
Publisher site: See Article on Publisher Site

Abstract

Hindawi Journal of Healthcare Engineering Volume 2022, Article ID 4653923, 25 pages https://doi.org/10.1155/2022/4653923 Review Article Machine Learning for Healthcare Wearable Devices: The Big Picture 1 1 1 2 Farida Sabry , Tamer Eltaras , Wadha Labda , Khawla Alzoubi , and Qutaibah Malluhi Computer Science and Engineering Department, Faculty of Engineering, Qatar University, Doha, Qatar Engineering Technology Department, Community College of Qatar, Doha, Qatar Correspondence should be addressed to Farida Sabry; faridasabry@qu.edu.qa Received 17 November 2021; Accepted 22 March 2022; Published 18 April 2022 Academic Editor: Yuxiang Wu Copyright © 2022 Farida Sabry et al. *is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Using artiﬁcial intelligence and machine learning techniques in healthcare applications has been actively researched over the last few years. It holds promising opportunities as it is used to track human activities and vital signs using wearable devices and assist in diseases’ diagnosis, and it can play a great role in elderly care and patient’s health monitoring and diagnostics. With the great technological advances in medical sensors and miniaturization of electronic chips in the recent ﬁve years, more applications are being researched and developed for wearable devices. Despite the remarkable growth of using smart watches and other wearable devices, a few of these massive research eﬀorts for machine learning applications have found their way to market. In this study, a review of the diﬀerent areas of the recent machine learning research for healthcare wearable devices is presented. Diﬀerent challenges facing machine learning applications on wearable devices are discussed. Potential solutions from the literature are presented, and areas open for improvement and further research are highlighted. features can be extracted for machine learning (ML) algo- 1. Introduction rithms to detect and learn useful patterns. *is can be very *e last few years have witnessed great advances in wearable useful in many healthcare and elderly care applications such technologies. Wearable devices include any device that can as activity detection for health state assessment, fall detec- be worn by humans such as wristwatches, glasses, chest tion, stress detection, ﬁtness tracking, vital signs monitoring, straps, rings, and prosthetic sockets. Wearable devices be- and diseases’ diagnosis. Using machine learning techniques long to the Internet of medical things (IoMT), together with to learn from human body signals, recorded by wearable the implantable, ambient, and stationary devices used in devices, has been an active research area in the last decade hospitals. *ese devices are typically connected to a network with a lot of published research studies. Despite this huge research eﬀort and the remarkable growth in using wearable and communicate remotely with mobile devices as shown in Figure 1. devices, especially smart watches, few machine learning Wearable devices may include diﬀerent types of sensors applications for wearable devices have found their way into to continuously monitor various human signals, e.g., tem- the market. perature sensors, accelerometers, optical sensors, and bio- Examples include irregular rhythm notiﬁcation feature metric sensors. Although the readings of some of these [3] in Apple Watch, which won U.S. Food and Drug Ad- sensors are not yet as accurate as stationary devices in ministration (FDA) approval with a long list of warnings and hospitals, they are sometimes considered acceptable [1, 2], precautions in 2018 (https://www.accessdata.fda.gov/ depending on the application. cdrh_docs/reviews/DEN180044.pdf), and Eko’s heart murmur detection algorithm, which has been recently Sensors in IoMT devices and human interaction with these devices are considered a big source of data from which published [4], which is not really for a personal wearable 2 Journal of Healthcare Engineering Smart Wearable devices phone Watch Helmet Glasses Middleware server Cloud-based infrastructure Chest strap On-device Edge/fog Cloud computing computing computing Figure 1: Wearable device application model. device but an electronic stethoscope. Additionally, some of the (2) What are the challenges facing machine learning for wearable devices that were used for monitoring, which were healthcare wearable devices? surveyed in [5], are no longer available in the market. Practical (3) What are the possible solutions for these challenges and reliable use of machine learning techniques in the domain in literature? of wearable devices is still facing many challenges. *us, the main focus of this study was to identify the Several review papers have discussed some challenges for challenges of developing machine learning applications for wearable devices. In a survey paper published in 2012 [6], the healthcare wearable devices and alternative solutions found authors focused on some features of wearable devices and in the literature. Diﬀerent categories of recent healthcare their types such as diseases that can be monitored, research machine learning research are reviewed while spotting the prototypes, and challenges such as system eﬃciency, user challenges and highlighting potential research areas and perception, cost, social inclusion, and ethical issues. In [5], applications that need further investigation. the authors provided a survey of commercially available *e rest of the review is organized as follows. In the next wearable devices at that time (2017). *ey focused on com- section, the necessary background for IoMTand the diﬀerent munication security issues, power eﬃciency, and wearable human body signals used in wearable devices research are computing. Neither of the surveys focused on the chal- presented. Moreover, applications for machine learning in lenges facing machine learning applications for healthcare IoMT are reviewed and categorized referencing some of the wearable devices speciﬁcally. recent research work published in each area. In Section 3, In this study, we review recent applied machine learning diﬀerent challenges facing machine learning research for research for wearable devices. We identify many challenges wearable devices are reviewed, relevant privacy and security facing machine learning applications on wearable devices aspects for machine learning applications in IoMT are from design to deployment, such as diﬀerent deployment discussed, and possible solutions in literature are presented. alternatives, storage, power consumption, user acceptance, In Section 4, we discuss these solutions, their applicability, reliability, communication, security, and privacy. We discuss and their shortcomings. Additionally, we highlight the main security and privacy both from the data and the model research gaps we perceived in the domain. Finally, the study perspectives listing potential solutions to keep subjects’ conclusions are provided in Section 5. personal data from wearable devices private and secure. Additionally, we review the diﬀerent privacy-preserving techniques used for machine learning training and inference 2. Wearable Devices and Machine Learning and discuss their applicability to the model of wearable *e wearable device domain is being actively researched for device usage shown in Figure 1. the sake of enhancing ease of use, comfort, and non- *e review includes the recent research papers in the invasiveness of monitoring physiological vital signs and ﬁeld of wearable devices that have been published from 2017 sometimes psychological or emotional state, which can be to December 2021 to answer the following questions: detected by analyzing data from diﬀerent sensors. Following (1) What are the healthcare machine learning tasks that the tremendous technological advances in the design of have been researched in the literature, the body system on chip (SoC), the development and use of wearable signals, and techniques used in these tasks? devices have remarkably achieved high growth rates in the Journal of Healthcare Engineering 3 last few years. *e wearable device market size was valued at are often used in a wide variety of applications to capture or USD 32.63 billion in 2019 and is expected to expand at high recognize body movement and activities, which can tell a lot rates in the next few years according to the wearable about the health and the lifestyle of the person [1, 25–39]. technology market industry report by Grand View Research Other signals that have been used sparingly in literature (https://www.grandviewresearch.com/industry-analysis/ include electrogastrogram (EGG), which records the elec- wearable-technology-market). trical activity of the stomach [40], and electrooculogram *e number of globally connected wearable devices is (EOG), which is generated by eye movements and can be about to reach 1 billion according to Statista (https://www. measured with electrodes placed around the eye [41]. statista.com/statistics/487291/global-connected-wearable- Sensory and especially olfactory inputs are challenging to devices/). Examples of wearable devices are smart watches, model, but it was observed that the human body has diﬀerent armbands, chest straps, shoes, helmets, glasses, lenses, rings, autonomic responses to diﬀerent odors, which can be an- patches, textiles, and hearing aids [6]. alyzed through the GSR and ECG signals [42]. *ese inputs Despite this grand growth, there is still a great need for can be used in applications including personalized treat- ongoing research in this area for enhancing the accuracy of ments based on odors and foods for neuropsychiatric and these devices, using diﬀerent body signals for new appli- eating disorders. Some other inputs may need visual cation areas, and dealing eﬀectively with the complexity of monitoring. Figure 3 summarizes the diﬀerent sensors used the human body. in the literature for diﬀerent machine learning healthcare tasks. Features are extracted from these signals to learn a model for either classiﬁcation or regression of a certain 2.1. Wearable Device’s Signals Used in Learning. *e human variable. Some literature studies use statistical values such as body can be seen in an abstract way as a group of systems mean, minimum, maximum, mode, variance, standard de- (circulatory system, nervous system, respiratory system, di- viation, entropy, and kurtosis. However, it is often hard to gestive system, etc.). It receives a group of inputs and releases a interpret how some of these statistical features aﬀect the set of outputs as shown in Figure 2. Inputs include the inhaled classiﬁcation or the outcome variable. Additionally, model’s air, water, food, visual input for all the scenes and objects seen accuracy is usually negatively aﬀected by adding more ir- during the day, auditory input for all sounds and voices heard, relevant features as more is not always better, and domain- sensory inputs for the things touched, and olfactory input for speciﬁc features that are expressive achieve better perfor- things smelled during the day. Outputs include the exhaled air, mance [43]. Estimating heart rate and breathing rate as excretions such as urine, feces, and sweat, skin moisture, body features from the PPG signal, change in acceleration mag- temperature, blood in case of injuries and laboratory tests, nitude, jerk of motion, and transient changes in skin re- energy released by the human in terms of body movements, or sistance for seizure detection are examples of domain- performing mental activities and voice output, which can be speciﬁc features. Some applications are concerned with normal speech, singing, or shouting. Analysis of the inputs and changes happening over a long time period, and some are outputscan, to some extent,predictthehealthstate of aperson, concerned with transient changes due to certain events such diagnose possible diseases/disorders, and assist with thera- as fall detection and emotion recognition. peutic suggestions. *ese inputs and outputs need to be monitored by wearable devices worn during the day. Wearable devices include any device mounted on the 2.2. Machine Learning for Wearable Devices. Machine body and can capture noninvasive signals from the human learning involves getting wearable devices to act/take de- body through the use of diﬀerent types of sensors. *ere are cisions without explicit programming for a speciﬁc scenario numerous well-known signals and signs that are read from through learning from past experiences. As it is well known, the human body in literature to identify the vital signs and machine learning is usually classiﬁed as either supervised, other information about the health or mental state of the unsupervised, semi-supervised, or reinforced according to subject. Examples of these sensors include skin temperature the type of the available training data. Learning from past sensor used in [7, 8] and electrodermal activity (EDA) sensor experiences is encoded in terms of data examples that can be or sometimes known as galvanic skin response (GSR) sensor either labeled or unlabeled. *e target variable for labeled used on the skin to record the skin conductance that varies data can be categorical or numerical. Among the tasks that with the sympathetic state of the subject [2]. Other examples involve machine learning are classiﬁcation in case of the include an electrocardiogram (ECG) sensor to capture categorical target output variable, regression in case of electrical changes in the skin corresponding to heartbeats numerical labels, and clustering for unlabeled data. Most used in [9–12]. To capture features of the electrical activity of machine learning research for wearable devices belongs to the brain and the health of muscles and the nerve cells, the classiﬁcation tasks, some are for clustering [44–46], and electroencephalogram (EEG) and electromyography (EMG) few can be tackled as regression problems [43]. sensors are used [13–19]. Blood volume pulse (BVP) can be Applied research to explore applying machine learning captured using an optical photoplethysmography (PPG) techniques using the body signals discussed in the last sensor to estimate heart rate and heart rate variation as in subsection for health monitoring, elderly care, and ﬁtness tracking has been growing over the last decade. Among the [1, 20–22]. PPG sensor [23] is also used to give an ap- proximation for the oxygen saturation in blood (SpO ) as in areas that got researchers’ attention are fall detection, seizure detection, vital sign monitoring and prediction [47], and [22, 24]. Accelerometers, gyroscopes, and magnetometers 4 Journal of Healthcare Engineering EEG EOG Skin Temperature Microphone Exhaled air Urine ECG Feces Sweat Air Skin moisture Water Temperatre Food tears Visual input Energy Auditory input (Body movement Sensory input + mental activities) Olfactory input Voice output Blood Accelerometer EDA signals EMG BVP Figure 2: Human body as a system and signals that can be used as a source of data for machine learning models. Eating Monitoring Fall Detection Activity Recognition Fitness Tracking accelerometer accelerometer gyroscope accelerometer accelerometer gyroscope magnetometer gyroscope gyroscope magnetometer microphone magnetometer magnetometer proximity sensor Seizure Detection Hydration Monitoring Arrhythmia Detection Sleep Monitoring EEG EDA, PPG, accelerometer ECG accelerometer accelerometer gyroscope, magnetometer PPG EDA gyroscope temperature acoustic sensor Temperature acoustic sensor PPG Stress Detection Emotion Recognition Disease Diagnosis Rehabilitation Tasks ECG EEG ECG EMG EDA EDA PPG inertial sensor Temperature PPG Temperature (accelerometer & PPG Temperature accelerometer gyroscope) Figure 3: Healthcare machine learning tasks and sensors used for each one in literature. activity recognition for ﬁtness tracking or identifying human machine learning technique(s), sensor(s), and dataset(s) daily activities. Additionally, wearable devices have been used in each study. researched for their use in stress detection and detection of heart rate arrhythmia and rehabilitation tasks. Tables 1–3 show the diﬀerent areas and a sample of the most recent 2.2.1. Fall Detection. *ree categories of fall detection re- research work done in each area. *e table also shows the search eﬀorts can be identiﬁed in the literature based on the Journal of Healthcare Engineering 5 Table 1: Machine learning research work for healthcare wearables for fall detection, activity recognition, eating monitoring, ﬁtness tracking, and stress detection. Research Task ML technique(s) Sensors/signals used Dataset(s) work J48 (96.7%), logistic regression 3D accelerometer and MobiAct (https://bmi.hmu.gr/the-mobifall- [48] (94.9%), MLP (98.2%) gyroscope in smartphone and-mobiact-datasets-2/) KNN (84.1), naive Bayes UMAFall dataset (https://ﬁgshare.com/articles/ Accelerometer, gyroscope, [49] (61.5%), SVM (68.25%), and dataset/UMA_ADL_FALL_Dataset_zip/ and magnetometer ANN (72%) 4214283) Temporal signal angle 12 features for 7 subjects performing 5 fall types Fall measurements Inertial measurement unit [30] detection (93.3%@200 Hz to 91.8% (IMU) (9 times each with 3 diﬀerent speeds) @10 Hz) KNN