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

Learn More →

Design of a Machine Learning-Assisted Wearable Accelerometer-Based Automated System for Studying the Effect of Dopaminergic Medicine on Gait Characteristics of Parkinson’s Patients

Design of a Machine Learning-Assisted Wearable Accelerometer-Based Automated System for Studying... Hindawi Journal of Healthcare Engineering Volume 2020, Article ID 1823268, 11 pages https://doi.org/10.1155/2020/1823268 Research Article Design of a Machine Learning-Assisted Wearable Accelerometer-Based Automated System for Studying the Effect of Dopaminergic Medicine on Gait Characteristics of Parkinson’s Patients 1 2 3 Satyabrata Aich , Pyari Mohan Pradhan, Sabyasachi Chakraborty , 1 4 5 6 1 Hee-Cheol Kim , Hee-Tae Kim, Hae-Gu Lee, Il Hwan Kim, Moon-il Joo, 1 7 Sim Jong Seong, and Jinse Park Institute of Digital Anti-Aging Healthcare, Inje University, Gimhae, Republic of Korea Department of Electronics and Communication Engineering, IIT, Roorkee, India Department of Computer Engineering, Inje University, Gimhae, Republic of Korea Department of Neurology, Hanyang University Hospital, College of Medicine, Seoul, Republic of Korea Department of Industrial Design, Kyoung Sung University, Busan, Republic of Korea Department of Oncology, Haeundae Paik Hospital, Inje University, Busan, Republic of Korea Department of Neurology, Haeundae Paik Hospital, Inje University, Busan, Republic of Korea Correspondence should be addressed to Jinse Park; jinsepark@gmail.com Received 24 September 2019; Revised 15 December 2019; Accepted 8 January 2020; Published 18 February 2020 Guest Editor: Chao Chen Copyright © 2020 Satyabrata Aich 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. In the last few years, the importance of measuring gait characteristics has increased tenfold due to their direct relationship with various neurological diseases. As patients suffering from Parkinson’s disease (PD) are more prone to a movement disorder, the quantification of gait characteristics helps in personalizing the treatment. +e wearable sensors make the measurement process more convenient as well as feasible in a practical environment. However, the question remains to be answered about the validation of the wearable sensor-based measurement system in a real-world scenario. +is paper proposes a study that includes an al- gorithmic approach based on collected data from the wearable accelerometers for the estimation of the gait characteristics and its validation using the Tinetti mobility test and 3D motion capture system. It also proposes a machine learning-based approach to classify the PD patients from the healthy older group (HOG) based on the estimated gait characteristics. +e results show a good correlation between the proposed approach, the Tinetti mobility test, and the 3D motion capture system. It was found that decision tree classifiers outperformed other classifiers with a classification accuracy of 88.46%. +e obtained results showed enough evidence about the proposed approach that could be suitable for assessing PD in a home-based free-living real- time environment. features of gait in PD are short-step, hypokinetic, slow gait 1. Introduction with decreased arm swing, and episodic gait, which includes +e most important symptom of Parkinson’s disease (PD) is freezing of gait (FOG) and festinating gait [3]. Despite the the disturbances in gait that directly affects the daily ac- clinical importance, most clinicians usually depend on tivities as well as the quality of life [1]. +e disturbances in neurological examination or self-questionnaire-based ex- gait characteristics in PD patients are categorized into amination for a short period of time. +erefore, it is very continuous gait and episodic gait disturbances [2]. Typical difficult to assess the PD patient’s gait status outside the 2 Journal of Healthcare Engineering consumption in wearable devices and hence a longer clinic and in a real-world environment. Objective quanti- fication of gait is crucial for the measurement of overall battery life for gait monitoring. condition as well as disease monitoring in PD. Several (4) +e validation study provides a low-cost alternative clinical scales such as Tinetti mobility test (TMT), Timed Up for assessing gait characteristics in the “on” state of PD and Go (TUG), and Unified Parkinson’s Disease Rating patients for both indoor and outdoor environments. Scales (UPDRS) are widely used to assess the PD and its (5) +e proposed study demonstrates that spatiotem- severity. In the last decade, numerous studies have inves- poral gait characteristics estimated by using only tigated the usefulness of gait analysis. Quantitative gait accelerometer data are highly correlated with those analysis includes infrared-based motion capture (three-di- obtained from a 3D motion capture system. Fur- mensional (3D) motion capture), pressure-based gait thermore, a high correlation was also found between analysis (GAITRite), and treadmill gait analysis [4–6]. De- results obtained from the proposed approach and spite their strength of accurate quantification of gait, clinical those obtained from the clinical TMT test. implication is still controversial due to high cost and large (6) +e proposed study proposed an automatic system space or laboratory required for system set up. that can classify PD patients and HOG with machine To overcome the previous limitations, an attempt has learning techniques based on gait characteristics. been made in this study to quantify the gait characteristics using the algorithmic-based approach with a wearable ac- +e structure of the paper is outlined in the following celerometer and its validation using a 3D motion capture way: Section 2 describes the past work related to this study. system as well as TMT. TMT is widely used for predicting the Section 3 describes the data collection methods as well as the fall risk of elderly people based on the balance and gait test proposed methodology. Section 4 presents the results and score. TMT test consists of two components such as the outcomes of the proposed approach. Section 5 provides the Tinetti balance scale and the Tinetti gait scale. +e balance discussion. Section 6 describes the conclusion. scale consists of 9 parameters, and each parameter has subparameters with a score of 0/1 or 0/1/2. +e total possible 2. Related Work score of the balance section is 16. +e gait scale consists of 8 parameters, and each parameter has subparameters with a +e gait analysis performed using a conventional way using a score of 0/1 or 0/1/2. +e total possible score of the gait qualitative analysis technique is usually performed in the section is 12. Each patient has to be assessed based on these clinics, and it required a complete medical history of the two scores. +e combined score determines the risk of falls patients to determine the gait characteristics. +e conven- in elderly people. According to Tinetti, a total score of ≤18 is tional method is relatively simple; however, it depends on treated as high risk, 19–23 is treated as moderate risk, and the expertise of the physicians, and it is relatively difficult to ≥24 is treated as low risk [7]. Since the Korean version of measure the parameters in a quantitative manner with high TMT has already been validated with the PD patients in the accuracy that could be useful for clinical applications. To laboratory [8], this version has been used in this study. address this aforementioned problem, a new method has +e contributions of the proposed study are as follows: been introduced in this paper to quantify the gait charac- teristics in an objective way by using quantitative mea- (1) +is study includes enrolment of a large number of surement techniques [10, 11]. Wearable devices are now participants with PD, higher than that recommended used for a wide range of healthcare observations as well as by the movement disorder society [9]. While the the measurement of gait. +e triaxial wearable accelerometer recommended minimum number of patients is 30, is known to be a useful tool for assessing gait as well as this study involves 48 PD patients to provide proper various motor symptoms in PD. It is not expensive as well as validity and reliability of the result. In addition, 40 can be used in a comfortable way by the user [12]. Beck et al. healthy older patients’ group has been included in proposed a new approach to quantify the gait smoothness the study for the classification of PD subjects from using accelerometer and gyroscope signals. +ey have healthy older group based on estimated gait char- implemented this