and RF SisFall dataset [51] Accelerometer and [50] (99.80% KNN and 96.82% for gyroscope (For falling and non-falling activities) falling activity recognition) SVM (97% F1 score and 99.7% Accelerometer and [52] Public fall detection dataset [27] recall) gyroscope CNN UCI-HAR dataset and study set Accelerometer and [25] (UCI-HAR dataset: 95.99%, gyroscope 21 participants and 6 ADLs study set: 93.77%) Locally linear embedding Accelerometer, [53] UCI-HAR dataset transfer learning magnetometer, gyroscope Tri-axis accelerometer, tri- Activity Sequence-to-sequence axis gyroscopes, Postures dataset, mini MobiAct, and UCI-HAR [26] recognition matching network magnetometer (depending dataset on the dataset) sEMG signals of the upper 6 males and 6 females for 3 motion states of [54] SVM: 90% limb by Delsys, virtual vehicle: left turn, stop, and right turn accelerometer Tri-axis accelerometer, tri- 6550 pieces of data for 4 activities: walking, [39] ATRCNN: 97% axis gyroscope sitting down, running, and climbing stairs A public dataset for performing diﬀerent [34] Proximity-based active learning 3D accelerometer activities including eating [34] One IMU and a proximity Two datasets: 12.5 hrs for 16 participants in Random forest (89.6% in the sensor on ear and one IMU semi-controlled setting with 6 labels and 3 hrs [55] laboratory and 72.2% outside on the upper back and a for each of 15 participants outside the laboratory the laboratory) microphone with chewing and non-chewing labels A public dataset for performing diﬀerent Eating [37] DBSCAN clustering 3D accelerometer activities including eating [34] monitoring A study dataset of 25 participants, 10 in a Random forest and DBSCAN Inertial sensor on the laboratory setting and 15 in the wild doing [56] clustering algorithm (average downside of the lower jaw diﬀerent activities including eating a meal of precision of 92.3%) diﬀerent food types Gyroscope and Gradient boosted decision tree [33] accelerometer in Apple 79 features for 16 subjects taking pills (80.27% accuracy) Watch Logistic regression (0.9356), random forest (0.9203), Study set of 39 participants with a total of extremely randomized trees 2 accelerometers (hip and [38] 55 days in which sport and jogging activities Fitness (ERT) (0.9177), and SVM ankle) were logged tracking (0.9328)—best accuracy reported in diﬀerent scenarios 3-Axis accelerometer and [57] L2-SVM 114 participants over 146 sessions 3-axis gyroscope Zephyr BioHarness for BN, SVM, KNN, J48, 2 participants with 324 instances [2] ECG RF and AB learning methods Shimmer3 GSR for EDA At rest and exercise sessions Neural network model (92% ECG, GSR, body Stress accuracy for metabolic [24] temperature, SpO2, glucose 312 biosignal records from 30 participants detection syndrome patients and 89% for level, and blood pressure the rest) HR and RR data for 44 children (26 with ASD LR (87% accuracy) and SVM [58] ECG sensor in a chest strap and 18 without ASD) while at rest (7 min) and (93%) while engaged in stressful tasks (9 min) 6 Journal of Healthcare Engineering Table 2: Machine learning research work for healthcare wearables for arrhythmia detection, seizure detection, rehabilitation tasks, and hydration monitoring. Research Task Techniques Sensors Dataset(s) work 14 subjects recordings for a 30-minute SVM and K-medoids clustering- [59] ECG and PPG sensors training session and a 30-minute testing based template learning session ECG sensor, PPG sensor [60] Deep learning (max 89% accuracy) Cleveland database on UCI (SpO2) ECG patch (from 91,232 single-lead ECGs from 53,549 [9] DNN (0.837 F1 score) iRhythm) patients 402 PPG recordings for 29 free-moving 50-Layer convolutional network subjects (13 with persistent AF) and the [61] PPG sensor (95% AUC) NSR dataset of 341 PPG recordings from Arrhythmia 53 healthy free-moving subjects detection PPG sensor in a ring-type 13,038 30-s PPG samples (5850 for SR [10] Deep learning (94.7%) device and 7188 for AF) Public available dataset from Computing in Cardiology Challenge (CinC) 2017 [11] SVM and bagging trees ECG (https://physionet.org/content/ challenge-2017/1.0.0/) 5878 deidentiﬁed audio recordings, totaling >rbin 34 hours from 5318 ResNet of 34 layers of 1D rectiﬁed [4] Acoustic recordings unique patients labeled by a majority linear unit vote of 3 cardiologists as heart murmur, no heart murmur, or inadequate signal SVM (97.31), RF (97.08), NB (95.08), UCI EEG sampled dataset for epileptic [62] K-nearest neighbor (90.01), and EEG seizures neural network (93.53) Accelerometer and SVM ((Sens > 92%) and bearable 135 patients with generalized tonic- [63] electrodermal activity FAR (0.2–1)) clonic seizures with 22 seizures from Empatica Embrace Accelerometer and 40 pediatric patients with generalized Seizure [64] Not mentioned electrodermal activity tonic-clonic seizures detection Two classiﬁers (the models are EDA and accelerometer 5,928 h of data of 55 convulsive [65] not mentioned) best sensitivity 95% from three wristbands Epileptic seizures from 22 patients and< 1 false alarm rate Temperature, 69 patients with epilepsy accelerometer [8] LSTM and 1DConv (total duration > 2311 hours, 452 Blood volume and EDA seizures) sEMG acquisition Muscle signals sEMG for 3 users doing 9 [66] SVM, RF module hand gestures 12 times IMU sensor module and 81654 samples for 10 people data, each K-means clustering, SVM, and [67] plantar pressure sample has 10 features calculated from Rehabilitation artiﬁcial neural network (ANN) measuring foot insoles 64 sensing nodes in the foot insole tasks Inertial features and anthropometric [68] Support vector regression (SVR) IMU in SportSole characteristics of 14 healthy subjects Multiple regression, inference tree, Two-sensor (fore and aft) Kinematic and pressure features for 30 [69] and RF insole (LoadsolTM) participants, each doing 120 steps SVM for drinking detection Acoustic sensor Frequency and cepstral domain [70] Gradient boosting decision tree for and inertial sensor Features are extracted from the signals activity recognition LDA, quadratic discriminant 51 hydrated samples and 17 dehydrated analysis, logistic regression, SVM, [21] EDA and PPG for 17 subjects with features from EDA Hydration Gaussian kernel, KNN, decision and PPG monitoring trees, ensemble of KNN ECG (not wearable (RR SVM (60%) and K-means clustering 10 minutes ECG for 16 athletes at rest, [71] interval, RMSSD, and (42%) post-exercise, and post-hydration SDRR recorded)) Shimmer (IMU, GSR, 3386 min for 11 subjects under fasting [43] DNN, RF, extra trees PPG, etc.) and non-fasting conditions Journal of Healthcare Engineering 7 Table 3: Machine learning research work for healthcare wearables for emotion recognition, sleep monitoring, and disease diagnosis. Research Task Techniques Sensors Dataset(s) work Liquid state machine (LSM)—above [16] 94% accuracy for valence, arousal, and EEG sensor DEAP dataset [72] liking recognition MUSE headband (EEG) and Emotion KNN (accuracy ranges from 53.6% [73] Shimmer GSR + device 54 subjects watching 24 pictures recognition to 69.9%) (SC and HR) Random forest, SVM, and logistic Respiratory belt (RB), PPG, and [74] regression—73.08% for arousal and DEAP dataset [72] ﬁngertip temperature sensor 72.18% for valence Auto-correlated wave detection with an adaptive threshold (ACAT), UCI-HAR dataset and study set of [1] Accelerometer and gyroscope accuracy for UCI-HAR dataset: 21 participants and 6 ADLs Sleep 95.99%, study set: 93.77% monitoring Accelerometer data during one night for 134 participants (70 with [32] Random forest (F1 score: 73.93%) Accelerometer in wristband sleep disorder and 64 good healthy sleepers) (MIMIC III) waveform database for ICU patients and a database of ResNet with LSTM for hypertension ECG, PPG, and invasive BP in patients with cardiac arrhythmias [75] detection ICU collected from Fuwai Hospital, Chinese Academy of Medical Sciences Everion wearable Disease (skin temperature, respiratory 200–1000 asymptomatic subjects diagnosis Machine learning techniques for early [76] rate, blood pressure, pulse rate, with close COVID-19 contact detection of COVID-19 blood oxygen saturation, and under quarantine in Hong Kong daily activities) Heart rate, heart rate variability, 34 patients with PCR-conﬁrmed Multivariate regression for case respiration rate, oxygen COVID-19 were admitted to the [22] deterioration saturation, blood pulse wave, isolation wards of Queen Mary skin temperature, and actigraphy Hospital used technology: (1) wearable devices that use accelerom- techniques is still questionable as the experiments were done in eters and magnetometers, (2) ambient devices such as ﬂoor a controlled environment with a limited number of participants and have the limitation of a high false alarm rate [77]. Another sensors and pressure sensors, and (3) vision-based devices that use monitoring cameras [77]. study to simulate fall data [31] was done to generate forward and syncope accelerometer data to form a larger dataset for fall In [50], the authors reported an accuracy of 99.80% using KNN classiﬁer and 96.82% for falling activity recognition using detection training. the random forest classiﬁer. Using tri-axial accelerometer de- vices [27], KNN, SVM, ANN, and RF classiﬁers were tested to get a mean average accuracy ranging from 48% to 98% 2.2.2. Activity Recognition. Activity recognition enables depending on the classiﬁer’s task. Some tasks involved dis- health professionals to get information about a patient’s tinguishing fall samples from daily activities’ samples. Other ability (or inability) to perform activities of daily living tasks were to distinguish between diﬀerent fall samples or (ADLs) as a measurement of their health status. Human diﬀerent daily activities. *e results showed that the classiﬁ- activity recognition has been researched using convolutional neural networks by the authors in [25], and an accuracy of cation results on raw data are better than depending solely on the magnitude as feature vector. On the contrary, the magnitude approximately 96% and 94% was achieved for the UCI-HAR performs better than raw data in the case of subject-indepen- dataset and their study dataset. However, the accuracy of dent evaluation. It was easier to distinguish between falls and no machine learning algorithms for activity recognition for falls and subject-independent evaluation testing showed that the human subjects greatly drops whenever a context of diﬀerent classiﬁer performance strongly depends on the subject data. *e data distribution compared with that of the training data is authors in [78] show the eﬀect of using an optimization confronted [53]. Personalized exercises may be inadequate technique to increase the accuracy of an SVM classiﬁcation to be directly used as training data for another subject so the model. authors in [53] applied a cross-subject transfer learning While the reported accuracy in most of the research done algorithm that can link source and target signals through the for fall detection is above 90% [28–30], the practicality of these construction of manifolds at the feature level. Another way 8 Journal of Healthcare Engineering with autism spectrum disorder (ASD) was investigated in to approach this problem is to build a personalized model for each subject, and this approach was investigated by the [58]. *e authors in [24] studied stress detection using a neural network for metabolic syndrome patients as the authors in [26] as they see that people perform activities in diﬀerent ways and that general models may average out increase in stress may result in chronic symptoms. important individual characteristics, besides that personal- Other mental disorders such as depression, anxiety, and ized models can learn from much fewer data and guarantee bipolar disorder [84] have also been studied in the literature better privacy for data collected from accelerometers and [85] using features from biosignals, eye sensors, micro- gyroscopes in wearable devices. Earable (ear-worn) devices phone, camera, or interactions with smartphone to assess can also be used for human activity recognition. *ey were social behaviors. found to have a superior signal-to-noise ratio under the inﬂuence of motion artifacts, and they are less susceptible to 2.2.4. Arrhythmia Detection. Heart rate tracking could be acoustic environment noise [79]. noticeably seen in some commercial wristband and smart Eating activity monitoring, sometimes also referred to as watches. *e detection of irregular heartbeats (arrhythmia) automated dietary monitoring (ADM), is essential for pa- is a relatively recent goal for commercial wearables. Fast tients’ diet assessment and following up with taking med- heartbeats (> rbin100 bpm) are called tachycardia, while ication [33] for elderly people by monitoring taking a pill slow heartbeats (< 60 bpm) are called bradycardia. Atrial activity. *is is considered an activity recognition task, but it ﬁbrillation is one type of arrhythmia that involves the rapid is added to a separate category “Eating Monitoring” in and irregular beating of the atrial chambers of the heart. Table 1. In [37], the authors proposed a proximity-based Apple conducted a clinical study to detect atrial ﬁbrillation active learning on accelerometer data obtained from a [3] in 419,297 participants using PPG sensors in Apple wrist wristband wearable device, which is a novel proximity-based watch patches, but they used non-machine learning algo- model to recognize eating gestures. In [36], the author rithm based on a proprietary threshold analyzed from data assessed using an EMG sensor and contact microphone for the degree of dispersion of inter-peak intervals to de- behind the ear near the jaw to record chewing sounds and termine irregularity. After a monitoring period and ana- detect eating activities. *ey used 8 features extracted for a 3- lyzing the results, participants with detected irregularities second window size for eating detection of crunchy and soft were notiﬁed to do ambulatory ECG monitoring using ECG food. A study for eating episode recognition [55] used two patches, of which only 34% responded (450 participants). IMUs, with one put on ear and the other one on the upper Similar to the clinical study done by Apple [3] to detect atrial back, and they trained a random forest with the sensors’ data ﬁbrillation, Huawei and Fitbit have recently launched their and labels. Another study [56] used features from inertial atrial ﬁbrillation study in mid-2020 (https:// sensor data placed on the downside of the lower jaw to detect cardiacrhythmnews.com/wearables-devices-the-new- eating episodes. A review of the research done until 2019 in frontier-in-arrhythmia-management/). the ﬁeld of eating detection comparing diﬀerent studies in *e authors in [59] used the SVM model to identify the terms of the used sensors, methods for collecting the data, raw heartbeats. *en, with an unsupervised dynamic time and evaluation metrics was discussed in [80]. *e authors warping (DTW)-based learning approach using the pointed out that most of the studies included accelerometer K-medoids clustering method, the distorted heartbeats are data from a wrist-worn device for accessibility and ease of identiﬁed and puriﬁed. SVM and bagging trees have been use, and they mentioned that the implementation of novel used in [11] to detect atrial ﬁbrillation from features from methods for naturally acquiring ground truth remains a ECG signals. challenge. A similar approach can be used for drinking In [10], the PPG signal was alternatively used. It was episode detection [81] and smoking cigarette detection [82]. recorded for patients with atrial ﬁbrillation using both a Fitness tracking is another application that can also be conventional oximeter and a cardiotracker ring, which considered as an activity recognition task. In [38], the au- generated comparable results. A convolutional neural net- thors were able to identify jogging periods using acceler- work achieved better results when compared to diﬀerent ometers and they concluded that there is no signiﬁcant SVM variants. A worst case accuracy of 94.7% was achieved beneﬁt from using accelerometers on both hip and ankle for 10-second recording periods. Although PPG signals have locations over using only one accelerometer. Segmentation limitations such as noise introduced by motion artifacts, the of exercise and non-exercise periods and recognizing which authors concluded that the ring PPG-based wearable has exercise is being performed were investigated in [57]. good diagnostic performance for atrial ﬁbrillation and can replace ECG-based detection. *ey also mentioned that considering longer periods for PPG signals may aﬀect the 2.2.3. Stress Detection. A survey for stress detection using performance due to false positives with atrial tachyar- diﬀerent signals such as heart rate (HR), blood volume rhythmia episodes. A deep learning model has also been pressure (BVP), inter-beat interval (IBI), electrodermal used in [60] but with the best accuracy of 89% achieved activity (EDA), temperature data, and behavioral features learning from both ECG and PPG sensor data. was conducted in [20]. *e authors found that the most distinctive features for detecting stress are EDA and HR. Remote monitoring of child safety through stress patterns 2.2.5. Seizure Detection. Epilepsy is a neurological disorder was tackled in [83]. Detecting stress and anxiety in children that aﬀects the central nervous system, causing seizures or Journal of Healthcare Engineering 9 [21, 70, 71]. In [71], the authors used heart rate variability periods of unusual behavior such as twitching in legs and arms and sometimes loss of consciousness. Detecting sei- (HRV) parameters: RR interval of the ECG signal, standard deviation of RR interval (SDRR), and root mean sum of zures is important to help the patient get help when needed by alarming a caregiver. Diﬀerent types of seizures and squares of diﬀerences between adjacent RR intervals wearable devices for detecting seizures are reviewed in [86]. (RMSSD), from the ECG signal as features with labels at rest, *e study in [63] included the highest number of partici- post-exercise, and post-hydration to detect dehydration. *e pants (135) and used Embrace Empatica Watch [64] with an authors in [21] used EDA and other heart rate variability accelerometer and EDA sensors. parameters extracted from PPG signal for mild dehydration Forecasting seizures can also be useful to alarm the identiﬁcation. *e authors in [43] used multimodal sources patient to rest and take protective measures. Seizure fore- of features from diﬀerent sensors to predict the last drinking casting has been investigated in [8] using deep learning on time of the user, which would ease collection of data and provide ways for personalization on-device. multimodal wristband sensor data from 69 patients with epilepsy (total duration >rbin 2,311 hours, 452 seizures). In [62], the authors investigated the use of support vector 2.2.8. Emotion Recognition. Emotional state monitoring for machine (SVM), random forest (RF), naive Bayes (NB), construction workers in a real worksite using a wearable K-nearest neighbor (KNN), and neural network (NN) to EEG sensor [15] was classiﬁed as positive (e.g., excitement, diagnose an epileptic seizure based on EEG sampled dataset happiness, contentment, or satisfaction) or negative (e.g., available at the UCI machine learning repository. Similarly, fear, anger, frustration, or depression) based on measuring the authors in [64, 65] have used both EDA and acceler- the EEG valence level and cortisol biochemical response as a ometer data for detecting seizures but with diﬀerent datasets reliable marker tested from saliva samples after each task. and techniques. *is can better be performed using machine learning techniques to replace the cumbersome cortisol testing. 2.2.6. Rehabilitation Tasks. Rehabilitation tasks involve Considering fear emotional state, fear level classiﬁcation tasks to improve abilities needed for daily life, which may be using diﬀerent machine learning techniques (KNN, RF, physical or mental abilities that have been lost or impaired LDA, SVM, and deep learning) has been researched in [89] depending on features extracted from EEG, GSR, HR, and due to injury, underlying disease, genetic disorders, or birth defect. One example for a rehabilitation task is foot strike subjective unit of distress (SUD) values in a virtual reality therapy setting. Using EEG raw data, the authors in [16] angle prediction, which was studied in [69], which can help in the coaching of running movements and consumer-based introduced using a liquid state model (LSM) for training to predict valence, arousal, and liking levels at diﬀerent du- shoe prescription. Diﬀerent machine learning techniques were compared, and random forest achieved the best ac- rations of the EEG input signal. curacy of 94.1%. Linear discriminant analysis (LDA) was used in [87] to 2.2.9. Sleep Monitoring. For sleep monitoring and sleep classify each subject as a normal or abnormal gait pattern. quality assessment, detecting diﬀerent sleep states (awake, *e authors used real-time acoustic feedback (RTAF) to rapid eye movement (REM) sleep stage, and non-REM support the subjects when they are performing the tasks stages) is a requirement. Classiﬁcation of sleep episodes has during the rehabilitation session, so that they are able to been studied in [32], where a random forest model was used adjust their motion pattern to the acoustic feedback. Support to detect diﬀerent sleep-wake states with an F1 score of vector regression (SVR) models yielded excellent intraclass 73.93% after being trained with the data from accelerometers correlation coeﬃcients (ICC) in the gait parameters (stride on the wrists of 134 subjects. length, stride velocity, and foot clearance) analyzed in [68] Sleep monitoring applications such as detection of sleep for both walking and running exercises. Similarly, the au- apnea episodes have been studied in [1]. Sleep apnea is a thors in [67] investigated the same problem on a diﬀerent problem accompanied by increased cardiovascular risk and dataset using the K-means clustering, SVM, and artiﬁcial decrease in the quality of life. *e authors compared auto- neural network (ANN). correlated wave detection with an adaptive threshold (ACAT) for both electrocardiogram (ECG) data and PPG 2.2.7. Hydration Monitoring. Hydration monitoring to sensor data to detect the cyclic variation of heart rate detect dehydration is another problem being researched for (CVHR). *e classiﬁer was able to discriminate sleep apnea its importance, especially for athletes, battleﬁeld soldiers, episodes from non-apnea episodes with 82% sensitivity, 89% workers in hot conditions, and elderly people who are not speciﬁcity, and 85% accuracy depending on PPG signals. able to communicate their need for water. *ere is an on- Electrodermal activity, accelerometer data, heart rate going progress in the development of biochemical sensors variability, and blood volume pressure during sleep have been used in [7] using wearable Empatica E4 smart watch for that can measure the concentration of diﬀerent electrolytes in sweat and hence determine the hydration state [88]. Side early detection of migraine from the quality of sleep to enable early alarms to take preventive medication. *ey by side, machine learning research studies try to learn from other diﬀerent body signals to detect dehydration based on achieved a balanced accuracy of over 84% for detecting the eﬀect of cognitive stress triggered by dehydration on the migraine attacks using quadratic discriminant analysis as a autonomic reactions of the body as the work done in classiﬁer. Another dataset is published in [90] that can be 10 Journal of Healthcare Engineering they use diﬀerent datasets, rely on diﬀerent features, and solve used for sleep stage prediction for which accelerometer and heart rate data were collected from Apple Watch, while the diﬀerent problems with multiple experiments. In Figure 4, a box plot is shown for the range of values reported for accuracy subjects underwent polysomnography during night sleep and it was used for sleep-wake classiﬁcation in [91] using of classiﬁcation models in research studies cited in Tables 1–3 ANN with accuracy 90%. under 5 group models (KNN, SVM, logistic regression, tree- based models (random forest, decision trees, extremely ran- (i) Disease diagnosis artiﬁcial intelligence research has domized trees), and deep learning (DNN, MLP, LSTM, been in health care and medical diagnosis of diseases CNN)). It can be seen that the best median average accuracy for a long time ago. Starting from expert systems in achieved is for using deep learning. Logistic regression, SVM, the 50 s of the last century, continuous eﬀorts have and tree-based models follow deep learning with very close been going on in this ﬁeld until recently applying values. All models are away from perfect classiﬁcation, but deep learning techniques for improved diagnosis of some are useful and there is a room for improvement over all diseases [92]. Examples include lung cancer diagnosis tasks using larger datasets, extracting more meaningful fea- based on CT scans and diagnosis of skin conditions tures, and modeling for personalization as body signals vary through scanning skin images, which has recently according to each person lifestyle, weight, height, and activity been announced by Google [93]. *e use of artiﬁcial level. intelligence and machine learning techniques through wearable devices for initial assisting diag- nosis and detection of symptoms is foreseen to be the 2.3. Datasets. PhysioNet (https://www.physionet.org/about/ upcoming future, especially in the COVID-19 pan- database/) is a big database that oﬀers large collections of demic circumstances we are passing through and the physiological and clinical data and related open-source quarantine protective requirements imposed in most software for research purposes in many areas such as sleep of the countries. apnea detection, arrhythmia recognition, stress detection, and human activity recognition. It was established by the With the onset of the COVID-19 pandemic, the authors National Institutes of Health (NIH) and is maintained by in [76] proposed a protocol for using a mobile health MIT Laboratory for Computational Physiology. platform to analyze the biosignals recorded by Everion wearable (skin temperature, respiratory rate, blood pressure, *ere are many other datasets available for human ac- tivity recognition and fall detection that are mentioned in pulse rate, blood oxygen saturation, and daily activities), together with a recording for the cough for early detection of [27]. MobiAct [48] (57 subjects/9 ADLs and 4 fall types) is the largest of them in terms of the number of subjects and is COVID patients. For the limited places for patients to re- ceive hospital care, which was observed by the spread of suitable for both fall detection and human activity recog- nition tasks. It is available upon request for research pur- COVID-19 in some countries, the researchers in [22] used poses. UCI-HAR dataset [96] provided by the University of machine learning-based analytic systems to detect early California Irvine is the most famous and cited dataset in the signs of clinical deterioration to schedule and guarantee domain of human activity recognition. *ey recorded the 3- resources’ optimization. *ey also used Everion biosensor to axis accelerometer data and 3-axis gyroscope angular ve- record many physiology parameters such as heart rate, heart locity time series at 50 Hz for 30 subjects while doing 6 rate variability, respiration rate, oxygen saturation, blood pulse, skin temperature, and actigraphy to monitor mild activities (WALKING, WALKING_UPSTAIRS, WAL- KING_DOWNSTAIRS, SITTING, STANDING, and LAY- COVID-19 cases and predict clinical deterioration accord- ingly. Hypertension diagnosis has also been studied in [75] ING) using smartphones banded on the waist. Friedrich-Alexander-University oﬀers many movement using deep learning for continuous monitoring of blood analysis datasets (https://www.mad.tf.fau.de/research/ pressure depending on one-channel ECG and PPG signal activitynet/) such as daily life activities, step activities, cy- that can be obtained from a wearable device. clic activities, gait analysis datasets, and energy expenditure *e use of electronic monitoring devices for asthma has estimation. A dataset of 3D accelerometer data speciﬁcally been reviewed in [94]. *e authors suggested that clinicians for eating activity recognition for 20 participants in labo- should evaluate asthma management applications to ensure high-quality and evidenced-based information before pa- ratory setting and 7 participants in free-living conditions is made available for research purposes by the authors of [56]. tients use them as current studies only analyzed asthma patients according to their sleep quality and physical activity A recent real-life human activity dataset was published by the University of A Coruna [97]. *ey recorded about measures. In [95], multiple features of motion and dexterity and 189 hours of measurements from the accelerometer, gyro- scope, magnetometer, and GPS of smartphones for 19 dif- sleep measures were collected using IMU sensors on the ferent subjects with no restriction for mobile position. *e chest, wrist, and ankle to correlate these measures with data have four labels that deﬁne diﬀerent activities (inactive measures of neurological disability in multiple sclerosis. for not carrying the mobile phone, active for carrying the As can be seen from Tables 1–3 and the review in this mobile phone and moving (making dinner, being in a subsection, no single model can be chosen for every problem as concert, etc.), walking for moving to a speciﬁc place whether it depends on the dataset size, the features extracted, and the problem being learned. It is very diﬃcult to compare diﬀerent jogging or running, and driving for moving in a car, bus, truck, etc.). techniques and the results reported in the research studies as Journal of Healthcare Engineering 11 k-NN SVM LR Tree-based Deep learning Machine learning models Figure 4: Box plot of accuracy for the machine learning techniques used in diﬀerent classiﬁcation problems for papers cited in Tables 1–3 with accuracy as the evaluation metric. On each box, the central mark is the median and the edges of the box are the 25th and 75th percentiles. *e small circles represent outliers. Sleep Data (https://sleepdata.org/datasets) have large 3. Challenges for ML Applications on collections of de-identiﬁed physiological signals and clinical Wearable Devices data elements that are oﬀered by the National Sleep Research Resource (NSRR) to help in sleep monitoring research. Some Developing machine learning applications in general follows of these signals can be obtained through wearable devices the Cross-Industry Standard Process for Data Mining Cycle using diﬀerent sensors’ types. (CRISP-DM 1999) [100]. *e development-to-deployment RecoFit [57] contains accelerometer and gyroscope re- process involves many challenges in collecting the data, cordings from over 200 participants performing various gym selecting the best features, selecting the libraries and exercises. framework [100], evaluating the trained model(s), selecting *e seizure gauge dataset (https://www.epilepsyeco the best model, and relying on the ML model decision since system.org/my-seizure-gauge-1) records long-term physio- practically no ML model is guaranteed to be 100% accurate. logical signals such as EMG, PPG, EEG, ECG, accelerometer ML learning models for health care are to be designed to signals, BVP, EDA, and temperature from diﬀerent people generalize well and deal with unseen examples while taking with epilepsy using three diﬀerent wearable devices. into account personal features, providing interpretation for iRhythm arrhythmia detection public test dataset (https:// the result, and communicating the results cautiously. Some irhythm.github.io/cardiol_test_set/) is a dataset used in [9] for issues can be handled through clinical and preclinical testing a model used for arrhythmia classiﬁcation for 336 studies, to provide a suitable user interface and a note for records of 30 s strips single-lead ECGs captured at 200 Hz from conﬁdence or reliance on the results to be provided as per 328 patients who used a single-lead ambulatory ECG moni- the regulatory requirements. *e model needs to be toring patch. Each record is annotated by a consensus label implemented and used in both retrospective and prospective obtained by a committee of three cardiologists. studies, and the clinical impact measured [101]. A database for emotion analysis using EEG signals (https:// In addition to the typical challenges facing any machine www.eecs.qmul.ac.uk/mmv/datasets/deap/) and peripheral learning application concerning the used data and model, physiological signals was collected, while the participants there are many challenges that developers of a machine watched 40 one-minute music videos [72]. A dataset for learning application for a wearable device should take care studying social stress using blood volume pulse (BVP) and of, all challenges are shown in Figure 5 and are presented in electrodermal activity (EDA) signals has been recently pub- the next subsections, and how they aﬀect the choices lished [98]. Cognitive load, eﬀect, and stress recognition have available for developers. been studied in [99] through recording the biosignals (ECG, PPG, EDA, and accelerometer data) of 62 healthy volunteers whileansweringmathproblems,logicproblems,andtheStroop 3.1. Data Availability and Reliability. Machine learning test. *e Stanford wearable dataset (http://ipop-data.stanford. approaches, in general, and especially in certain healthcare applications, require the availability of enough data for edu/wearable_data/Stanford_Wearables_data.tar) was used in [9] for arrhythmia detection and classiﬁcation. training to generalize well for unseen data. As presented in Accuracy 12 Journal of Healthcare Engineering To ensure data reliability, conducting a wide range of clinical experiments while reporting the results transpar- Data ently is critical for evaluating diﬀerent techniques [103] and Availability & ﬁnding promising research directions. Medicolegal aspects Reliability Model Storage need to be well-deﬁned and regulated [104]. As an