method in PD patients as well as healthy acteristics. Due to a large number of subjects, the controls, and they found clear differentiation in terms of proposed study could be recommended for a real-life smoothness between two groups. For validation, they have scenario. used the correlation technique by comparing their algorithm (2) +e proposed study focuses on the PD patients when spectral arc length measure (SPARC) with traditional gait they are clinically in “on” state, i.e., after taking measures and the UPDRS scale. +is is one of the potential dopaminergic medicine. “On” state is the state where use cases for using wearable sensors; however, their method the effect of the medicine is present, and the im- did not use 3D motion capture and TMT gait scale for provement in the gait characteristics is closer to the correlation [13]. Hausdorff et al. mentioned that quantifi- healthy older group. cation of gait characteristics was possible using a wearable (3) +e good accuracy found by using only accelerometer device. +ey have collected the accelerometer data during data for estimating spatiotemporal gait characteristics the tandem walking and also validated the method. +is indicates that the gyroscope data could be excluded method also mentioned the potential of using wearable for these kinds of studies. +is will lead to low power devices for gait analysis. +ey have not implemented this Journal of Healthcare Engineering 3 good level of correlation with a correlation coefficient that method to the PD patients, and at the same time, they have not used any other methods such as 3D motion capture or ranges from 0.961 to 0.984. +e study also proposed a machine learning-based approach to distinguish PD with clinical scale for correlation of their method [14]. Gazit et al. proposed a method for quantifying gait initiations using FoG from PD with no FoG, and an accuracy of 88% has been wearable sensors. +ey have used only one IMU sensor for found using SVM classifier. In this research, the effect of evaluating the gait initiations and found good results. +ey dopaminergic medicine has not been considered, and the have validated the method with the ground truth and found correlation analysis has not been performed with clinical that the interclass correlation coefficient with one wearable scale [21]. Mikos et al. proposed a method for FoG detection sensor ranges from 0.75 to 0.96. +ey have tested this using a single sensor node. +ey have developed a system using machine learning based on the extracted features from method on the data collected from younger and older adults. +ey have not used the 3D motion capture system for the signals. +ey have found a classification accuracy of 92.9% in average of sensitivity and specificity when validation of their results and also not implemented for PD patients [15]. Anwary et al. proposed a method to find the exploiting its patient adaptive learning capability. +is re- search has given enough evidence that a single sensor can be best location in the foot to place wearable sensors. +ey have used accelerometer data and gyroscopic data for deter- used for the detection of FoG and machine learning systems mining the gait features. For validation of this method, they for the classification of FoG [22]. Jeon et al. proposed a study have used a quality motion capture system. +ey have done that used the wearable device to detect the severity of tremor this analysis for healthy groups and mentioned that wearable in PD. +e wearable device used in this study consists of an sensors have the potential to quantify the gait characteristics accelerometer and gyroscope. +is study also used machine with high accuracy [16]. Qiu et al. proposed a method that learning techniques to classify the severity of tremor based on the score of UPDRS. It was found that the decision tree used body-worn sensors to collect the gait data for the as- sessment of stroke patients. +ey have found that the gait outweighs other classifiers with an accuracy of 85.5%. +is research has provided enough evidence that the wearable analysis has a huge contribution towards the diagnosis and treatment of the stroke patients and mentioned that a sensors can be used for the diagnosis of PD, and machine learning techniques can be used to automate the system [23]. wearable sensor-based gait analysis system has the potential for supporting rehabilitation in the clinics and hospitals [17]. Sama` et al. proposed a study using wearable accelerometer Byun et al. have proposed a method that uses the wearable that can detect freezing of gait at real-time environment accelerometer to measure the gait characteristics of older using a set of features which are related to the previous people having normal cognition. +e gait characteristics are approaches mentioned by the previous researchers. +ese quantified using the signal-processing algorithm. Validation features were trained using machine learning classifiers and of the measurement method was carried out using the used to detect the FoG with an improvement over the previous methods. +is research suggested that the wearable GAITRite system. +e two methods show a good level of correlation with a correlation coefficient that ranges from sensor has the potential to be used for measuring the gait characteristics, and machine learning techniques could be 0.91 to 0.96. +ey have not used the 3D motion capture system, and this method was not tested for PD patients [18]. used for the detection of the PD group [24]. Pham et al. have proposed a technique that used an inertial +e past works mentioned above provide a strong measurement system which consists of a gyroscope and an recommendation about the use of the wearable device in the accelerometer to detect the gait patterns such as toe-off and field of PD as well as the effective use of machine learning heel strike in the patients with PD as well as older adults techniques for autodetection of gait patterns in PD and when they were encountered with turning as well as straight HOG. +e proposed approach has got a lot of inspiration walking. An algorithm based on continuous wavelet from the previous pieces of literature cited by different transform is used to detect the gait patterns, and the vali- researchers. In this study, an algorithmic-based approach has been developed, and it was validated using clinical test dation study was carried out using the optoelectronic sys- tem. +ey have not used any clinical scale for comparing the and well-known measuring instruments, and a machine learning-based approach has been proposed to detect the PD result. 3D motion capture has not been used in this research [19]. Del Din et al. have used the wearable accelerometer for from the healthy older group using estimated gait charac- measuring the gait characteristics of older adults as well as teristics. +is system is developed by keeping in mind that it PD patients. Signal processing of the collected accelerometer can be used in the home environment as well as in clinical data provides gait characteristics, and the validation was environments. carried out using the instrumented walkway. Fourteen gait characteristics were compared; it was found that four 3. Proposed Methodology characteristics show a good amount of correlation, another four gait characteristics show an agreement of moderate 3.1. Data Collection. +is study was performed clinically in level, and the rest six characteristics show an agreement of a the “on” state, i.e., after taking dopaminergic medicine for poor level. +is paper does not have any correlation analysis the PD group of patients. “On” state is the state where there of gait characteristics with the clinical scale, and they did not is an effect of the medicine. In this state, there is an im- use the 3D motion capture system [20]. Aich et al. proposed provement in gait characteristics. +e resulting gait char- a method that used a wearable accelerometer that can detect acteristics are very similar to those of the healthy older FoG, and the validation study was performed that shows a group. +e accelerometer data for PD patients have been 4 Journal of Healthcare Engineering Table 1: Details of the PD group. collected in the “on” state so as to study the difference between two groups, i.e., PD patients and healthy older M/F (n � 48) 25/23 Age 70.61 ± 9.51 group when they are in a similar state. +is study was UPDRS part III 20.9 ± 12.31 performed at Haeundae Paik Hospital located at Busan, H&Y stage 2.10 ± 0.74 South Korea. +e approval was taken from the review board Disease duration (months) 35.49 ± 27.07 of the institute (IRB No. 2017-01-028). Prior