example, Selection & Limitations Reliability the authors in [105] provide guidelines for future study data collection and design for heart rate data, data cleaning and processing, analysis, and reporting that may help alleviate the data reliability challenge. Power Challenges to ML Consumption Application Deployment Limitations on Wearable alternatives 3.2. Model Selection and Reliability. For reliably calculating the accuracy of machine learning models, the use of cross- validation techniques is considered one approach to achieve this by testing the model on unseen data that have not been used in training. In [106], the authors reviewed the research Security Communication work of using either record-wise or subject-wise cross- Utility Privacy & user validation. *ey experimented with a publicly available acceptance dataset for activity recognition and simulation data to ﬁnd that using record-wise cross-validation overestimates the prediction accuracy of machine learning algorithms. *is Figure 5: Challenges to healthcare ML applications on wearable result agrees with the research ﬁndings in [27]. Diﬀerently, devices. some of the authors in [107] criticized the work done in [106] with arguments that no within-subject dependence between observations can be detected so record-wise cross-validation Tables 1–3 in the last section, the study with the maximum can be used. *e authors also suggest avoiding leave-one-out number of subjects [9] included 53,877 patients in a retro- cross-validation and recommend keeping the number of spective study that was funded by a commercial company for folds large enough while following strategies such as re- manufacturing ECG patches. *e rest of the studies depend peated test-train split, shuﬄe-split, repeated K-fold, or on far fewer data since the health data collection is expensive, Monte Carlo cross-validation to avoid overﬁtting and ensure which makes their results questionable for reliability. generalization. For wearable devices, since the models *e data obtained from wearable technologies have to be usually represent. reliable as well with deﬁnite conﬁdence and clear warnings to For model selection, there are many criteria that aﬀect ask for medical staﬀ help for any concern as human health is the the decision when it comes to wearable devices [43]. One of ultimate goal. *e authors in [102] investigated the sources of them is to maximize the evaluation metric used to report inaccuracy in diﬀerent wearable optical heart rate sensors. *ey accuracy for classiﬁcation or minimize the error metric used explored heart rate and PPG data from consumer and research- in regression problems. Usually training an ensemble of grade wearables while doing diﬀerent activities for diﬀerent skin diﬀerent models achieves the best performance. Interpret- tone subjects. According to their ﬁndings, there was statistically ability or explainability of the model [108] is another cri- no signiﬁcant diﬀerence in accuracy across skin tones, but terion as most of the applications for wearable devices target signiﬁcant diﬀerences between devices and between activity healthcare application for which the result of classiﬁcation/ types were remarkable with an average absolute error of 30% regression or clustering is to be explainable and makes sense more than during rest. *e reliability of data in health care is so for the user. Tree-based models are seen as more inter- important for the patient and physician to rely on the device pretable than neural network-based models [43]. *e size of readings to take the most appropriate medical decision, which the model to ﬁt on the wearable device with limited memory may in some cases threaten the life of a human. *is what led is among the criteria. Additionally, the computational Verily Life Sciences (formerly Google) to discontinue their complexity for inference and for online training on the glucose-sensing lens project (https://www.business-standard. device for personalization is a concern due to the limited com/article/news-ians/google-halts-project-to-build- computation power for wearable devices until now. On- glucose-sensing-contact-lens-118111800398_1.html) when device deep learning and transfer learning for personali- their ﬁndings reveal that there is insuﬃcient consistency in the zation have been researched in [43, 109–111]. correlation between tear glucose and blood glucose concen- trations. *eranos is another example that went dreadful after being charged for wire fraud (https://www.fda.gov/inspections- 3.3. Deployment Alternatives. *ere are three deployment compliance-enforcement-and-criminal-investig alternatives for the machine learning model for the wearable ations/press-releases/june-15-2018-theranos-founder-and- device scenario, either to deploy the model on the wearable device, or on an edge device or on the cloud as shown in former-chief-operating-oﬃcer-charged-alleged-wire-fraud- schemes) when investigations found they advertised rapid Figure 1. Each deployment alternative has some advantages and disadvantages that might make it impractical in some blood test devices that they knew were likely to contain inac- curate and unreliable results for diﬀerent blood tests. cases. Deploying the machine learning model on the device Journal of Healthcare Engineering 13 alternatives and maintaining the continuous update of the has the advantages of keeping the data private and de- creasing the latency for the prediction/classiﬁcation as there ML cycle. A survey for the diﬀerent automated machine learning approaches to automate feature extraction and is no need to transmit large amounts of data from the device to the cloud. In particular, for healthcare applications, selection, hyperparameter optimization, pipeline optimizers, having the patient’s data and the machine learning model on and neural architecture search for healthcare systems can be the device is more safe from privacy-preserving perspective. found in [115]. Low-latency and real-time feedback is also required for many healthcare applications that require immediate alert for the users or their caregivers such as fall detection and 3.4. Power Consumption. Power consumption is the main limitation of wearable devices due to their limited battery stress detection. On the other side, the main disadvantages of on-device computing include the limited device computing lifetime in general. For machine learning applications on wearable devices, the power consumption is greatly aﬀected by power, storage, and battery life. With these limitations, oﬄoading computations to be the need to send physiological data measured by the device’s sensors to the cloud to perform computations on the cloud. At done on one or more edge devices (e.g., smartphones or locally on a hospital/house/oﬃce terminal/gateway) is one the time of writing this study, the best commercial smart watch solution [112]. Edge/fog computing has also many advan- battery lifetime is just a few weeks, which monitors walking and tages over the cloud computing alternative [112] in terms of running activities and give an approximate measure of the pulse security, latency, power consumption, real-time processing, rate and oxygen saturation. *is could be far less in practice and and bandwidth load [113]. Edge computing can reduce data could be as low as a few hours for wearables that monitor transmission to the cloud and consequently reduce power multiple vital signs continuously for alerting users to abnormal situations (e.g., alerting for abnormal heart rhythm or detecting consumption and improve privacy by analyzing sensitive private data on a local gateway, ﬁltering it, and compressing fall). *e elements that aﬀect the power consumption in it, instead of doing it on a cloud away from the user’s control. However, this depends on the size of the machine learning wearable devices include the board, its components of diﬀerent biosensors and their sampling rate, the operating model and the data streams to be used in training and testing, the need for online training and real-time prediction, system and other software running on the board, the and the computational power needed for training and wearable display, the rate of logging data on the device, and testing. the amount of data transmitted over the communication *e main advantage of adopting cloud computing is the channel (e.g., Bluetooth or Wi-Fi) to be sent to the edge/ ﬂexibility of storage and computational resource on-demand cloud. scalability. *is comes as a trade-oﬀ for higher costs, power Transmission and reception of data are thought to consume more energy than sensing and logging data. consumption, latency, and challenges for preserving the privacy of both the data and the machine learning model as Research in the area of reduction in the power con- sumption can be seen to go in diﬀerent directions, de- will be discussed later. Data drift (how data distribution could change over veloping special embedded hardware for running time) and continuous integration and delivery are other machine learning algorithms [116, 117], reducing data to aspects that determine the decision of which deployment be transferred [118–120], compression [121] or sched- alternative to employ in machine learning applications for uling of the data to be transferred [122], computational wearable devices. oﬄoading [123, 124], and developing self-powered *e development process for wearable machine learn- wearable devices [125, 126]. ing-based software requires the same operations for any One approach suggested by the authors in [127] to software with some speciﬁc operations related to machine save the consumed power by the data transfer over the wireless connection is to perform embedded machine learning applications such as data collection, cleaning and preprocessing, continuous (re)training of the ML model, learning on the device, i.e., following the tinyML ap- proach. According to the analysis in their work, this can and continuous (re)-deployment of the updated model to the device or to the edge nodes or to the cloud service [114]. increase the battery lifetime by more than 70%. Re- Figure 6 shows the typical machine learning operations searchers in [128] proposed a hybrid approach of using (MLOps) in the development process of wearable software less battery, low sampling rate, and wearable RFID tags, for ML-based applications with the diﬀerent deployment which can be powered intermittently by a reader with alternatives (device, edge node, cloud service). *e feedback additional passive RF tags that capture the presence and arrows from the deployment process are orchestrated based use of speciﬁc objects for daily activities’ recognition. on the performance of the model on edge nodes or cloud As previously mentioned, another way to reduce service after getting feedback from users or updated models power consumption is to reduce the data stored and transmitted to the cloud, and the authors in [129] pro- received in case of federated learning scenario, which will be discussed in Section 3.8.3 to ensure continuous integration posed a variant of symbolic aggregation approximation (SAX) tested for compressing heart rate data, which (CI) and continuous deployment (CD) requirements. Tools such as Apache Airﬂow, Kubeﬂow, and Google Cloud proved to achieve the best trade-oﬀ between diﬀerent AutoML support the software lifecycle operations of ML performance metrics for systems that require short components by orchestrating the diﬀerent deployment latency. 14 Journal of Healthcare Engineering Orchestration Containerization Scope Collect Train Deploy and and Extract machine model in design clean features learning production the project the data model Cloud Edge Device Figure 6: MLOps for wearable device application. 3.5. Storage and Memory. Typically, existing wearable de- *us, research goes on in many directions to overcome vices have limited memories (e.g., Apple Watch Series 6 these factors. From the data perspective, data selection and released in September 2020 has only 1 GB RAM) due to dimensionality reduction techniques are employed. From the small device size and weight requirements. Wearable and model perspective, designing new models with acceptable IoT devices use nonvolatile memory (e.g., ﬂash, EEPROM, prediction accuracy while minimizing model size and pre- diction costs such as Bonsai [132] is another approach. Com- MRAM, and F-RAM) to ensure resilient system recovery on sudden shutdown with the limited battery lifetime and to pressionofmodelscantakeplacebypruning(usinglessnumber of weights), quantization (using less bits per weight) [133], and ensure short boot time. While ﬂash-based storage is con- sidered a de facto storage standard for IoT devices for its encoding. *e authors in [134] reviewed model compression speed and stability [130], F-RAM is commonly used for techniques. Some of these techniques are implemented in medical wearables for its low power operation and high- TensorFlow Lite (https://www.tensorﬂow.org/lite). write cycle endurance, which allow it to reliably and eﬃ- ciently store more data logs from sensors [129]. EEPROM is sometimes also used since it is more reliable and smaller 3.6. Utility and User Acceptance. Users of wearable devices have been growing over the past few years, especially ﬁtness than ﬂash memories for applications that do not require frequent write operations and requires less power. In [131], trackers. Nevertheless, there is still a lack of user acceptance to adopt other wearable devices incorporating AI solutions the authors proposed using battery-backed RAM on for healthcare tasks. wearable devices and eﬃciently oﬄoad energy-intensive According to [135], 35% of 1,183 adult patients in France tasks to the smartphone/edge device to perform small and would refuse using wearable monitoring devices and AI- energy-eﬃcient tasks locally using battery-backed RAM. In addition to the development of memory archi- based tools in their care. Another study in the United States [136] examined the response of 307 consumers to the per- tectures (in-memory computing) and hardware (appli- cation-speciﬁc integrated circuits (ASICs)) that are ceived beneﬁts and risks of AI medical devices with clinical decision support (CDS) features. *e results of the study show capable of running machine learning applications on battery-operated devices, tinyML Foundation (https:// that performance/accuracy and communication, besides the ethical and regulatory concerns to keep the data private and www.tinyml.org/), which started in 2019, has also fo- secure, signiﬁcantly contribute to the perceived risks of using cused on signiﬁcant progress on algorithms, networks, AI applications in health care. Regulatory agencies should and models down to 100 kB and below to perform on- establish a standard and evaluation guidelines for the device analytics at extremely low power, thus minimizing implementation and use of AI in health care. Privacy and bandwidth and latency concerns while providing higher security concerns are among the major concerns raised for the privacy. *e practicality of deploying a machine learning application use of wearables. For example, there are security concerns raised for using Google Glass for recording people data on a wearable device or an edge device depends on many factors: the size of the device, the data size (features and time without their permission. It has been proven to be a serious issue since it can be used (like any recording device) to steal span of physiological data used for prediction), the complexity of the model (no. of parameters and layers), and use of batch or passwords by recording and analyzing the shadows of ﬁnger movements on a screen while typing a password [137]. *us, real-time processing. A model with high accuracy often requires the ﬁrst version of Google Glass failed to gain social accep- more memory for the number of parameters and layers in the tance [138] before releasing its second version and funding model than lower accuracy models. Depending on the machine some research studies about its usability, for example, its learning application, some machine learning models can reach desirability for a sample of school children with autism [139]. up to an order of 100 megabytes or even gigabyte (speciﬁcally Another important factor for user acceptance is how those including image inputs), which cannot ﬁt on the best wearable device along with the memory needed for doing the comfortable the device is for daily use. Design guidelines for wearable devices are identiﬁed in [140]. For example, computations. Journal of Healthcare Engineering 15 [144]. For example, accelerometer and gyroscope data on a designing a wearable should follow the anatomical structure of the body, take into consideration diﬀerent gender re- smart watch can be analyzed to reveal passwords and credit card information (https://securelist.com/trojan-watch/ quirements, and choose materials that are comfortable for the body and do not cause irritation to the skin. Addi- 85376/). Other attacks on IoMT devices can be life-threat- tionally, it is preferable to be used in a free-moving envi- ening such as attacks disrupting the medical service, e.g., ronment and it is required to be as easy as possible to use denial-of-service attacks (DoS) and ransomware attacks. without the need for many setup and conﬁguration steps. Whether wearable devices are used for health monitoring or *us, a wearable device should be compact and simple to for ﬁtness tracking, sensors’ data and other personal data are operate and maintain while providing secure and private being exchanged and analyzed by machine learning services experience for both the wearer and the people around him. to detect patterns and do classiﬁcation/prediction based on More awareness endeavors of the wearable technology to the the data. While it seems to be “a no problem to share” for some users, most end users are skeptical about how their public need to take place and the advertisers should abide by honest marketing about the product’s actual impact. personal data exchanged with such services is being used and how secure they are against diﬀerent types of attacks. *e issue of security and privacy of personally identiﬁable in- 3.7. Communication. In case of edge computing model, the formation and medical data in wearable and other IoMT intra-communication between the wearable device and the devices’ applications is critical and is regulated by diﬀerent edge device can be done over one of the standards such as data protection standards across the globe. Bluetooth, Zigbee, RFID, NFC, and UWB. Usually, light- In the case of wearables, the connection is usually done weight Bluetooth is employed for its low power consump- over lightweight Bluetooth as mentioned earlier and as tion [141]. However, according to Bluetooth 5 speciﬁcation, shown in Figure 1. Security guidelines for Bluetooth pro- the Bluetooth protocol allows up to 7 devices’ simultaneous vided in [145] consider wearable sensor devices as “Class 1.5 connections to a device and practically performance de- Low Energy” devices with a maximum output power of grades and pairing problems arise when there are multiple 10 mW that can operate for up to 30 meters distance but are connections to a smartphone. Other factors that aﬀect the typically used within 5 meters. *e guidelines show that for choice of communication technology are the maximum this class, each service request can have its own security distance between the wearable and the edge device, the requirements. It recommends the use of Security Mode 1 required data rate for the wearable-to-edge device, and the Level 4 for medical devices, which requires low energy secure required latency [142]. *e intercommunications in the connections authenticated pairing and encryption using wearable model over the Internet run between the edge AES-CMAC and P-256 elliptic curve to the edge device. device and the remote service or directly between the *e main challenge for edge computing is to incorporate wearable device and the remote service are two-way data security into the design of wearable devices through using communication channels over transmission control proto- encryption and providing solutions to manage, update, and col (TCP) or user datagram protocol (UDP) at the transport secure the wearable devices. Security risks include but are layer with the Internet protocol (IP) at the network level. not limited to malicious hardware or software injections, TCP/IP is mostly adopted for lossless transmission of health denial-of-service attacks, and diﬀerent routing and physical data or machine learning model parameters over wide area attacks. Some of these attacks can be defended using ap- network (WAN). propriate administrative policy settings and incorporating At the application layer, hypertext transfer protocol diﬀerent ML-based solutions for detecting diﬀerent attacks (HTTP) is commonly used as the request-response model that may compromise the communication network, com- from the edge to the cloud services. TLS is often employed to putations, battery consumption, or storage [146]. secure HTTP communication over TCP; however, HTTP is Additionally, securing the data stored on the cloud, resource-intensive and is more suitable to be used for edge which is fed to the machine learning inference model, and or fog devices with high power and storage capabilities. securing the model itself represent a big challenge [147]. Not Other less-weight application layer protocols include con- only the medical data itself and the machine learning model strained application protocol (CoAP), message queuing are considered prone to privacy attacks, but also the social telemetry transfer (MQTT), and advanced message queuing dynamics and interactions with other users can be analyzed protocol (AMQP) [143]. MQTT is a well-known publish- as done in [148]. subscribe model standard used for IoT and wearable devices Potential solutions for privacy-preserving ML are dis- for being a lightweight protocol. It can facilitate one-to- cussed in detail in [149, 150]. *ese include techniques for many communications between wearable device(s) with low achieving diﬀerential privacy, cryptographic techniques, and power and storage and the edge device on the other side. client-based federated learning techniques. *e following *e two communication channels with their running provides a brief discussion of these methods. protocols at diﬀerent network layers are susceptible to the diﬀerent well-known network security attacks. 3.8.1. Diﬀerential Privacy. *e diﬀerential privacy concept 3.8. Security and Privacy. User data captured on wearable was ﬁrst introduced in [151] and refers to the process of devices and sent to machine learning cloud services as shown protecting private data by adding noise based on Laplace, in Figure 1 are subject to many security and privacy threats exponential, or Gaussian distributions. *e noise is added in 16 Journal of Healthcare Engineering technique to use a clustering model over encrypted such a way that enables data analytics while providing privacy guarantees of the perturbed data. Diﬀerential privacy data. *ey employed the mean-shift algorithm and homomorphic encryption for the arithmetic of ap- can be useful for applications such as health care due to its useful properties such as group privacy, composition, and proximate numbers. To overcome the computational robustness to auxiliary information. With diﬀerential pri- load of the mean-shift algorithm, they performed vacy, healthcare applications that employ machine learning each iteration on a sample of the data instead of the algorithms can still learn from the distribution of data whole dataset. without revealing the actual data of the patients. However, (2) Trusted Execution Environments (TEEs): TEE is a researchers in [152] concluded that privacy compromises secure area located inside the main processor in must be made to preserve utility, especially in the chal- particular architectures. It ensures the conﬁdentiality lenging multi-class classiﬁcation tasks based on their ex- and integrity of the data and code within the TEE. periments on two datasets with membership attack and Examples of TEEs are Software Guard Extensions attribute inference attack. *is utility-privacy trade-oﬀ has (SGX) from Intel and TrustZone from Arm. Intel’s also been discussed in [153], where the authors found that as SGX provides a trusted execution environment, the privacy level increases, the machine learning algo- called an enclave, which trusts only the CPU and the rithm—diﬀerentially private stochastic gradient descent in on-chip cache [166]. A user program (code and data) their case—targets the body of the distribution but loses must be partitioned into an untrusted portion and a important information about minority classes such as dying trusted portion that will run inside the enclave. SGX patients and minority ethnicity that are usually represented protects the conﬁdentiality and integrity of code and in the tail of distributions. data during execution within the enclave from Another challenge for practically using diﬀerential pri- malicious programs that may be running alongside vacy in healthcare wearable applications is that it is best used it, including privileged programs, such as the OS and for high-dimensional balanced big datasets. *is is not the hypervisor. Hunt et al. [167] employed the SGX to case in some personalized healthcare wearable applications build their system for privacy-preserving outsourced such as a fall detector, which only learns from accelerometer machine learning called Chiron to protect the signals where falls are considered of low frequency. training algorithm and the user data. Segarra et al. [168] employed SGX to present a secure streaming processing system speciﬁcally ﬁtted for medical data. 3.8.2. Cryptography-Based Methods. Traditional cryptogra- phy is valuable and eﬃcient to achieve conﬁdentiality when (3) Secure Multiparty Computation (SMPC): SMPC used in secure communication between parties and out- oﬀers cryptographic protocols in which the com- putation is distributed across multiple parties where sourcing the data for storage, but it is not valid when we need to perform the computation on conﬁdential data as it needs no individual party can see the other parties’ data preliminary data decryption. Here, we introduce some [169]. Two common approaches to achieve SMPC methods employed to perform computations on sensitive are garbled circuits and secret sharing. data without violating privacy. (i) Garbled Circuits: in this technique, two (or (1) Homomorphic Encryption (HE): the idea behind HE more) parties can jointly evaluate a function over is to use special encryption functions that enable the their private inputs [170]. *e main idea behind computation of encrypted data [154]. HE ensures this technique is to use a Boolean circuit to that the result from performing operations on represent the function that needs to be evaluated. *e gates of the function are garbled by one encrypted data, when it gets decrypted, is equivalent to the result of performing the same operations party, and the private inputs are garbled and without any encryption. HE has the drawback of exchanged using an oblivious transfer protocol. being impractically slow. However, it has been A garbled circuit can provide a solution for getting more practical and standardized over the last privacy-preserving computations [171]. For ex- few years. HE can play a very useful role in healthcare ample, consider a patient who wants to use a applications where privacy is crucial, and using the diagnosis service without revealing his data and a data is subject to regulations. Many works in liter- service provider also wants to hide his algorithm ature have demonstrated the idea of using homo- parameters, which are considered trade secrets. morphic encryption for privacy-preserving machine In this case, a service provider can convert his learning in medical applications [155]. Research algorithm into a Boolean circuit, garble the circuit, and send it to the patient to be evaluated studies in [156–161] have presented diﬀerent tech- niques to train a logistic regression model over without loss of privacy. encrypted data using homomorphic encryption. In (ii) Secret Sharing: in this technique, an entity can [162–164], techniques of using the naive Bayes preserve the privacy of its sensitive data by classiﬁer model without leaking privacy information breaking it up into multiple shares and distrib- by applying homomorphic encryption have been uting the shares to a set of non-colluding parties presented. Cheon et al. [165] have presented a where each party computes a partial result Journal of Healthcare Engineering 17 depending on the shares it received [172]. Fi- artiﬁcial intelligence and machine learning still face some nally, one of the parties can receive these partial challenges in medical wearable devices as presented in this results and combine them to get the ﬁnal result. review. In this section, we will discuss brieﬂy a summary for the main perceived pitfalls or diﬃculties facing applying SMPC protocols are widely used to provide privacy- machine learning research for wearable devices and high- preserving in machine learning applications. However, light the related machine learning research directions that SMPC fails to protect against exploratory attacks. Explor- need further development. atory attacks act by performing several queries on a fully *e training data input to a machine learning model is trained model to leak some information about the model considered the most crucial element in the machine learning parameters and its training data, such as if a speciﬁc example process as garbage in mean garbage out. *e ﬁrst step is to was used in the training set or not. With this information, choose well-calibrated sensors that are better validated the attacker can gradually train a substitute model that against benchmarked devices used in hospitals that have reproduces the same prediction of the target model [147]. To undergone plenty of clinical experiments or other gold address these kinds of attacks, Kesarwani et al. [173] pro- standard devices [180]. Care should be taken as some of the posed a monitoring scheme called extraction monitor to research wearable devices provide raw data [102] that re- track the queries issued by the user, evaluate the information quire clean-up of the signals for removing noise and motion that a user might leak from these queries, and give a warning artifacts as the work done in [181]. Identifying the inac- when the user exceeds the average number of queries needed curacies in the data collected and considering that most of to reconstruct the model. the sensors are only accurate during rest [102] have to be taken into account as this has implications on the drawn conclusions and health-related decisions using wearable 3.8.3. Federated Learning Methods. Federated learning was devices in research. Most of the research works that have ﬁrst introduced in [174]. Federated learning is a machine been cited in Section 2 use research-grade wearable devices learning setting in which many devices collaborate in and do not mention the preprocessing and cleaning up steps training a model in a centralized manner while keeping the of data. Signal processing techniques are better to be training data private and decentralized [175]. In the cross- employed to cure these signals and remove motion artifacts device federated learning setting, the server sends out an [182]. Moreover, clinical experiments are to be done to help initial model to the devices, and the clients then train the in deﬁning the reference signal/ground truth and obtaining model on-device with their data locally and send the updated clinical evidence. device model to the server. Updated models are combined at For some applications, the ground truth signal is not the server using federated averaging to update the initial known due to the complexity of the human body’s response model. *is process goes on by sending the updated com- and the diﬀerent responses for each individual. For this bined model until the metrics are satisfactory. reason, the collection of data from as many subjects as possible *is approach was tested in [176] by applying federated is recommended to develop algorithms to clean the data and learning to heart activity data collected from multiple smart build more accurate models that generalize well. However, the bands in a stress-level monitoring scenario. *e authors process of data collection is an expensive and time-taking achieved comparable accuracy while preserving the privacy process that most academic research work cited in Section 4, of the data and reducing the communication burden by only which was not funded by companies, depending on data from communicating the models’ parameters. Additional privacy- a relatively small set of subjects. A transparent and repro- preserving protections such as secure multiparty compu- ducible process for collection of data from wearables, training, tation or diﬀerential privacy may further be included in the and evaluation of models is recommended for gaining trust in federated learning setting to keep data and model statistics the research results and eﬀectively building over accumu- private from malicious clients [177]. In the research done by lating research eﬀorts. *e authors in [183] pointed out [178], the authors utilized both SMC and diﬀerential privacy recommendations for reporting machine learning results in to balance the trade-oﬀ between vulnerability to inference clinical research and similar guidelines are to be followed for and low accuracy in a federated learning setting. machine learning research for wearable devices, especially For most of the healthcare applications, the machine those used in healthcare as they aﬀect human life. learning model is better to be personalized as per the bio- One cheap approach for big data collection is crowd- signals for each patient. Model aggregation with federated sourcing data collection such as the one initiated by a re- averaging as mentioned above does not provide this per- search group at Stanford University (https://innovations. sonalization. *e authors in [179] applied transfer learning stanford.edu/wearables), which collects data from wear- in the federated learning setting so that each device can train able devices remotely