approval has Timed-up and go 20.87 ± 15.78 been taken from all participants before joining this study. Tinetti gait scale 9.86 ± 2.56 +e details about the PD group are shown in Table 1. +e healthy older group comprises normal persons with no signs of PD. No medication has been given to them prior Table 2: Details of the healthy group. to this study. +e healthy group consists of 22 males and 18 M/F (n � 40) 22/18 females. All the subjects in the healthy group were age- Age 69.36 ± 7.42 matched. +e details regarding the patients belonging to the UPDRS part III 0 healthy group are shown in Table 2. Disease duration (months) 0 UPDRS and H&Y represent Unified Parkinson’s Disease Rating Scale and Hoehn and Yahr scale, respectively. UPDRS is widely used for checking the severity of the disease [25]. H&Y scale is a clinical rating scale, which is used to define different categories of motor functions in the PD [26]. Tinetti gait scale is widely used for predicting the fall risk of elderly people. +e participants were asked to wear the accelerometer on the left knee as well as the right knee. Two wearable triaxial accelerometers with a sampling frequency of 32 Hz (Fit Meter, Fit. Life, Suwon, Korea) were used. +e triaxial ac- celerometer measures body movements in all directions: anterior-posterior, mediolateral, and vertical. It is small and lightweight (35 mm × 35 mm × 13 mm and 13.7 gm). It is sensitive to acceleration from − 8 g to 8 g, allowing for Figure 1: Location of the accelerometers specified for the proposed monitoring of almost all human physical activities. All the study. participants wore the accelerometers at a distance of 34 cm from the ground, as shown in Figure 1. All the participants were asked to walk along a six-meter track. For validation of acceleration signal. +e filtered signal was integrated for gait the proposed approach, the gait characteristics were also event detection. +e objective was to detect the initial measured by using the 3D motion analysis system (VICON, contacts (ICs) of the leg, which are also termed as the heel- Oxford, UK). +e motion was captured during the walking strike event in a gait cycle. +e locations of ICs were detected process. Five important gait characteristics were measured from the points of minima in the smoothed signal by de- that include step time, stride time, step length, stride length, termining the first-order derivative using the Gaussian and walking speed. For estimating gait status more objec- continuous wavelet transform. +e flowchart of the pro- tively, the Korean version of the Tinetti gait scale [7] was posed algorithmic approach based on the accelerometer data used. +e gait characteristics obtained from the 3D motion is shown in Figure 2. system and the Tinetti gait scale were used for validation of In this study, five gait characteristics such as step time, the proposed approach. stride time, step length, stride length, and walking speed were estimated for the feasibility study of objective assessment of PD using wearable accelerometer data. +ese five charac- 3.2. Estimation of Gait Characteristics. A variant of the teristics have received great attention from the researchers in method proposed by Del Din et al. [12] was used to detect the gait-related study and its effectiveness for the assessment of gait cycle. +e measured acceleration values along X-, Y-, PD. Five major domains of gait study have been proposed by and Z-axes represent linear accelerations along the medial- Hollman et al. using the factor analysis: (1) step time and lateral (ML), anterior-posterior (AP), and vertical (V) di- stride time represented by the rhythm domain; (2) tempor- rections, respectively. +e corrections are needed to over- ophasic domain of gait cycle represented by the phase do- come the effect of gravitational component, error due to main; (3) step variability represented by the variability imprecise position of wearable accelerometer, etc. [27]. +e domain; (4) step length, stride length, and gait speed rep- dynamic tilt correction approach proposed in [27] was used resented by the pace domain; (5) step width represented by to transform the acceleration from ML and AP directions to the base of the support domain [28]. +e aforementioned five characteristics have also been used recently to detect the FoG a global horizontal-vertical coordinate system. +e resulting vertical acceleration signal was used hereafter for gait event [20]. Walking speed, stride length, and stride time have been identification. A low-pass fourth-order Butterworth filter given high importance by Schlachetzki et al. [29] for the with a cutoff frequency of 15 Hz was used to filter the vertical discrimination of healthy subjects from the PD subjects. Journal of Healthcare Engineering 5 Collection of acceleration data using triaxial wearable accelerometer at knee of left and right legs Preprocessing of raw acceleration data to eliminate the offsets and misalignments Filtering of acceleration data for the Filtering of acceleration data for the left leg to remove noise right leg to remove noise Detection of peaks (minima) Detection of peaks (minima) representing initial contact (IC) of the representing initial contact (IC) left leg using continuous wavelet of the right leg using continuous transform wavelet transform Calculation of spatiotemporal Calculation of spatiotemporal parameters for the left leg using the parameters for the right leg using above data for left leg the above data for right leg Comparison of result with  3D motion analysis system (gold standard) Data Interpretation and validation of results using statistical analysis Figure 2: Flowchart of the algorithmic approach for the estimation of gait characteristics. Bertoli et al. estimated the spatiotemporal parameters such as the inverted pendulum model [21, 31], as shown in Figure 3. stride time, step time, swing time, stance time, stride length, +e step length and stride length can be computed as and gait velocity for the quantitative assessment of PD, mild follows: 􏽱������������ cognitive impairment patients, and healthy older adults [30]. step length � K ∗ 2 2W H − H 􏼁, I h +e step time can be calculated based on the IC events (3) [9] as follows: stride length � 2 ∗ step length, step time(i) � IC(i + 1) − IC(i). (1) where W represents the distance from the ground to the wearable accelerometer and H represents the change in Similarly, the stride time can also be computed based on height of the wearable sensor between two consecutive IC the IC events [9] as follows: events. +is is computed by finding the difference between the maximum and minimum values of the double integrated stride time(i) � IC(i + 2) − IC(i), (2) vertical acceleration signal between two IC events. +e generic multiplying factor K is used for mapping the center where i denotes the index of the IC event in the signal. In the proposed approach, the step length has been estimated using of mass in an inverted pendulum model with that of the K W 1 h 6 Journal of Healthcare Engineering 3.3.2. Support Vector Machine (SVM) Classifier. SVM is one of the classifiers suitable to deal with binary classification problems. +e classifier tries to maximize the margin ar- ithmetically between two input datasets by defining a surface in an input space, which is multidimensional in nature [34]. In another way, SVM selects the hyperplane with the highest possible margin between two classes while separating them. It is impossible for a hyperplane to separate the data between two classes, but it tries to separate as much data as possible to provide good accuracy [35]. In this study, the radial basis kernel function is used, which provides good accuracy compared to other available kernel functions. 