through a mobile application. How- a personalized model tailored to the user’s data by utilizing ever, this approach is susceptible to many privacy issues that the cloud model and data and the local data. existing commercial smart watch entities do not handle and the data collection process will be only protected by the 4. Discussion privacy policy of the application. Privacy-aware sharing of *e use of artiﬁcial intelligence research has clearly been data and learning from it without revealing the actual users’ data employing some privacy-preserving techniques men- rapidly growing in healthcare applications. However, for healthcare wearable devices, it can be seen that practical tioned in the last Section 3 is an active research area. An 18 Journal of Healthcare Engineering wearable healthcare domain to give insight into the conﬁ- example for that is the work by the authors in [184] for privacy-preserving data collection using a local diﬀerential dence in the ML model, which would be helpful, especially in tasks such as seizure detection and diseases’ diagnosis de- privacy technique over salient data to protect users’ data. Research work in diﬀerential privacy has also open issues to vices. Besides uncertainty modeling in wearable ML appli- investigate new learning solutions, which can learn from cations, joining upregulation among key stakeholders in the data distribution tails (data that represent minority class) ﬁeld of healthcare wearables is a key to make sure that ML is eﬀectively while maintaining an acceptable privacy loss as introduced in wearables with more transparency from tech suggested by [153]. Federated learning is also a relatively new companies and for gaining better users’ trust and accept- privacy-preserving method that needs further attention and ability. As it is pretty obvious, ML applications for healthcare wearables are a multidisciplinary ﬁeld that requires stan- exploration in the ﬁeld of machine learning for wearable devices, which can also promote the personalization of the dards about naming conventions, evaluation metrics, ethical reporting of research results, and clinical impact as sug- local model. Another approach for augmenting the data input to the gested in [101] as these devices may directly aﬀect human life. Mentioning evaluation of machine learning models, it ML model is generating training data using generative adversarial network (GAN) variants that may help train was found that the use of subject-based cross-validation is good quality models without exposing users’ wearable data recommended since subject data represent a clinically more and signals used in training or without even using any real relevant scenario for disease diagnosis application than data but only simulated data as in [185]. using records from a subject in training while using some For guaranteeing some level of privacy for users of other records for the same subject in testing the machine wearable devices, the machine learning application, which learning algorithm. *is can let the machine learning al- gorithm learn an association between unique features of the usually holds identiﬁable information at the edge device (e.g., smartphone running the application), should follow a same subject and accordingly may fall into overﬁtting. *is raises a question about whether building personalized set of regulations to gain users’ trust. Besides compliance with HIPAA (Health Insurance Portability and Account- models can be more eﬀective as it learns from signals of the same individual to avoid averaging out important individual ability Act: https://www.hhs.gov/hipaa/for-professionals/ index.html), GDPR (General Data Protection Regulation: characteristics such as age, sex, weight, height, eating style, https://gdpr.eu/), HITECH (Health Information Technology and way of doing an exercise [26]. Furthermore, person- for Economic and Clinical Health: https://www.hhs.gov/ alized models can learn from much less data and guarantee hipaa/for-professionals/special-topics/hitech-act- better privacy for data. enforcement-interim-ﬁnal-rule/index.html), and act regu- Deciding upon the time window of the signal to learn lations, machine learning applications for wearable devices from is yet another challenging decision as more data do not necessarily mean better results. It faces memory limitations should follow the OWASP security standards (Open Web Application Security Project (OWASP): https://owasp.org/ on the wearable device and power limitations for sending this amount of sampled data from the device to the edge or www-project-top-ten/). Among other data challenges in machine learning for to the cloud. Consequently, the sampling rate for diﬀerent wearables is identifying which data to be collected from these physiological signals needs to be optimized as per the ap- devices. Diﬀerent modalities can increase the accuracy of the plication to optimize the use of resources and decrease the model as usually many signals can be used for a single task. power consumption. Nevertheless, exploring tinyML em- Another big challenge is how to model the uncertainty bedded solutions and models’ optimization techniques in arising from the complex input received by the human body IoT is a recent research area that is open to some applications as shown in Figure 2, which may aﬀect the accuracy of any in healthcare wearables as well. However, for computa- model. For example, considering the stress detection task is a tionally intensive applications, full computation oﬄoading can be eﬀective while for data-intensive applications oﬀ- complex task that involves many inputs and can aﬀect many body systems. Using multimodal sources is believed to be a loading techniques that oﬄoad the processing of some of the data will be more suitable as suggested by [187] while very rich source of information that can help in health monitoring by identifying the emotional state, stress level, preserving the privacy of users’ data. and diagnosis of some diseases. As an example, monitoring Most of the research works for applying machine audio signals from the user to detect laughing, crying, learning for healthcare wearable devices tasks that we shouting, or coughing can help in these applications on reviewed in Section 3 are experiments for learning from data wearables, but it faces other challenges [186]. EEG signal, obtained from one or more sensors for detection or rec- despite the diﬃculty of capturing it, can also hold a lot of ognition of some pattern. Complete analysis of the proposed models in terms of memory requirements and amount of information, which can help in achieving higher accuracy in many applications such as dehydration detection, emotion communicated data in case of edge or cloud deployment, which greatly aﬀects power consumption, is better to be recognition, and mental disorders detection. Moreover, depending on learning from only few body provided. Overcoming these diﬃculties with AI solutions, together signals, which may also be noisy, would lead to uncertain decision from the ML model. As handling and estimating with the ongoing research and development in the ﬁeld of uncertainty in ML modeling remain an active area of re- medical sensors, storage, SoCs, and power-eﬃcient man- search in ML, we recommend applying its techniques in the agement and generation for wearable devices, would ensure Journal of Healthcare Engineering 19 digital stethoscope platform,” Journal of American Heart having AI-enabled healthcare wearable devices that can help Association, vol. 10, no. 9, Article ID e019905, 2021. reliably with remote patient monitoring, detect problems [5] S. Seneviratne, Y. Hu, T. Nguyen et al., “A survey of wearable with the human body earlier and assist in diagnosis, perform devices and challenges,” IEEE Communications Surveys & elderly people care and monitoring, and act as a lifestyle Tutorials, vol. 19, no. 4, pp. 2573–2620, 2017. guide and much more. A lot of many other AI applications [6] M. Chan, D. Esteve, ` J.-Y. Fourniols, C. Escriba, and can receive user acceptance by solving problems with the E. Campo, “Smart wearable systems: current status and reliability, comfort, and privacy of the wearable devices. future challenges,” Artiﬁcial Intelligence in Medicine, vol. 56, no. 3, pp. 137–156, 2012. [7] P. Siirtola, H. Koskimaki, ¨ H. Monttinen, ¨ and J. Roning, ¨ 5. Conclusion “Using sleep time data from wearable sensors for early de- *e growth of using wearable devices over the recent years is tection of migraine attacks,” Sensors, vol. 18, no. 5, 2018. clearly noticeable. With the huge amount of research eﬀorts [8] C. Meisel, R. El Atrache, M. Jackson, S. Schubach, C. Ufongene, and T. Loddenkemper, “Machine learning to employ artiﬁcial intelligence solutions in healthcare-re- from wristband sensor data for wearable, noninvasive seizure lated tasks for wearable devices, their use is expected to grow forecasting,” Epilepsia, vol. 61, no. 12, pp. 2653–2666, 2020. even more and change from a “nice to have” devices with fun [9] A. Y. Hannun, P. Rajpurkar, and M. Haghpanahi, “-level applications to necessary devices for remote patient moni- arrhythmia detection and classiﬁcation in ambulatory toring and detection of any irregularities with the human electrocardiograms using a deep neural network,” Nature body. In this review, we presented ML tasks that have been Medicine, vol. 25, pp. 65–69, 2019. researched in the healthcare wearable devices ﬁeld, the [10] S. Kwon, J. Hong, E. Choi et al., “Detection of atrial ﬁ- machine learning techniques used, the diﬀerent modalities brillation using a ring-type wearable device (CardioTracker) used, and the available datasets in the ﬁeld. *e diﬀerent and deep learning analysis of photoplethysmography signals: challenges facing machine learning applications on wearable prospective observational proof-of-concept,” Study J Med devices (deployment alternatives, power consumption, Internet Res, vol. 22, no. 5, p. 2020. [11] Z. Mei, X. Gu, H. Chen, and W. Chen, “Automatic atrial storage and memory, utility and user acceptance, data ﬁbrillation detection based on heart rate variability and availability and reliability, communication, security and spectral features,” IEEE Access, vol. 6, Article ID 53566, 2018. privacy) were discussed while identifying possible solutions [12] N. Rashid and M. A. Al Faruque, “Energy-eﬃcient real-time found in the literature. Finally, the study highlights issues myocardial infarction detection on wearable devices,” Annu that require further research concerning data availability, Int Conf IEEE Eng Med Biol Soc, vol. 2020, pp. 4648–4651, reliability, and privacy to enable eﬀective and eﬃcient learning from data generated by wearable devices. [13] R. Buettner, J. Frick, and T. Rieg, “High-performance De- Contents of the research are solely the responsibility of tection of Epilepsy in Seizure-free EEG Recordings,” in the authors and do not necessarily represent the oﬃcial Proceedings of the International Conference on Information views of the Qatar National Research Fund (QNRF). systems, p. 12, Munich, Germany, December, 2019. [14] C. Ieracitano, N. Mammone, A. Hussain, C. Francesco, and Morabito, “A novel multi-modal machine learning based Conflicts of Interest approach for automatic classiﬁcation of EEG recordings in dementia,” Neural Networks, vol. 123, pp. 176–190, 2020. *e authors declare that there are no conﬂicts of interest in [15] S. Hwang, H. Jebelli, B. Choi, M. Choi, and S.H. Lee, this study. “Measuring workers’ emotional state during construction tasks using wearable EEG,” Journal of Construction Engi- Acknowledgments neering and Management, vol. 144, p. 01, 2018. [16] O. Al Zoubi, M. Awad, K. Nikola, and Kasabov, “Anytime *is publication was made possible by a grant from the multipurpose emotion recognition from EEG data using a Qatar National Research Fund (QNRF), project number Liquid State Machine based framework,” Artiﬁcial Intelli- ECRA 01-006-1-001. gence in Medicine, vol. 86, pp. 1–8, 2018. [17] A. Jaramillo-Yanez, ´ M. E. Benalcazar, ´ and E. Mena-Mal- donado, “Real-time hand gesture recognition using surface References electromyography and machine learning: a systematic lit- erature review,” Sensors, vol. 20, no. 9, 2020. [1] J. Hayano, H. Yamamoto, I. Nonaka et al., “Quantitative [18] C. Fang, Y. Wang, and S. Gao, “Utilizing wearable GRF and detection of sleep apnea with wearable watch device,” PLoS EMG sensing system and machine learning algorithms to ONE, vol. 15, Article ID e0237279, 2020. enable locomotion mode recognition for in-home rehabili- [2] F. Delmastro, F. D. Martino, and C. Dolciotti, “Cognitive training and stress detection in MCI frail older people tation,” in Proceedings of the 2020 IEEE International Conference on Flexible and Printable Sensors and Systems, through wearable sensors and machine learning,” IEEE Access, vol. 8, Article ID 65573, 2020. vol. 1–4, Manchester, UK, August, 2020. [19] B. Milosevic, S. Benatti, and E. Farella, “Design challenges for [3] M. V. Perez, K. W. Mahaﬀey, H. Hedlin et al., “Large-Scale Assessment of a smartwatch to identify atrial ﬁbrillation,” wearable EMG applications,” in Proceedings of the Design, Automation Test in Europe Conference Exhibition, New England Journal of Medicine, vol. 381, no. 20, pp. 1909–1917, 2019. pp. 1432–1437, Lausanne, Switzerland, March 2017. [20] Y. S. Can, B. Arnrich, and C. Ersoy, “Stress detection in daily [4] J. S. Chorba, A. M. Shapiro, L Le et al., “Deep learning al- gorithm for automated cardiac murmur detection via a life scenarios using smart phones and wearable sensors: a 20 Journal of Healthcare Engineering survey,” Journal of Biomedical Informatics, vol. 92, Article ID International Symposium on Wearable Computers, ISWC ’15, 103139, 2019. pp. 91-92, New York, NY, USA, September, 2015. [21] H. F. Posada-Quintero, N. Reljin, A. Moutran et al., “Mild [36] S. Bi, T. Wang, E. Davenport et al., “Toward a wearable dehydration identiﬁcation using machine learning to assess sensor for eating detection,” in Proceedings of the 2017 autonomic responses to cognitive stress,” Nutrients, vol. 23, Workshop on Wearable Systems and Applications, WearSys no. 12, p. 42, 2019. ’17, pp. 17–22, Association for Computing Machinery, New [22] K. C. Un, C. K. Wong, and Y. M. Lau, “Observational study York, NY, USA, June, 2017. on wearable biosensors and machine learning-based remote [37] M. Nourollahi, S. A. Rokni, P. Alinia, and G. Hassan, monitoring of COVID-19 patients,” Scientiﬁc Reports, “Proximity-based active learning for eating moment rec- vol. 11, p. 4388, 2021. ognition in wearable systems,” in Proceedings of the 6th ACM [23] D. Castaneda, A. Esparza, M. Ghamari, C. Soltanpur, and Workshop on Wearable Systems and Applications, WearSys H. Nazeran, “A review on wearable photoplethysmography ’20, pp. 7–12, Association for Computing Machinery, New sensors and their potential future applications in healthcare,” York, NY, USA, June, 2020. International journal of biosensors & bioelectronics, vol. 4, [38] E. Zdravevski, B. R. Stojkoska, M. Standl, and H. Schulz, pp. 195–202, 2018. “Automatic machine-learning based identiﬁcation of jogging [24] F. P. Akbulut, B. Ikitimur, and A. Akan, “Wearable sensor- periods from accelerometer measurements of adolescents based evaluation of psychosocial stress in patients with under ﬁeld conditions,” PLOS ONE, vol. 12, no. 9, pp. 1–28, metabolic syndrome,” Artiﬁcial Intelligence in Medicine, vol. 104, Article ID 101824, 2020. [39] D. Zhang, H. Zhang, and C. Zhao, “Prediction of behavior [25] C. T. Yen, J. X. Liao, and Y. K. Huang, “Human daily activity and activity of patients through wearable devices based on recognition performed using wearable inertial sensors deep learning,” Journal of Sensors, vol. 2022, Article ID combined with deep learning algorithms,” IEEE Access, 3067840, 12 pages, 2022. vol. 8, Article ID 174105, 2020. [40] A. A. Gharibans, B. L. Smarr, and D. C. Kunkel, “Artifact [26] C. Paul, G. Lefebvre, S. Duﬀner, and C. Garcia, “Learning rejection methodology enables continuous, noninvasive personalized ADL recognition models from few raw data,” measurement of gastric myoelectric activity in ambulatory Artiﬁcial Intelligence in Medicine, vol. 107, Article ID 101916, subjects,” Scientiﬁc Reports, vol. 8, p. 5019, 2018. [41] J. Heo, H. Yoon, and K. S. Park, “A novel wearable forehead [27] E. Casilari, J. A. S. Ramon, ´ and J. M. C. Garc´ıa, “Analysis of EOG measurement system for human computer interfaces,” public datasets for wearable fall detection systems,” Sensors, Sensors, vol. 17, no. 7, 2017. vol. 17, no. 7, 2017. [42] A. Tonacci, L. Billeci, I. Mambro, R. Marangoni, [28] A. Harris, H. True, Z. Hu, J. Cho, N. Fell, and M. Sartipi, “Fall C. Sanmartin, and F. Venturi, “Wearable sensors for recognition using wearable technologies and machine learning assessing the role of olfactory training on the autonomic algorithms,” in Proceedings of the 2016 IEEE International response to olfactory stimulation,” Sensors, vol. 21, no. 3, Conference on Big Data, pp. 3974–3976, (Big Data), Washington, p. 770, 2021. DC, USA, December, 2016. [43] F. Sabry, T. Eltaras, W. Labda, F. Hamza, K. Alzoubi, and [29] K. Liu, C. Hsieh, S. J. Hsu, and C. Chan, “Impact of sampling Q. Malluhi, “Towards on-device dehydration monitoring rate on wearable-based fall detection systems based on using machine learning from wearable device’s data,” Sen- machine learning models,” IEEE Sensors Journal, vol. 18, sors, vol. 22, no. 5, 2022. no. 23, pp. 9882–9890, 2018. [44] S. Park, S. W. Lee, S. Han, and M. Cha, “Clustering insomnia [30] S. B. Kwon, J. Park, C. Kwon, H. J. Kong, J. Y. Hwang, and patterns by data from wearable devices: algorithm devel- H. C. Kim, “An energy-eﬃcient algorithm for classiﬁcation opment and validation study,” JMIR mHealth and uHealth, of fall types using a wearable sensor,” IEEE Access, vol. 7, vol. 7, no. 12, 2019. Article ID 31321, 2019. [45] A. Abedin, F. Motlagh, Q. Shi, R. Hamid, and D. Ranasinghe, [31] A. Collado-Villaverde, M. Cobos, P. Muñoz, and “Towards deep clustering of human activities from wear- D. F. Barrero, “A simulator to support machine learning- ables,” in Proceedings of the 2020 International Symposium based wearable fall detection systems,” Electronics, vol. 9, on Wearable Computers, ISWC ’20, pp. 1–6, Association for no. 11, 2020. Computing Machinery, New York, NY, USA, September, [32] K. Sundararajan, S. Georgievska, B. Te Lindert et al., “Sleep classiﬁcation from wrist-worn accelerometer data using [46] Po-T. Lee, W.