3.3.3. Na¨ıve Bayes (NB) Classifier. NB classifier is one of the simple probabilistic classifiers based on the Bayes’ theorem. +is classifier selects mutually independent variables. +is kind of classifier can be employed in the complex real-life scenario as it can be trained efficiently using the supervised IC le Step length IC right learning technique. +e advantage of this algorithm is that it Figure 3: Extended inverted pendulum model [20] for estimation needs less amount of data for training purposes to perform of step length. the classification task. In this study, the classification task has been performed by using the Bayes’ rule to calculate the probability of class label PD or a healthy group [36]. wearable sensor. +e value of K will change based on the value of W . +erefore, to avoid the time-consuming task of mapping for each participant that requires determining K 3.3.4. Decision Tree Classifier. +e decision tree classifier for each participant, W has been fixed at 34 cm, and cor- works on the basis of conditional statements and its possible respondingly, K � 4 has been chosen for this study. Walking consequences. It is a tree-like model. Nodes and branches speed is calculated as follows [21]: are the primary components to build a decision tree model. +ree steps are followed for building a well-designed de- mean step length walking speed � . (4) cision tree model. +e first step is splitting, followed by mean step time stopping, and then finally pruning. +e continuation of the splitting process stops when the model reaches the desired +ese aforementioned five estimated gait characteristics stopping criteria. +e stopping rule is used to avoid the were used as features for the classification of PD groups and problem of overfitting and underfitting. If the stopping rule healthy older group. does not work well, the pruning method is used to improve the overall classification accuracy [37]. A planned-designed PD detection framework should be 3.3. Machine Learning Classifiers and Its Effectiveness for 8is efficient and quick enough to perform the binary classifi- Study. In this study, comparative performance analysis has cation for the classification of PD patients from the healthy been carried out between four machine learning classifiers that have been employed to perform the classification task older group. Accuracy, sensitivity, and specificity are widely used to measure the effectiveness of the system. +e amount between the PD patients and the healthy control adults. of correctness required for the distinction of PD patients from the healthy older group could be measured using the 3.3.1. 8e k-Nearest Neighbour Classifier (k-NN). +e k-NN term accuracy. +e potential to identify PD is measured by classifier performs the classification process based on the sensitivity, and it is usually expressed as the ratio of true positives to the total number of PD patients [21]. +e po- proximity of a data point to the nearest training data points. It generally measures the Euclidean distance to measure the tential to identify PD when the system identifies the PD can be measured by the term specificity. +e subjects belong to closeness between them. +e local data structure has a strong influence on the k-NN algorithm. +ere is no standardized the PD group, correctly identified as PD subject, and are rule to define the value of k. +e classes are selected based on represented as true positives. +e subjects belong to a the majority rule from among the selected number of k- healthy older group, correctly identified as healthy older nearest neighbors, where k is always greater than zero and an groups, and are represented as true negatives. +e subjects integer. +e instability in the result, as well as an increase in belong to the healthy older group but wrongly identified as the variance, can be seen with the smaller values of k. +e PD subjects are represented as false positives. +e subjects belong to the PD group but wrongly identified as the healthy reduction in sensitivity, as well as increasing bias, can be seen with the higher values of k. In general, the k values are older group are represented as false negatives. In this study, the objective is to reduce the false negatives as it affects the chosen depending on the dataset. In this study, a value of k � 5 is chosen as it provides good accuracy [32, 33]. effectiveness of the system. h Journal of Healthcare Engineering 7 Table 3: Mean value of gait characteristics and average error rate for the left and right legs. Sl. no. Parameters Mean value (3D motion capture) Mean value (algorithm) Mean error rate (%) Left leg 1 Step time (s) 0.57 0.54 6.94 ± 2.82 2 Stride time (s) 1.17 1.13 4.76 ± 3.55 3 Step length (m) 0.37 0.34 6.35 ± 2.85 4 Stride length (m) 0.74 0.71 6.51 ± 2.92 5 Walking speed (m/s) 0.64 0.61 7.12 ± 2.74 Right leg 1 Step time (s) 0.54 0.56 7.14 ± 2.52 2 Stride time (s) 1.18 1.14 5.25 ± 3.62 3 Step length (m) 0.37 0.34 6.15 ± 2.81 4 Stride length (m) 0.74 0.70 6.35 ± 2.71 5 Walking speed (m/s) 0.69 0.66 6.72 ± 3.14 plots between the algorithmic approach as well as a 3D 4. Results motion capture system are shown in Figures 4–8. +e mean +e mean value of five estimated gait characteristics based error rate was calculated based on the formula [21] as on the accelerometer data as well as the mean error rate follows: between the algorithmic approach and the 3D motion capture system are highlighted in Table 3. +e correlation (value estimated from acc) − (value estimated from 3D capture) average error rate(%) � ∗ 100. (5) (value estimated from 3D capture) +e total number of subjects including both the groups is 88. In this paper, Tinetti mobility test (TMT) gait scale is used to assess the spatiotemporal gait characteristics such Out of 88 subjects, 62 subjects belong to the training group, as step time, stride time, step length, stride length, and and the rest 26 belong to the validation group. Out of 26 walking speed and its importance in terms of clinical subjects, which belong to the validation group, 14 subjects practices by comparing the score with the result obtained belong to the PD group (PDG) and 12 subjects belong to the using other methods, in this case, computerized gait healthy older group (HOG). Moreover, a 5 split cross-vali- analysis using accelerometer data and 3D motion capture dation was also performed based on the subject’s data to check system. +e true changes in the gait characteristics can be the generalizability of the model. +e cross-validation was easily understand based on the accuracy of the clinical performed in such a way where the data of 62 random subjects observation measures, and it is an important step in clinical were used to train a classifier and the rest data of 26 subjects practices. So, in this paper, we have used Pearson’s cor- were used for checking the testing accuracy. Test set 1, test set 2, test set 3, and test set 4 consist of 26 subjects each. +e relation coefficient to analyze the relationship between the TMT gait scale score, and spatiotemporal gait character- cross-validation was performed using 4 different classifiers, istics derived objectively used computerized gait analysis namely, KNN, SVM, Naive Bayes, and decision tree. +e using accelerometer data and 3D motion capture system. implementation of four different algorithms was done to +e correlation plots between TMT gait scale and gait perform a comparative analysis between the classifiers. After characteristics measured from the 3D motion capture successful cross-validation, it was found that the decision tree system are shown in Figures 4–8. We have found strong plotted the best set of results by prompting a maximum correlations between them, and the results were mentioned accuracy of 88.46%, sensitivity of 92.86%, and specificity of as follows: step time (0.96, p < 0.01), stride time (0.97, 90.91%, respectively. Table 4 shows the results for the cross- p < 0.01), step length (0.98, p < 0.01), stride length (0.99, validation. p < 0.01), and walking speed (0.99, p < 0.01). Similarly, the +e classifiers’ performance has been evaluated using three parameters such as accuracy, sensitivity, and speci- correlation plots between TMT gait scale and gait char- acteristics obtained from the wearable accelerometer data ficity. +e classification results are shown in Table 5. +e are shown in Figures 9–13. We have found moderate to decision tree classifier could able to provide the highest strong correlations between them, and the results were accuracy of 88.46% with a sensitivity of 0.9286 and specificity mentioned as follows: step time (0.57, p < 0.01), stride time of 0.9091. From 14 subjects belonging to PDG, the proposed (0.54, p < 0.01), step length (0.84, p < 0.01), stride length model correctly identified 13 as PDG. Similarly, from the 12 (0.84, p < 0.01), and walking speed (0.75, p < 0.01). subjects belonging to HOG, the proposed model correctly +is study used the split named as stratified train-vali- identified 10 as HOG. +e confusion matrix is shown in dation [21] with a ratio of 70 : 30 for training and validation. Figure 14. 8 Journal of Healthcare Engineering 1 1.4 r = 0.96 r = 0.99 1.2 0.9 0.8 0.8 0.7 0.6 0.6 0.4 0.5 0.2 0.4 0 0.2 0.4 0.6 0.8 1 1.2 Stride length, accelerometer 0.3 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Figure 7: Stride length correlation plot between the accelerometer- Step time, accelerometer based approach and 3D motion system ( p < 0.01). Figure 4: Step time correlation plot between the accelerometer- based approach and 3D motion system ( p < 0.01). 