-C. Chiu, Y.-H. Ho, Y.-C. Tai, C.-C. K. Lin, and random forests,” Scientiﬁc Reports, vol. 11, no. 24, 2021. C.-L. Lin, “Development of wearable device and clustering [33] A. Cheon, S. Y. Jung, C. Prather, M. Sarmiento, K. Wong, based method for detecting falls in the elderly,” in Pro- and D. M. k. Woodbridge, “A machine learning approach to ceedings of the 2021 IEEE 10th Global Conference on Con- detecting low medication state with wearable technologies,” sumer Electronics (GCCE), pp. 231-232, Kyoto, Japan, in Proceedings of the 42nd Annual International Conference Octobar, 2021. of the IEEE Engineering in Medicine Biology Society, [47] J. Dunn, L. Kidzinski, and R. Runge, “Wearable sensors pp. 4252–4255, 2020. enable personalized predictions of clinical laboratory mea- [34] E. *omaz, I. Essa, and D. Gregory, “Abowd. A practical surements,” Nature Medicine, vol. 27, 2021. approach for recognizing eating moments with wrist- [48] C. Chatzaki, M. Pediaditis, V. George, and M. Tsiknakis, mounted inertial sensing,” in Proceedings of the 2015 ACM “Human daily activity and fall recognition using a smart- International Joint Conference on Pervasive and Ubiquitous Computing, pp. 1029–1040, Osaka, Japan, 2015. phone’s acceleration sensor,” in Proceedings of the Infor- mation and Communication Technologies for Ageing Well [35] A. Bedri, A. Verlekar, E. *omaz, V. Avva, and T. Starner, “A wearable system for detecting eating activities with proximity and e-Health, pp. 100–118, Springer International Publish- sensors in the outer ear,” in Proceedings of the 2015 ACM ing, Cham, Switcher land, June, 2017. Journal of Healthcare Engineering 21 [49] A. Ramachandran, R. Adarsh, P. Pahwa, and K. R. Anupama, [62] P. Resque, A. Barros, D. Rosario, ´ and E. Cerqueira, “An “Machine learning-based techniques for fall detection in investigation of diﬀerent machine learning approaches for epileptic seizure detection,” in Proceedings of the 15th In- geriatric healthcare systems,” in Proceedings of the 2018 9th ternational Wireless Communications Mobile Computing International Conference on Information Technology in Conference (IWCMC), pp. 301–306, Tangier, Morocco, June, Medicine and Education (ITME), pp. 232–237, Hangzhou, China, Octobar, 2018. [63] G. Regalia, F. Onorati, M. Lai, C. Caborni, W. Rosalind, and [50] F. Hussain, F. Hussain, M. Ehatisham-ul-Haq, and Picard, “Multimodal wrist-worn devices for seizure detection M. A. Azam, “Activity-aware fall detection and recognition and advancing research: focus on the Empatica wristbands,” based on wearable sensors,” IEEE Sensors Journal, vol. 19, Epilepsy Research, vol. 153, pp. 79–82, 2019. no. 12, pp. 4528–4536, 2019. [64] C. Caborni, F. Onorati, G. Regalia et al., “Prospective [51] J. D. Lopez, ´ A. Sucerquia, and J. F. V.-B. SisFall, “A fall and evaluation of seizure detection performance of the Embrace movement dataset,” Sensors, vol. 17, no. 1, p. 198, 2017. wristband on pediatric patients in the epilepsy monitoring [52] D. Giuﬀrida, B. Guido, D. De Martini, and T. Facchinetti, unit,” in Proceedings of the 2nd International Congress on “Fall detection with supervised machine learning using Mobile Devices & Seizure Detection in Epilepsy, Lausanne, wearable sensors,” in Proceedings of the 2019 IEEE 17th Switzer land, September, 2019. International Conference on Industrial Informatics (INDIN), [65] F. Onorati, G. Regalia, C. Caborni et al., “Multicenter Clinical vol. 1, pp. 253–259, Helsinki, Finland, July, 2019. Assessment of Improved Wearable Multimodal Convulsive [53] R. Saeedi, K. Sasani, S. Norgaard, and A. H. Gebremedhin, Seizure Detectors,” Epilepsia, vol. 58, 2017. “Personalized human activity recognition using wearables: a [66] G. Yang, J. Deng, G. Pang et al., “An IoT-enabled stroke manifold learning-based knowledge transfer,” in Proceedings rehabilitation system based on smart wearable armband and of the 2018 40th Annual International Conference of the IEEE machine learning,” IEEE Journal of Translational Engineering Engineering in Medicine and Biology Society (EMBC), in Health and Medicine, vol. 6, no. 1–10, 2018. pp. 1193–1196, Honolulu, HI, USA, July, 2018. [67] S. Potluri, A. B. Chandran, C. Diedrich, and L. Schega, [54] X. Li, Z. Zhou, J. Wu, and Y. Xiong, “Human posture de- “Machine learning based human gait segmentation with tection method based on wearable devices,” Journal of wearable sensor platform,” in Proceedings of the 2019 41st Healthcare Engineering, vol. 2021, Article ID 8879061, Annual International Conference of the IEEE Engineering in 8 pages, 2021. Medicine and Biology Society (EMBC), pp. 588–594, Berlin, [55] A. Bedri, R. Li, M. Haynes et al., “EarBit: using wearable Germany, July, 2019. sensors to detect eating episodes in unconstrained envi- [68] H. Zhang, Y. Guo, and D. Zanotto, “Accurate ambulatory ronments,” Proceedings of the ACM on interactive, mobile, gait analysis in walking and running using machine learning wearable and ubiquitous technologies, vol. 1, no. 3, p. 37, models,” IEEE Transactions on Neural Systems and Reha- bilitation Engineering, vol. 28, no. 1, pp. 191–202, 2020. [56] K. San Chun, H. Jeong, R. Adaimi, and E. *omaz, “Eating [69] S. R. Moore, C. Kranzinger, J. Fritz, T. Stggl, J. Krll, and episode detection with jawbone-mounted inertial sensing,” H. Schwameder, “Foot strike angle prediction and pattern in Proceedings of the 2020 42nd Annual International Con- classiﬁcation using LoadsolTM wearable sensors: a com- ference of the IEEE Engineering in Medicine Biology Society parison of machine learning techniques,” Sensors, vol. 20, (EMBC), pp. 4361–4364, Montreal, QC, Canada, July, 2020. no. 23, 2020. [57] D. Morris, S. Scott, A. Guillory, and I. Kelner, “RecoFit: using [70] Y. Mengistu, M. Pham, M. Do, and W. Sheng, “Autohydrate: a wearable sensor to ﬁnd, recognize, and count repetitive A Wearable Hydration Monitoring System,” in Proceedings exercises,” in Proceedings of the SIGCHI Conference on of the 2016 IEEE/RSJ International Conference on Intelligent Human Factors in Computing Systems, pp. 3225–3234, ACM, Robots and Systems (IROS), pp. 1857–1862, Daejeon, Korea Toranto, Canada, April, 2014. (South), Octobar, 2016. [58] A. J. Masino, D. Forsyth, H. Nuske et al., “m-Health and [71] A. Alvarez, E. Severeyn, J. Velasquez, ´ S. Wong, G. Perpiñan, Autism: recognizing stress and anxiety with machine and M. . Huerta, “Machine learning methods in the classi- learning and wearables data,” in Proceedings of the 2019 IEEE ﬁcation of the athletes dehydration,” in Proceedings of the 32nd International Symposium on Computer-Based Medical 2019 IEEE Fourth Ecuador Technical Chapters Meeting Systems (CBMS), pp. 714–719, Cordoba, Spain, June, 2019. (ETCM), pp. 1–5, Guayaquil, Ecuador, November, 2019. [59] Q. Zhang, X. Zeng, W. Hu, and D. Zhou, “A machine [72] S. Koelstra, C. Muehl, M. Soleymani et al., Deap, A database learning-empowered system for long-term motion-tolerant for emotion analysis using physiological signals,” IEEE wearable monitoring of blood pressure and heart rate with Transactions on Aﬀective Computing, vol. 3, no. 1, pp. 18–31, ear-ECG/PPG,” IEEE Access, vol. 5, Article ID 10547, 2017. [60] S. Tuli, N. Basumatary, S. S. Gill et al., “HealthFog: an en- [73] R. Laureanti, M. Bilucaglia, M. Zito et al., “Emotion as- semble deep learning based smart healthcare system for sessment using Machine Learning and low-cost wearable automatic diagnosis of heart diseases in integrated IoT and devices,” in Proceedings of the 42nd Annual International fog computing environments,” Future Generation Computer Conference of the IEEE Engineering in Medicine Biology Systems, vol. 104, pp. 187–200, 2020. Society (EMBC), pp. 576–579, Montreal, QC, Canada, July, [61] Y. Shen, M. Voisin, A. Aliamiri, A. Anand, A. Hannun, and A. Ng, “Ambulatory atrial ﬁbrillation monitoring using [74] D. Ayata, Y. Yaslan, and M. E. Kamasak, “Emotion recog- wearable photoplethysmography with deep learning,” in nition from multimodal physiological signals for emotion Proceedings of the 25th ACM SIGKDD International Con- aware healthcare systems,” Journal of Medical and Biological ference on Knowledge Discovery Data Mining, KDD ’19, Engineering, vol. 40, pp. 149–157, 2020. pp. 1909–1916, Association for Computing Machinery, New [75] F. Miao, Bo Wen, Z. Hu et al., “Continuous blood pressure York, NY, USA, July, 2019. measurement from one-channel electrocardiogram signal 22 Journal of Healthcare Engineering using deep-learning techniques,” Artiﬁcial Intelligence in [90] O. Walch, “Motion and Heart Rate from a Wrist-Worn Medicine, vol. 108, Article ID 101919, 2020. Wearable and Labeled Sleep from Polysomnography,” [76] C. K. Wong, D. T. Y. Ho, A. R. Tam, M. Zhou, Y. M. Lau, and PhysioNet, vol. 101, 2019. [91] O. Walch, Y. Huang, D. Forger, and C. Goldstein, “Sleep M. O. Y. Tang, “Artiﬁcial intelligence mobile health platform stage prediction with raw acceleration and photo- for early detection of COVID-19 in quarantine subjects using plethysmography heart rate data derived from a consumer a wearable biosensor: protocol for a randomised controlled wearable device,” Sleep, vol. 42, no. 12, p. 08, 2019. trial,” BMJ Open, vol. 10, 2020. [92] D. Houfani, S. Slatnia, O. Kazar, H. Saouli, and A. Merizig, [77] S. Rastogi and J. Singh, “A Systematic Review on Machine “Artiﬁcial intelligence in healthcare: a review on predicting Learning for Fall Detection System,” Computational Intel- clinical needs,” International Journal of Healthcare Man- ligence, vol. 37, 2021. agement, vol. 0, no. 0, pp. 1–9, 2021. [78] M. Kilany, A. E. Hassanien, and A. Badr, “Accelerometer- [93] A. Jain, D. Way, V. Gupta, Yi Gao, G. O. Marinho, and based human activity classiﬁcation using Water Wave Op- J. Hartford, “Development and assessment of an artiﬁcial timization approach,” in Proceedings of the 2015 11th In- intelligence–based tool for skin condition diagnosis by ternational Computer Engineering Conference (ICENCO), primary care physicians and nurse practitioners in tele- pp. 175–180, Cario, Egypt, December, 2015. dermatology practices,” JAMA Network Open, vol. 4, no. 4, [79] C. Min, A. Mathur, and F. Kawsar, “Exploring audio and Article ID e217249, 2021. kinetic sensing on earable devices,” in Proceedings of the 4th [94] J. Greiwe and S. M. Nyenhuis, “Wearable technology and ACM Workshop on Wearable Systems and Applications, how this can Be implemented into clinical practice,” Current WearSys ’18, pp. 5–10, Association for Computing Ma- Allergy and Asthma Reports, vol. 20, no. 8, p. 36, 2020. chinery, New York, NY, USA, June, 2018. [95] T. Chitnis, B. I. Glanz, and C. Gonzalez, “Quantifying [80] B. M. Bell, R. Alam, and N. Alshurafa, “Automatic, wearable- neurologic disease using biosensor measurements in-clinic based, in-ﬁeld eating detection approaches for public health and in free-living settings in multiple sclerosis,” Npj Digital research: a scoping review,” Npj Digital Medicine, vol. 3, Medicine, vol. 2, no. 123, 2019. no. 38, 2020. [96] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes- [81] S. Bae, D. Ferreira, B. Suﬀoletto et al., “Detecting drinking Ortiz, “A Public Domain Dataset for Human Activity episodes in young adults using smartphone-based sensors,” Recognition Using Smartphones,” in Proceedings of the 21st Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.vol.1, European Symposium on Artiﬁcial Neural Networks, Com- no. 2, 2017. putational Intelligence And Machine Learning, Bruges, [82] M. H. Imtiaz, R. I. Ramos-Garcia, S. Wattal, T. Stephen, and Belgium, April 2013. E. Sazonov, “Wearable sensors for monitoring of cigarette [97] D. Garcia-Gonzalez, D. Rivero, E. Fernandez-Blanco, and smoking in free-living: a systematic review,” Sensors, vol. 19, M. R. Luaces, “A public domain dataset for real-life human no. 21, 2019. activity recognition using smartphone sensors,” Sensors, [83] Y. Choi, Yu-Mi Jeon, L. Wang, and K. Kim, “A biological vol. 20, no. 8, 2020. signal-based stress monitoring framework for children using [98] R. Meziatisabour, Y. Benezeth, P. De Oliveira, Julien Chappe, wearable devices,” Sensors, vol. 17, no. 9, 2017. and Y. Fan, “UBFC-phys: a multimodal database for psy- [84] A. Greco, G. Valenza, A. Lanata, G. Rota, and E. P. Scilingo, chophysiological studies of social stress,” IEEE Transactions “Electrodermal activity in bipolar patients during aﬀective on Aﬀective Computing, p. 1, 2021. elicitation,” IEEE Journal of Biomedical and Health Infor- [99] V. Markova, T. Ganchev, and K. Kalinkov, “A database for matics, vol. 18, no. 6, pp. 1865–1873, 2014. cognitive load, aﬀect and stress recognition,” in Pro- [85] E. Garcia-Ceja, M. Riegler, T. Nordgreen, P. Jakobsen, ceedings of the First International Conference on Bio- K. J. Oedegaard, and Jim Tørresen, “Mental health moni- medical Innovations and Applications, Article ID 8967457, toring with multimodal sensing and machine learning: a Varna, Bulgaria, 2019. survey,” Pervasive and Mobile Computing, vol. 51, pp. 1–26, [100] G. Nguyen, S. Dlugolinsky, and M. Bobak, “Machine learning and deep learning frameworks and libraries for [86] T. Rukasha, S. I. Woolley, T. Kyriacou, and T. Collins, large-scale data mining: a survey,” Artiﬁcial Intelligence “Evaluation of wearable electronics for epilepsy: a systematic Review, vol. 52, 2019. review,” Electronics, vol. 9, no. 6, 2020. [101] P. C. Chen, Y. Liu, and L. Peng, “How to develop machine [87] J. C. Alcaraz, S. Moghaddamnia, N. Poschadel, and J. Peissig, learning models for healthcare,” Nature Materials, vol. 18, “Machine learning as digital therapy assessment for mobile no. 5, pp. 410–414, 2019. gait rehabilitation,” in Proceedings of the 2018 IEEE 28th [102] B. Bent, B. Goldstein, K. Warren, and J. Dunn, “Investigating International Workshop on Machine Learning for Signal sources of inaccuracy in wearable optical heart rate sensors,” Processing (MLSP), pp. 1–6, Aalborg, Denmark, September, Npj Digital Medicine, vol. 3, no. 18, 2020. [103] B. A. Mateen, J. Liley, and A. K. Denniston, “Improving the [88] R. Ghaﬀari, J. A. Rogers, and T. R. Ray, “Recent progress, quality of machine learning in health applications and challenges, and opportunities for wearable biochemical clinical research,” Nature Machine Intelligence, vol. 2, sensors for sweat analysis,” Sensors and Actuators B: pp. 554–556, 2020. Chemical, vol. 332, Article ID 129447, 2021. [104] A. S. Keerthy and S. Manju Priya, Artiﬁcial Intelligence in [89] O. Blan, G. Moise, A. Moldoveanu, M. Leordeanu, and Healthcare Databases, Springer International Publishing, F. Moldoveanu, “An investigation of various machine and Cham, Switcher land, pp. 19–34, 2021. deep learning techniques applied in automatic fear level [105] B. Nelson, C. Low, N. Jacobson, P. Arean, J. Torous, and detection and acrophobia virtual therapy,” Sensors, vol. 20, N. Allen, “Guidelines for wrist-worn wearable assessment of no. 2, 2020. heart rate in biobehavioral research,” Nature, vol. 3, 2020. Journal of Healthcare Engineering 23 [106] S. Saeb, A. Jayaraman, D. C. Mohr, K. P. Kording, and [122] R. Simon Sherratt, B. Janko, T. Hui et al., “Task scheduling to L. Lonini, “*e need to approximatee the use-case in clinical constrain peak current consumption in wearable healthcare machine learning,” GigaScience, vol. 6, no. 5, pp. 1–9, 2017. sensors,” Electronics, vol. 8, no. 7, 2019. [107] M. A. Little, G. Varoquaux, S. Saeb et al., “Using and un- [123] Alma Secerbegovic, A. Gogic, N. Suljanovic, M. Zajc, and derstanding cross-validation strategies. Perspectives on Saeb A. Mujcic, “Computational balancing between wearable etal,” GigaScience, vol. 6, no. 5, p. 03, 2017. sensor and smartphone towards energy-eﬃcient remote [108] A. B. Arrieta, N. D´ıaz-Rodr´ıguez, A. Bennetot et al., “Ex- healthcare monitoring,” Advances in Electrical and Computer plainable Artiﬁcial Intelligence (XAI): concepts, taxonomies, Engineering, vol. 18, no. 3–10, p. 11, 