1.4 r = 0.99 1.2 r = 0.97 1.8 1.6 0.8 0.6 1.4 0.4 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Walking speed, accelerometer 0.8 Figure 8: Walking speed correlation plot between the acceler- 0.6 ometer-based approach and 3D motion system ( p < 0.01). 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 Stride time, accelerometer Figure 5: Stride time correlation plot between the accelerometer- 0.9 ∗ ∗ r = 0.57 based approach and 3D motion system ( p < 0.01). 0.8 0.7 r = 0.98 1.4 1.2 0.6 0.5 0.8 0.4 0.6 0.3 0.4 02468 10 12 0.2 Tinetti gait scale Figure 9: Step time correlation plot between the accelerometer- 0 0.2 0.4 0.6 0.8 1 1.2 ∗ based approach and Tinetti gait scale ( p < 0.01). Step length, accelerometer Figure 6: Step length correlation plot between the accelerometer- learning approach can be used for automatic detection and based approach and 3D motion system ( p < 0.01). differentiation of PD patients from the healthy older group. Although wearable sensors have been widely used in many fields, these have not been given enough importance in 5. Discussion PD-related assessment due to distorted gait pattern. Sijobert +is study proposes an algorithmic approach to estimate the et al. [38] have proposed a technique that provides a mean gait characteristics of PD subjects as well as the healthy older error rate of 10.3% for the PD group and 6% for the healthy groups. +e approach is validated using measuring instru- group. +e comparison has been made based on the esti- ments and clinical scale. It is also proposed that the machine mated stride length calculated using wearable sensor data Step length, 3D motion capture Stride time, 3D motion capture Step time, 3D motion capture Step time Walking speed, 3D motion capture Stride length, 3D motion capture Journal of Healthcare Engineering 9 1.4 ∗ 1.4 r = 0.54 1.3 r = 0.75 1.2 1.2 1.1 1 0.8 0.9 0.8 0.6 0.7 0.4 0.6 0.2 0.5 0 2 4 6 8 10 12 14 Tinetti gait scale 02468 10 12 14 Figure 10: Stride time correlation plot between the accelerometer- Tinetti gait scale based approach and Tinetti gait scale ( p < 0.01). Figure 13: Walking speed correlation plot between the acceler- ometer-based approach and Tinetti gait scale ( p < 0.01). Table 4: 5 split cross-validation. Performance (%) KNN SVM NB Decision tree Accuracy test set 1 82.11 81.36 84.52 86.28 0.7 r = 0.84 Sensitivity test set 1 0.8746 0.7801 0.8225 0.9152 0.6 Specificity test set 1 0.8452 0.725 0.8654 0.8833 Accuracy test set 2 83.64 84.25 81.20 84.31 0.5 Sensitivity test set 2 0.8055 0.8139 0.8558 0.8631 0.4 Specificity test set 2 0.8519 0.8687 0.8411 0.8551 Accuracy test set 3 86.32 84.93 85.31 82.28 0.3 Sensitivity test set 3 0.9025 0.8755 0.9032 0.8111 Specificity test set 3 0.8947 0.9054 0.8748 0.8364 0.2 Accuracy test set 4 85.57 87.23 84.41 88.46 0.1 Sensitivity test set 4 0.9125 0.9189 0.8956 0.9286 Specificity test set 4 0.8836 0.8997 0.8735 0.9091 02468 10 12 14 Accuracy test set 5 87.26 84.39 79.32 87.32 Tinetti gait scale Sensitivity test set 5 0.8568 0.8793 0.8178 0.9025 Specificity test set 5 0.9034 0.8998 0.8227 0.9131 Figure 11: Step length correlation plot between the accelerometer- based approach and Tinetti gait scale ( p < 0.01). Table 5: Classification results. Performance k-NN SVM NB Decision tree Accuracy (%) 85.57 87.23 84.41 88.46 Sensitivity (%) 0.9125 0.9189 0.8956 0.9286 Specificity (%) 0.8836 0.8997 0.8735 0.9091 1.4 r = 0.84 1.2 PDG HOG 0.8 0.6 13 1 PDG 0.4 0.2 02468 10 12 14 210 HOG Tinetti gait scale Figure 12: Stride length correlation plot between the accelerom- eter-based approach and Tinetti gait scale ( p < 0.01). Predicted labels Figure 14: Confusion matrix. Step length Stride length Stride time Walking speed True labels 10 Journal of Healthcare Engineering and further validated using the GAITRite-based walkway sys- to get recommended for the clinicians to use in the labo- tem. +e estimated mean error rate for the five gait charac- ratory as well as in the home environment. teristics is found to be less than 8% with our proposed approach, In the future, we will collect gait data from a large which used the wearable accelerometer to collect the data. +e number of PD patients to summarize the gait characteristics results of our proposed approach provide the feasibility of our in a better way so that it could be promoted for clinical ap- approach when compared with the previous study. +e pro- plications. We would also like to combine brain EEG signals posed study provides some new ideas that are as follows: with the gait data to understand more about the relation and detect the symptoms like freezing of gait before it happens. We (i) +e results obtained in this study include various would like to combine the MRI image with the gait data for phenotypes and severity of PD due to the large more accurate diagnosis and early detection of PD. sample size. (ii) +is study uses the Tinetti gait scale [7] and the 3D Data Availability motion capture system [39] for validation of gait status. +e previous report [38] has demonstrated +e data used to support the experiments and the finding of the validation using GAITRite that can assess the study have been duly included in Section 3 of the paper. spatiotemporal data by using pressure parameters. Section 3.1 clearly describes about the data. (iii) +e gait characteristics estimated using our pro- posed approach have been compared with the Conflicts of Interest clinical scale, and the result shows a good level of agreement, which makes the method feasible to be +e authors declare that they have no conflicts of interest. implemented in the real-life environment. (iii) It is observed from the study that only acceler- Acknowledgments ometer data can provide enough information for performing the gait analysis, which leads to re- +is research was supported by the National Research dundancy of gyroscopic data, which indirectly saves Foundation (NRF) of Korea grant funded by the Korea battery power, time, and cost. government (MSIT) (Grant number 2019R1C1C1011197) and also funded by Ministry of Trade, Industry and Energy +e strength and possibility of the wearable sensor-based (MOTIE), Korea, through the Education program for Cre- PD assessment are aimed at long-term monitoring of gait. ative and Industrial Convergence (Grant number N0000717). Gait disturbance usually gets aggravated in specific cases such as starting time, meeting narrow space, or obstacle [40]. +e hospital has limited space, and therefore, it is difficult to References replicate gait disturbance observed in a real-world scenario. [1] C. C. Walton, J. M. Shine, J. M. Hall et al., “+e major impact +e battery mounted in our device lasts up to 72 hours. +e of freezing of gait on quality of life in Parkinson’s disease,” limitation of this study is that the patients in the relatively Journal of Neurology, vol. 262, no. 1, pp. 108–115, 2014. early stage of PD were not enrolled, and therefore, this study [2] J. M. Hausdorff, “Gait dynamics in Parkinson’s disease: is not intended at assessing various balance status in PD such common and distinct behavior among stride length, gait as a freeze of gait (FOG) or hypokinetic and short-step gait. variability, and fractal-like scaling,” Chaos: An Interdisci- +e classifiers used in this study have shown good accuracy. plinary Journal of Nonlinear Science, vol. 19, no. 2, Article ID All the classifiers showed the acceptable results in terms of 026113, 2009. performance parameters such as accuracy, sensitivity, and [3] N. Giladi and Y. Balash, “+e clinical approach to gait dis- turbances in Parkinson’s disease; maintaining independent specificity. mobility,” Journal of Neural Transmission, vol. 70, pp. 327– 332, 2006. 6. Conclusion [4] M. Morris, R. Iansek, J. Mcginley, T. Matyas, and F. Huxham, “+ree-dimensional gait biomechanics in Parkinson’s disease: +e proposed study highlights the feasibility of wearable evidence for a centrally mediated amplitude regulation dis- accelerometers for gait analysis of PD patients. +e algo- order,” Movement Disorders, vol. 20, no. 1, pp. 40–50, 2004. rithmic approach used in this study is able to estimate the [5] T. R. Beijer, S. R. Lord, and M. A. D. Brodie, “Comparison of handheld video camera and GAITRite measurement of gait gait characteristics with an acceptable mean error rate. +e impairment in people with early stage Parkinson’s disease: a validation study was performed to compare the estimated pilot study,” Journal of Parkinson’s Disease, vol. 3, no. 2, values from the algorithmic approach with those obtained pp. 199–203, 2013. from the 3D motion capture system and TMT. +e proposed [6] T. G. Hampton and I. Amende, “Treadmill gait analysis approach is a low-cost approach for the detection of PD as characterizes gait alterations in Parkinson’s disease and well as able to distinguish PD subjects from the healthy older amyotrophic lateral sclerosis mouse models,” Journal of Motor group. It is also observed that the proposed classification Behavior, vol. 42, no. 1, pp. 1–4, 2009. model could able to achieve an accuracy of 88.46% with a [7] M. E. Tinetti, T. Franklin Williams, and R. Mayewski, “Fall sensitivity of 0.9286 and a specificity of 0.9091. +e objective risk index for elderly patients based on number of chronic of reducing the false negatives as much as possible could be disabilities,” 8e American Journal of Medicine, vol. 80, no. 3, achieved. +e proposed approach showed enough potential pp. 429–434, 1986. Journal of Healthcare Engineering 11 [8] J. Park, S.-B. Koh, H. J. Kim et al., “Validity and reliability [24] A. Sama, ` C. Perez-L ´ opez, ´ D. Rodr´ıguez-Mart´ın et al., “A double closed loop to enhance the quality of life of Parkinson’s study of the Korean Tinetti mobility test for Parkinson’s disease,” Journal of Movement Disorders, vol. 11, no. 1, Disease patients: REMPARK system,” Studies in Health Technology and Informatics, vol. 207, pp. 115–124, 2015. pp. 24–29, 2018. [9] W. Maetzler, J. Klucken, and M. Horne, “A clinical view on [25] S. Fahn, “Members of the UPDRS development committee. Unified Parkinson’s disease rating scale,” Recent Develop- the development of technology-based tools in managing ments in Parkinson’s Disease, vol. 2, pp. 293–304, 1987. Parkinson’s disease,” Movement Disorders, vol. 31, no. 9, [26] R. Bhidayasiri and D. Tarsy, “Parkinson’s disease: Hoehn and pp. 1263–1271, 2016. Yahr scale,” in Movement Disorders: A Video Atlas, pp. 4-5, [10] J. Taborri, E. Palermo, S. Rossi, and P. Cappa, “Gait parti- Humana Press, Totowa, NJ, USA, 2012. tioning methods: a systematic review,” Sensors, vol. 16, no. 1, [27] A. Millecamps, K. A. Lowry, J. S. Brach, S. Perera, p. 66, 2016. M. S. Redfern, and E. Sejdic, ´ “Understanding the effects of [11] W. Zhao, M. A. Reinthal, D. D. Espy, and X. Luo, “Rule-based pre-processing on extracted signal features from gait accel- human motion tracking for rehabilitation exercises: realtime erometry signals,” Computers in Biology and Medicine, vol. 62, assessment, feedback, and guidance,” IEEE Access, vol. 5, pp. 164–174, 2015. pp. 21382–21394, 2017. [28] J. H. Hollman, E. M. Mcdade, and R. C. Petersen, “Normative [12] W. Tao, T. Liu, R. Zheng, and H. Feng, “Gait analysis using spatiotemporal gait parameters in older adults,” Gait & wearable sensors,” Sensors, vol. 12, no. 2, pp. 2255–2283, 2012. Posture, vol. 34, no. 1, pp. 111–118, 2011. [13] Y. Beck, T. Herman, M. Brozgol, N. Giladi, A. Mirelman, and [29] J. C. M. Schlachetzki, J. Barth, F. Marxreiter et al., “Wearable J. M. Hausdorff, “SPARC: a new approach to quantifying gait sensors objectively measure gait parameters in Parkinson’s smoothness in patients with Parkinson’s disease,” Journal of disease,” PLoS One, vol. 12, no. 10, Article ID e0183989, 2017. Neuroengineering and Rehabilitation, vol. 15, no. 1, p. 49, 2018. [30] M. Bertoli, A. Cereatti, D. Trojaniello et al., “Estimation of [14] J. M. Hausdorff, N. Ganz, E. Gazit, A. Hadad, A. S. Buchman, spatio-temporal parameters of gait from magneto-inertial and A. Mirelman, “Automatic quantification of tandem measurement units: multicenter validation among Parkinson, walking using a wearable device: validity of the instrumented mildly cognitively impaired and healthy older adults,” Bio- tandem walk,” Innovation in Aging, vol. 3, no. Supplement_1, medical Engineering Online, vol. 17, no. 1, p. 58, 2018. p. S335, 2019. [31] W. Zijlstra and A. L. Hof, “Displacement of the pelvis during [15] E. Gazit, A. S. Buchman, R. Dawe et al., “What happens before human walking: experimental data and model predictions,” the first step? A new approach to quantifying gait initiation Gait & Posture, vol. 6, no. 3, pp. 249–262, 1997. using a wearable sensor,” Gait & Posture, vol. 76, pp. 128–135, [32] L.-Y. Hu, M.-W. Huang, S.-W. Ke, and C.-F. Tsai, “+e distance function effect on k-nearest neighbour classification [16] A. R. Anwary, H. Yu, and M. Vassallo, “Optimal foot location for medical datasets,” Springer Plus, vol. 5, no. 1, 2016. for placing wearable IMU sensors and automatic feature [33] A. Ozdemir and B. Barshan, “Detecting falls with wearable extraction for gait analysis,” IEEE Sensors Journal, vol. 18, sensors using machine learning techniques,” Sensors, vol. 14, no. 6, pp. 2555–2567, 2018. no. 6, pp. 10691–10708, 2014. [17] S. Qiu, Z. Wang, H. Zhao, L. Liu, and Y. Jiang, “Using body- [34] X.-Y. Wang, Q.-Y. Wang, H.-Y. Yang, and J. Bu, “Color image worn sensors for preliminary rehabilitation assessment in segmentation using automatic pixel classification with sup- stroke victims with gait impairment,” IEEE Access, vol. 6, port vector machine,” Neurocomputing, vol. 74, no. 18, pp. 31249–31258, 2018. pp. 3898–3911, 2011. [18] S. Byun, J. W. Han, T. H. Kim, and K. W. Kim, “Test-Retest [35] S. Akbarzadeh, A. Paap, S. Ahderom, B. Apopei, and reliability and concurrent validity of a single tri-axial accel- K. Alameh, “Plant discrimination by Support Vector Machine erometer-based gait analysis in older adults with normal classifier based on spectral reflectance,” Computers and cognition,” PLoS One, vol. 11, no. 7, Article ID e0158956, 2016. Electronics in Agriculture, vol. 148, pp. 250–258, 2018. [19] M. H. Pham, M. Elshehabi, L. Haertner et al., “Validation of a [36] B. M. Gayathri and C. P. Sumathi, “An automated technique step detection algorithm during straight walking and turning using Gaussian na¨ıve Bayes classifier to classify breast cancer,” in patients with Parkinson’s disease and older adults using an International Journal of Computer Applications, vol. 148, inertial measurement unit at the lower back,” Frontiers in no. 6, pp. 16–21, 2016. Neurology, vol. 8, 2017. [37] Y. Y. Song and L. U. Ying, “Decision tree methods: appli- [20] S. D. Din, A. Godfrey, and L. Rochester, “Validation of an cations for classification and prediction,” Shanghai Archives of accelerometer to quantify a comprehensive battery of gait Psychiatry, vol. 27, pp. 130–135, 2015. characteristics in healthy older adults and Parkinson’s disease: [38] B. Sijobert, M. Benoussaad, J. Denys, R. Pissard-Gibollet, toward clinical and at home use,” IEEE Journal of Biomedical C. Geny, and C. A. Coste, “Implementation and validation of a and Health Informatics, vol. 20, no. 3, pp. 838–847, 2016. stride length estimation algorithm, using a single basic inertial [21] S. Aich, P. Pradhan, J. Park, N. Sethi, V. Vathsa, and sensor on healthy subjects and patients suffering from Par- H.-C. Kim, “A validation study of freezing of gait (FoG) kinson’s disease,” Health, vol. 7, no. 6, pp. 704–714, 2015. detection and machine-learning-based FoG prediction using [39] D. Meldrum, C. Shouldice, R. Conroy, K. Jones, and estimated gait characteristics with a wearable accelerometer,” M. Forward, “Test-retest reliability of three dimensional gait Sensors, vol. 18, no. 10, p. 3287, 2018. analysis: including a novel approach to visualising agreement [22] V. Mikos, C.-H. Heng, A. Tay et al., “A wearable, patient- of gait cycle waveforms with Bland and Altman plots,” Gait & adaptive freezing of gait detection system for biofeedback Posture, vol. 39, no. 1, pp. 265–271, 2014. cueing in Parkinson’s disease,” IEEE Transactions on Bio- [40] Y. Okuma, “Freezing of gait in Parkinson’s disease,” Journal of medical Circuits and Systems, vol. 13, no. 3, pp. 503–515, 2019. Neurology, vol. 253, pp. 27–32, 2006. [23] H. Jeon, W. Lee, H. Park et al., “Automatic classification of tremor severity in Parkinson’s disease using a wearable de- vice,” Sensors, vol. 17, no. 9, p. 2067, 2017. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Healthcare Engineering Hindawi Publishing Corporation

Design of a Machine Learning-Assisted Wearable Accelerometer-Based Automated System for Studying the Effect of Dopaminergic Medicine on Gait Characteristics of Parkinson’s Patients

Download PDF
11 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/design-of-a-machine-learning-assisted-wearable-accelerometer-based-3egShX0Ia1

References (43)

Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2020 Satyabrata Aich 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
10.1155/2020/1823268