2018. opportunities and challenges toward responsible AI,” In- [124] M. Antonio Scrugli, D. Loi, L. Raﬀo, and P. Meloni, “A formation Fusion, vol. 58, pp. 82–115, 2020. runtime-adaptive cognitive IoT node for healthcare moni- [109] N. Mairittha, T. Mairittha, and S. Inoue, “On-device deep toring,” in Proceedings of the 16th ACM International personalization for robust activity data collection,” Sensors, Conference on Computing Frontiers, CF ’19, pp. 350–357, vol. 21, 2020. Association for Computing Machinery, New York, NY, USA, [110] Z. Fu, X. He, E. Wang, J. Huo, J. Huang, and D. Wu, May, 2019. “Personalized human activity recognition based on inte- [125] T. Takao, “Design consideration of sustainable self-suﬃcient grated wearable sensor and transfer learning,” Sensors, energy system using ferroelectric material for wearable vol. 21, no. 3, 2021. healthcare,” International Journal of Energy Research, vol. 40, [111] A. Ferrari, D. Micucci, M. Mobilio, and P. Napoletano, no. 15, pp. 2176–2186, 2016. “Deep learning and model personalization in sensor-based [126] J. Yuan and R. Zhu, “A fully self-powered wearable moni- human activity recognition,” Journal of Reliable Intelligent toring system with systematically optimized ﬂexible ther- Environments, vol. 54, 2022. moelectric generator,” Applied Energy, vol. 271, Article ID [112] L. Sun, X. Jiang, H. Ren, and Y. Guo, “Edge-cloud computing 115250, 2020. and artiﬁcial intelligence in internet of medical things: ar- [127] X. Fafoutis, L. Marchegiani, A. Elsts, J. Pope, R. Piechocki, chitecture, technology and application,” IEEE Access, vol. 8, and I. Craddock, “Extending the Battery Lifetime of Article ID 101079, 2020. Wearable Sensors with Embedded Machine Learning,” in [113] F. A. Kraemer, A. E. Braten, N. Tamkittikhun, and D. Palma, Proceedings of the IEEE 4th World Forum on Internet of “Fog computing in healthcare – a review and discussion,” Gings (WF-IoT), pp. 269–274, Singapore, Febrauary, 2018. IEEE Access, vol. 5, pp. 9206–9222, 2017. [128] D. Jaiswal, A. Gigie, T. Chakravarty, A. Ghose, and A. Misra, [114] D. A. Tamburri, “Sustainable MLOps: trends and chal- “Table of interest: activity recognition and behaviour analysis lenges,” in Proceedings of the 2020 22nd International using a battery less wearable sensor,” in Proceedings of the 4th Symposium on Symbolic and Numeric Algorithms for Sci- ACM Workshop on Wearable Systems and Applications, entiﬁc Computing (SYNASC), pp. 17–23, Timisoara, Roma- WearSys ’18, pp. 39–44, Association for Computing Ma- nia, September, 2020. chinery, New York, NY, USA, June, 2018. [115] J. Waring, C. Lindvall, and R. Umeton, “Automated machine [129] L. Klus, E. S. Lohan, C. Granell, and J. Nurmi, “Lossy Com- learning: review of the state-of-the-art and opportunities for pression Methods for Performance-Restricted Wearable healthcare,” Artiﬁcial Intelligence in Medicine, vol. 104, Devices,” in Proceedings of the 2020 International Conference Article ID 101822, 2020. on Localization and GNSS (ICL-GNSS), Tampere, Finland, [116] J. K. Kim, H. Jung, G. B. Hwang, O. S. Gwon, and S. E. Lee, June, 2020, https://doi.org/10.5281/zenodo.3908174. “Design of low-power SoC for wearable healthcare device,” [130] S. H. Baek and K. W. Park, “A Durable Hybrid RAM Disk Journal of Circuits, Systems, and Computers, vol. 29, no. 06, with a Rapid Resilience for Sustainable IoT Devices,” Sensors, Article ID 2050085, 2020. vol. 20, 2020. [117] E. Torti, C. D’Amato, Giovanni Danese, and F. Leporati, “An [131] J. Huang, A. Badam, and R. Chandra, Nightingale, Wear- hardware recurrent neural network for wearable devices,” in Drive: fast and energy-eﬃcient storage for wearables,” in Proceedings of the 2020 23rd Euromicro Conference on Digital Proceedings of the Annual Technical Conference (ATC’15). System Design (DSD), pp. 293–300, Kranj, Slovenia, August, USENIX - Advanced Computing Systems Association, Best 2020. Paper Award, Berkeley, CA, USA, July, 2015. [118] M. Faustino, J. Calado, J. Sarraipa, and R. Jardim-Gonçalves, [132] A. Kumar, S. Goyal, and M. Varma, “Resource-eﬃcient “Adaptable power consumption proﬁles for wearable lo- machine learning in 2 KB RAM for the internet of things,” in calization devices,” in Proceedings of the SEIA’2019 Con- Proceedings of the 34th International Conference on Machine ference Proceedings, p. 200, Tenerife, Canary, September, Learning - Volumevol. 70, pp. 1935–1944, JMLR.org, 2017. 2022. [133] M. Sek, “Chai. Quantization-guided training for compact [119] L. Zhang, J. Hu, C. Guo, and H. Xu, “Dynamic power op- TinyML models,” in Proceedings of the Research Symposium timization for secondary wearable biosensors in e-healthcare on Tiny Machine Learning, San Francisco, USA, March, 2021. leveraging cognitive WBSNs with imperfect spectrum [134] M. Merenda, C. Porcaro, and D. Iero, “Edge machine sensing,” Future Generation Computer Systems, vol. 112, learning for AI-enabled IoT devices: a review,” Sensors, pp. 67–92, 2020. vol. 20, no. 9, 2020. [120] M. Lewandowski, B. O. Płaczek, and M. Bernas, “Classiﬁer- [135] V. T. Tran, C. Riveros, and P. Ravaud, “Patients’ views of Based Data Transmission Reduction in Wearable Sensor wearable devices and AI in healthcare: ﬁndings from the Network for Human Activity Monitoring,” Sensors, vol. 21, ComPaRe e-cohort,” Npj Digital Medicine, vol. 2, 2019. no. 1, 2021. [136] P. Esmaeilzadeh, “Use of AI-based tools for healthcare [121] J. Zhou and C. Wang, “An ultra-low power turning angle purposes: a survey study from consumers’ perspectives,” based biomedical signal compression engine with adaptive BMC Medical Informatics and Decision Making, vol. 20, threshold tuning,” Sensors, vol. 17, p. 1809, 2017. p. 170, 2020. 24 Journal of Healthcare Engineering [137] Q. Yue, Z. Ling, X. Fu, B. Liu, W. Yu, and W. Zhao, “My [153] V. M. Suriyakumar, N. Papernot, A. Goldenberg, and Google Glass Sees Your Passwords! Black Hat USA,” Sensors, M. Ghassemi, “Challenges of diﬀerentially private prediction in healthcare settings,” in Proceedings of the IJCAI 2021 [138] D. Weiz, G. Anand, and C. P. H. Ernst, Ge Inﬂuence of Workshop on AI for Social Good, Montreal, Canada, 2021. [154] R. L. Rivest, L. Adleman, and M. L. Dertouzos, “On data Subjective Norm on the Usage of Smartglasses, Springer, banks and privacy homomorphisms,” Foundations of secure Cham, Switzerland. [139] N. T. Sahin, N. U. Keshav, J. P. Salisbury, and computation, vol. 4, no. 11, pp. 169–180, 1978. [155] A. Wood, K. Najarian, and D. Kahrobaei, “Homomorphic A. Vahabzadeh, “Second version of google glass as a wearable encryption for machine learning in medicine and bio- socio-aﬀective aid: positive school desirability, high usability, informatics,” ACM Computing Surveys, vol. 53, no. 4, 2020. and theoretical framework in a sample of children with [156] M. Kim, Y. Song, S. Wang, Y. Xia, and X. Jiang, “Secure autism,” JMIR human factors, vol. 5, no. 1, p. 1, 2018. logistic regression based on homomorphic encryption: de- [140] K. L. LaBat and K. S. Ryan, Human Body A Wearable Product sign and evaluation,” JMIR medical informatics, vol. 6, no. 2, Designer’s Guide, CRC Press, London, UK, 1 edition, 2019. p. e19, 2018. [141] T. Zhang, J. Lu, F. Hu, and H. Qi, “Bluetooth low energy for [157] Y. Aono, T. Hayashi, L. T. Phong, and L. Wang, “Privacy- wearable sensor-based healthcare systems,” in Proceedings of preserving logistic regression with distributed data sources the2014 IEEE Healthcare Innovation Conference (HIC), via homomorphic encryption,” IEICE - Transactions on Info pp. 251–254, Seattle, USA, September 2014. and Systems, vol. 99, no. 8, pp. 2079–2089, 2016. [142] C. Balakrishna, E. Rendon-Morales, R. Aviles-Espinosa, [158] H. Chen, R. Gilad-Bachrach, K. Han et al., “Logistic re- D. Henry, and Z. Luo, “Challenges of wearable health gression over encrypted data from fully homomorphic en- monitors: a case study of foetal ECG monitor,” in Pro- cryption,” BMC Medical Genomics, vol. 11, no. 4, pp. 3–12, ceedings of the 2019 global IoT summit (GIoTS), pp. 1–6, Aarhus, USA, June 2019. [159] Y. Aono, T. Hayashi, L. T. Phong, and L. Wang, “Scalable and [143] E. Adi, A. Anwar, and Z. Baig, “Machine learning and data secure logistic regression via homomorphic encryption,” in analytics for the IoT,” Neural Computing & Applications, Proceedings of the Sixth ACM Conference on Data and Ap- vol. 32, Article ID 16205, 2020. plication Security and Privacy, pp. 142–144, Washington, [144] F. Alsubaei, A. Abuhussein, and S. Shiva, “Security and USA, April 2016. privacy in the internet of medical things: taxonomy and risk [160] J. L. H. Crawford, G. Craig, S. Halevi, D. Platt, and V. Shoup, assessment,” in Proceedings of the 2017 IEEE 42nd Conference “Doing real work with FHE: the case of logistic regression,” on Local Computer Networks Workshops (LCN Workshops), in Proceedings of the 6th Workshop on Encrypted Computing pp. 112–120, Singapore, Octobar, 2017. & Applied Homomorphic Cryptography, pp. 1–12, Toranto, [145] J. Padgette, J. Bahr, M. Batra et al., “NIST Special Publication Canada, October 2018. 800-121 Revision 2,” Guide to BluetoothSecurity. [161] S. Carpov, N. Gama, M. Georgieva, and J. R. T. Pastoriza, [146] L. Xiao, X. Wan, X. Lu, Y. Zhang, and Di Wu, “IoT security “Privacy-preserving semi-parallel logistic regression training techniques based on machine learning: how do IoT devices with fully homomorphic encryption,” BMC Medical Geno- use AI to enhance security?” IEEE Signal Processing Maga- mics, vol. 13, no. 7, pp. 1–10, 2020. zine, vol. 35, no. 5, pp. 41–49, 2018. [162] S. Kim, M. Omori, T. Hayashi, T. Omori, L. Wang, and [147] R. N. Reith, T. Schneider, and O. Tkachenko, “Eﬃciently S. Ozawa, “Privacy-preserving naive bayes classiﬁcation stealing your machine learning models,” in Proceedings of the using fully homomorphic encryption,” in Proceedings of the 18th ACM Workshop on Privacy in the Electronic Society, International Conference on Neural Information Processing, WPES’19, pp. 198–210, Association for Computing Ma- pp. 349–358, Springer, Siem Reap, Cambodia, December chinery, New York, NY, USA, November 2019. [148] K. Hansel, ¨ K. Katevas, O. Guido, D. C. Richardson, [163] Y. Yasumura, I. Yu, and H. Yamana, “Secure Na¨ıve Bayes A. Alomainy, and H. Haddadi, “*e potential of wearable classiﬁcation protocol over encrypted data using fully ho- technology for monitoring social interactions based on in- momorphic encryption,” in Proceedings of the 21st Inter- terpersonal synchrony,” in Proceedings of the 4th ACM national Conference on Information Integration and Web- Workshop on Wearable Systems and Applications, WearSys based, pp. 45–54, Applications & Services, Munich, Ger- ’18, pp. 45–47, Association for Computing Machinery, New many, December 2019. York, NY, USA, November 2018. [164] Li Jing, X. Kuang, S. Lin, Ma Xu, and Yi Tang, “Privacy [149] B. Amine, A. Derhab, and Y. Challal, “A review of privacy- preservation for machine learning training and classiﬁcation preserving techniques for deep learning,” Neurocomputing, based on homomorphic encryption schemes,” Information vol. 384, pp. 21–45, 2020. Sciences, vol. 526, pp. 166–179, 2020. [150] A. Qayyum, J. Qadir, M. Bilal, and A. Al-Fuqaha, “Secure and [165] H. Jung, D. Kim, and J. H. Park, “Towards a practical cluster robust machine learning for healthcare: a survey,” IEEE analysis over encrypted data,” in Proceedings of the Inter- Reviews in Biomedical Engineering, vol. 14, pp. 156–180, 2021. national Conference on Selected Areas in Cryptography, [151] C. Dwork, “Diﬀerential privacy,” in Proceedings of the 33rd pp. 227–249, Springer, Waterloo, Canada, August 2019. International Colloquium on Automata, Languages and [166] V. Costan and S. Devadas, “Intel SGX explained,” IACR Programming, part II (ICALP 2006), pp. 1–12, Springer- Cryptol. ePrint Arch.vol. 2016, no. 86, pp. 1–118, 2016. Verlag, Venice, Italy, July 2006. [167] T. Hunt, C. Song, R. Shokri, V. Shmatikov, and E. Witchel, [152] B. Jayaraman and D. Evans, “Evaluating diﬀerentially private “Chiron: privacy-preserving machine learning as a service,” machine learning in practice,” in Proceedings of the 28th 2018. USENIX Security Symposium (USENIX Security 19), [168] C. Segarra, R. Delgado-Gonzalo, M. Lemay, P.-L. Aublin, pp. 1895–1912, USENIX Association, Santa Clara, CA, USA, P. Peter, and V. Schiavoni, “Using trusted execution envi- August 2019. ronments for secure stream processing of medical data,” in Journal of Healthcare Engineering 25 Proceedings of the IFIP International Conference on Dis- ’19, pp. 41–46, Association for Computing Machinery, New York, NY, USA, June, 2019. tributed Applications and Interoperable Systems, pp. 91–107, Springer, Kongens Lyngby, Denmark, June 2019. [186] D. Liaqat, R. Wu, A. Gershon, H. Alshaer, R. Frank, and E. de Lara, “Challenges with real-world smartwatch based [169] M. Nassar, A. Erradi, F. Sabri, and Q. M. Malluhi, “Secure Outsourcing of Matrix Operations as a Service,” in Pro- audio monitoring,” in Proceedings of the 4th ACM Workshop on Wearable Systems and Applications, WearSys ’18, ceedings of the 2013 IEEE Sixth International Conference on pp. 54–59, Association for Computing Machinery, New Cloud Computing, pp. 918–925, Santa Clara, CA, USA, June York, NY, USA, November, 2018. [187] M. R. Nakhkash, T. N. Gia, I. Azimi, A. Anzanpour, [170] R. Lazzeretti and M. Barni, “Private computing with garbled A. M. Rahmani, and P. Liljeberg, “Analysis of performance circuits [applications corner],” IEEE Signal Processing and energy consumption of wearable devices and mobile Magazine, vol. 30, no. 2, pp. 123–127, 2013. gateways in IoT applications,” in Proceedings of the Inter- [171] Y. Wang, Q. M. Malluhi, and K. M. D. Khan, “Garbled national Conference on Omni-Layer Intelligent Systems, computation in cloud,” Future Generation Computer Sys- COINS ’19, pp. 68–73, Association for Computing Ma- tems, vol. 62, pp. 54–65, 2016. chinery, New York, NY, USA, May, 2019. [172] Adi Shamir, “How to share a secret,” Communications of the ACM, vol. 22, no. 11, pp. 612-613, 1979. [173] M. Kesarwani, B. Mukhoty, V. Arya, and S. Mehta, “Model extraction warning in MLaaS paradigm,” in Proceedings of the 34th Annual Computer Security Applications Conference, pp. 371–380, San Juan, PR, USA, December 2018. [174] H. Brendan McMahan, E. Moore, D. Ramage, and y A. Blaise Aguera, ¨ “Federated Learning of Deep Networks Using Model Averaging,” CoRR, abs/1602, Article ID 05629, 2016. [175] K. Peter, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, and A. N. Bhagoji, Advances and Open Problems in Federated Learning, 2019, Article ID 04977. [176] Y. S. Can and C. Ersoy, Privacy-Preserving Federated Deep Learning for Wearable IoT-Based Biomedical Monitoring, vol. 21, no. 1, 2021. [177] J. Xu, B. S. Glicksberg, and C. Su, “Federated learning for healthcare informatics,” J Healthc Inform Res, vol. 5, no. 1–19, 2021. [178] S. Truex, N. Baracaldo, A. Ali et al., “A hybrid approach to privacy-preserving federated learning,” in Proceedings of the 12th ACM Workshop on Artiﬁcial Intelligence and Security, pp. 1–11, Association for Computing Machinery, New York, NY, USA, November, 2019. [179] Y. Chen, X. Qin, J. Wang, C. Yu, and W. Gao, “FedHealth: a federated transfer learning framework for wearable health- care,” IEEE Intelligent Systems, vol. 35, no. 4, pp. 83–93, 2020. [180] D. Batista, P. Hugo, A. Fred, C. Moreira, M. Reis, and H. A. Ferreira, “Benchmarking of the BITalino biomedical toolkit against an established gold standard,” Healthcare Technology Letters, vol. 6, no. 32–36, 2019. [181] M. Huang, D. Chen, and X. Fan, “An eﬀective adaptive ﬁlter to reduce motion artifacts from ECG signals using accel- erometer,” in Proceedings of the 2019 9th International Conference on Biomedical Engineering and Technology, pp. 83–88, Association for Computing Machinery, New York, NY, USA, March 2019. [182] D. Pollreisz and N. Taherinejad, “Detection and Removal of Motion Artifacts in PPG Signals,” Mobile Networks and Applications, vol. 24, 2019. [183] L. M. Stevens, B. J. Mortazavi, R. C. Deo, L. Curtis, and D. P. Kao, “Recommendations for reporting machine learning analyses in clinical research,” Circulation: Cardio- vascular Quality and Outcomes, vol. 13, no. 10, Article ID e006556, 2020. [184] J. W. K. S. Moon, S.U. Kang, and B. Jang, “Eﬀective privacy- preserving collection of health data from a user’s wearable device,” Applied Sciences, vol. 10, no. 18, 2020. [185] C. Kang, H. Jung, and Y. Lee, “Towards machine learning with zero real-world data,” in Proceedings of the Ge 5th ACM Workshop on Wearable Systems and Applications, WearSys

Journal

Journal of Healthcare Engineering – Hindawi Publishing Corporation

Published: Apr 18, 2022

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Machine Learning for Healthcare Wearable Devices: The Big Picture

Machine Learning for Healthcare Wearable Devices: The Big Picture

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Machine Learning for Healthcare Wearable Devices: The Big Picture

Machine Learning for Healthcare Wearable Devices: The Big Picture

References (197)

Abstract

Journal

Recommended Articles

There are no references for this article.

Our policy towards the use of cookies