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Healthcare Engineering Volume 2020, Article ID 1823268, 11 pages https://doi.org/10.1155/2020/1823268 Research Article Design of a Machine Learning-Assisted Wearable Accelerometer-Based Automated System for Studying the Effect of Dopaminergic Medicine on Gait Characteristics of Parkinson’s Patients 1 2 3 Satyabrata Aich , Pyari Mohan Pradhan, Sabyasachi Chakraborty , 1 4 5 6 1 Hee-Cheol Kim , Hee-Tae Kim, Hae-Gu Lee, Il Hwan Kim, Moon-il Joo, 1 7 Sim Jong Seong, and Jinse Park Institute of Digital Anti-Aging Healthcare, Inje University, Gimhae, Republic of Korea Department of Electronics and Communication Engineering, IIT, Roorkee, India Department of Computer Engineering, Inje University, Gimhae, Republic of Korea Department of Neurology, Hanyang University Hospital, College of Medicine, Seoul, Republic of Korea Department of Industrial Design, Kyoung Sung University, Busan, Republic of Korea Department of Oncology, Haeundae Paik Hospital, Inje University, Busan, Republic of Korea Department of Neurology, Haeundae Paik Hospital, Inje University, Busan, Republic of Korea Correspondence should be addressed to Jinse Park; jinsepark@gmail.com Received 24 September 2019; Revised 15 December 2019; Accepted 8 January 2020; Published 18 February 2020 Guest Editor: Chao Chen Copyright © 2020 Satyabrata Aich 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. In the last few years, the importance of measuring gait characteristics has increased tenfold due to their direct relationship with various neurological diseases. As patients suffering from Parkinson’s disease (PD) are more prone to a movement disorder, the quantification of gait characteristics helps in personalizing the treatment. +e wearable sensors make the measurement process more convenient as well as feasible in a practical environment. However, the question remains to be answered about the validation of the wearable sensor-based measurement system in a real-world scenario. +is paper proposes a study that includes an al- gorithmic approach based on collected data from the wearable accelerometers for the estimation of the gait characteristics and its validation using the Tinetti mobility test and 3D motion capture system. It also proposes a machine learning-based approach to classify the PD patients from the healthy older group (HOG) based on the estimated gait characteristics. +e results show a good correlation between the proposed approach, the Tinetti mobility test, and the 3D motion capture system. It was found that decision tree classifiers outperformed other classifiers with a classification accuracy of 88.46%. +e obtained results showed enough evidence about the proposed approach that could be suitable for assessing PD in a home-based free-living real- time environment. features of gait in PD are short-step, hypokinetic, slow gait 1. Introduction with decreased arm swing, and episodic gait, which includes +e most important symptom of Parkinson’s disease (PD) is freezing of gait (FOG) and festinating gait [3]. Despite the the disturbances in gait that directly affects the daily ac- clinical importance, most clinicians usually depend on tivities as well as the quality of life [1]. +e disturbances in neurological examination or self-questionnaire-based ex- gait characteristics in PD patients are categorized into amination for a short period of time. +erefore, it is very continuous gait and episodic gait disturbances [2]. Typical difficult to assess the PD patient’s gait status outside the 2 Journal of Healthcare Engineering consumption in wearable devices and hence a longer clinic and in a real-world environment. Objective quanti- fication of gait is crucial for the measurement of overall battery life for gait monitoring. condition as well as disease monitoring in PD. Several (4) +e validation study provides a low-cost alternative clinical scales such as Tinetti mobility test (TMT), Timed Up for assessing gait characteristics in the “on” state of PD and Go (TUG), and Unified Parkinson’s Disease Rating patients for both indoor and outdoor environments. Scales (UPDRS) are widely used to assess the PD and its (5) +e proposed study demonstrates that spatiotem- severity. In the last decade, numerous studies have inves- poral gait characteristics estimated by using only tigated the usefulness of gait analysis. Quantitative gait accelerometer data are highly correlated with those analysis includes infrared-based motion capture (three-di- obtained from a 3D motion capture system. Fur- mensional (3D) motion capture), pressure-based gait thermore, a high correlation was also found between analysis (GAITRite), and treadmill gait analysis [4–6]. De- results obtained from the proposed approach and spite their strength of accurate quantification of gait, clinical those obtained from the clinical TMT test. implication is still controversial due to high cost and large (6) +e proposed study proposed an automatic system space or laboratory required for system set up. that can classify PD patients and HOG with machine To overcome the previous limitations, an attempt has learning techniques based on gait characteristics. been made in this study to quantify the gait characteristics using the algorithmic-based approach with a wearable ac- +e structure of the paper is outlined in the following celerometer and its validation using a 3D motion capture way: Section 2 describes the past work related to this study. system as well as TMT. TMT is widely used for predicting the Section 3 describes the data collection methods as well as the fall risk of elderly people based on the balance and gait test proposed methodology. Section 4 presents the results and score. TMT test consists of two components such as the outcomes of the proposed approach. Section 5 provides the Tinetti balance scale and the Tinetti gait scale. +e balance discussion. Section 6 describes the conclusion. scale consists of 9 parameters, and each parameter has subparameters with a score of 0/1 or 0/1/2. +e total possible 2. Related Work score of the balance section is 16. +e gait scale consists of 8 parameters, and each parameter has subparameters with a +e gait analysis performed using a conventional way using a score of 0/1 or 0/1/2. +e total possible score of the gait qualitative analysis technique is usually performed in the section is 12. Each patient has to be assessed based on these clinics, and it required a complete medical history of the two scores. +e combined score determines the risk of falls patients to determine the gait characteristics. +e conven- in elderly people. According to Tinetti, a total score of ≤18 is tional method is relatively simple; however, it depends on treated as high risk, 19–23 is treated as moderate risk, and the expertise of the physicians, and it is relatively difficult to ≥24 is treated as low risk [7]. Since the Korean version of measure the parameters in a quantitative manner with high TMT has already been validated with the PD patients in the accuracy that could be useful for clinical applications. To laboratory [8], this version has been used in this study. address this aforementioned problem, a new method has +e contributions of the proposed study are as follows: been introduced in this paper to quantify the gait charac- teristics in an objective way by using quantitative mea- (1) +is study includes enrolment of a large number of surement techniques [10, 11]. Wearable devices are now participants with PD, higher than that recommended used for a wide range of healthcare observations as well as by the movement disorder society [9]. While the the measurement of gait. +e triaxial wearable accelerometer recommended minimum number of patients is 30, is known to be a useful tool for assessing gait as well as this study involves 48 PD patients to provide proper various motor symptoms in PD. It is not expensive as well as validity and reliability of the result. In addition, 40 can be used in a comfortable way by the user [12]. Beck et al. healthy older patients’ group has been included in proposed a new approach to quantify the gait smoothness the study for the classification of PD subjects from using accelerometer and gyroscope signals. +ey have healthy older group based on estimated gait char- implemented this method in PD patients as well as healthy acteristics. Due to a large number of subjects, the controls, and they found clear differentiation in terms of proposed study could be recommended for a real-life smoothness between two groups. For validation, they have scenario. used the correlation technique by comparing their algorithm (2) +e proposed study focuses on the PD patients when spectral arc length measure (SPARC) with traditional gait they are clinically in “on” state, i.e., after taking measures and the UPDRS scale. +is is one of the potential dopaminergic medicine. “On” state is the state where use cases for using wearable sensors; however, their method the effect of the medicine is present, and the im- did not use 3D motion capture and TMT gait scale for provement in the gait characteristics is closer to the correlation [13]. Hausdorff et al. mentioned that quantifi- healthy older group. cation of gait characteristics was possible using a wearable (3) +e good accuracy found by using only accelerometer device. +ey have collected the accelerometer data during data for estimating spatiotemporal gait characteristics the tandem walking and also validated the method. +is indicates that the gyroscope data could be excluded method also mentioned the potential of using wearable for these kinds of studies. +is will lead to low power devices for gait analysis. +ey have not implemented this Journal of Healthcare Engineering 3 good level of correlation with a correlation coefficient that method to the PD patients, and at the same time, they have not used any other methods such as 3D motion capture or ranges from 0.961 to 0.984. +e study also proposed a machine learning-based approach to distinguish PD with clinical scale for correlation of their method [14]. Gazit et al. proposed a method for quantifying gait initiations using FoG from PD with no FoG, and an accuracy of 88% has been wearable sensors. +ey have used only one IMU sensor for found using SVM classifier. In this research, the effect of evaluating the gait initiations and found good results. +ey dopaminergic medicine has not been considered, and the have validated the method with the ground truth and found correlation analysis has not been performed with clinical that the interclass correlation coefficient with one wearable scale [21]. Mikos et al. proposed a method for FoG detection sensor ranges from 0.75 to 0.96. +ey have tested this using a single sensor node. +ey have developed a system using machine learning based on the extracted features from method on the data collected from younger and older adults. +ey have not used the 3D motion capture system for the signals. +ey have found a classification accuracy of 92.9% in average of sensitivity and specificity when validation of their results and also not implemented for PD patients [15]. Anwary et al. proposed a method to find the exploiting its patient adaptive learning capability. +is re- search has given enough evidence that a single sensor can be best location in the foot to place wearable sensors. +ey have used accelerometer data and gyroscopic data for deter- used for the detection of FoG and machine learning systems mining the gait features. For validation of this method, they for the classification of FoG [22]. Jeon et al. proposed a study have used a quality motion capture system. +ey have done that used the wearable device to detect the severity of tremor this analysis for healthy groups and mentioned that wearable in PD. +e wearable device used in this study consists of an sensors have the potential to quantify the gait characteristics accelerometer and gyroscope. +is study also used machine with high accuracy [16]. Qiu et al. proposed a method that learning techniques to classify the severity of tremor based on the score of UPDRS. It was found that the decision tree used body-worn sensors to collect the gait data for the as- sessment of stroke patients. +ey have found that the gait outweighs other classifiers with an accuracy of 85.5%. +is research has provided enough evidence that the wearable analysis has a huge contribution towards the diagnosis and treatment of the stroke patients and mentioned that a sensors can be used for the diagnosis of PD, and machine learning techniques can be used to automate the system [23]. wearable sensor-based gait analysis system has the potential for supporting rehabilitation in the clinics and hospitals [17]. Sama` et al. proposed a study using wearable accelerometer Byun et al. have proposed a method that uses the wearable that can detect freezing of gait at real-time environment accelerometer to measure the gait characteristics of older using a set of features which are related to the previous people having normal cognition. +e gait characteristics are approaches mentioned by the previous researchers. +ese quantified using the signal-processing algorithm. Validation features were trained using machine learning classifiers and of the measurement method was carried out using the used to detect the FoG with an improvement over the previous methods. +is research suggested that the wearable GAITRite system. +e two methods show a good level of correlation with a correlation coefficient that ranges from sensor has the potential to be used for measuring the gait characteristics, and machine learning techniques could be 0.91 to 0.96. +ey have not used the 3D motion capture system, and this method was not tested for PD patients [18]. used for the detection of the PD group [24]. Pham et al. have proposed a technique that used an inertial +e past works mentioned above provide a strong measurement system which consists of a gyroscope and an recommendation about the use of the wearable device in the accelerometer to detect the gait patterns such as toe-off and field of PD as well as the effective use of machine learning heel strike in the patients with PD as well as older adults techniques for autodetection of gait patterns in PD and when they were encountered with turning as well as straight HOG. +e proposed approach has got a lot of inspiration walking. An algorithm based on continuous wavelet from the previous pieces of literature cited by different transform is used to detect the gait patterns, and the vali- researchers. In this study, an algorithmic-based approach has been developed, and it was validated using clinical test dation study was carried out using the optoelectronic sys- tem. +ey have not used any clinical scale for comparing the and well-known measuring instruments, and a machine learning-based approach has been proposed to detect the PD result. 3D motion capture has not been used in this research [19]. Del Din et al. have used the wearable accelerometer for from the healthy older group using estimated gait charac- measuring the gait characteristics of older adults as well as teristics. +is system is developed by keeping in mind that it PD patients. Signal processing of the collected accelerometer can be used in the home environment as well as in clinical data provides gait characteristics, and the validation was environments. carried out using the instrumented walkway. Fourteen gait characteristics were compared; it was found that four 3. Proposed Methodology characteristics show a good amount of correlation, another four gait characteristics show an agreement of moderate 3.1. Data Collection. +is study was performed clinically in level, and the rest six characteristics show an agreement of a the “on” state, i.e., after taking dopaminergic medicine for poor level. +is paper does not have any correlation analysis the PD group of patients. “On” state is the state where there of gait characteristics with the clinical scale, and they did not is an effect of the medicine. In this state, there is an im- use the 3D motion capture system [20]. Aich et al. proposed provement in gait characteristics. +e resulting gait char- a method that used a wearable accelerometer that can detect acteristics are very similar to those of the healthy older FoG, and the validation study was performed that shows a group. +e accelerometer data for PD patients have been 4 Journal of Healthcare Engineering Table 1: Details of the PD group. collected in the “on” state so as to study the difference between two groups, i.e., PD patients and healthy older M/F (n � 48) 25/23 Age 70.61 ± 9.51 group when they are in a similar state. +is study was UPDRS part III 20.9 ± 12.31 performed at Haeundae Paik Hospital located at Busan, H&Y stage 2.10 ± 0.74 South Korea. +e approval was taken from the review board Disease duration (months) 35.49 ± 27.07 of the institute (IRB No. 2017-01-028). Prior approval has Timed-up and go 20.87 ± 15.78 been taken from all participants before joining this study. Tinetti gait scale 9.86 ± 2.56 +e details about the PD group are shown in Table 1. +e healthy older group comprises normal persons with no signs of PD. No medication has been given to them prior Table 2: Details of the healthy group. to this study. +e healthy group consists of 22 males and 18 M/F (n � 40) 22/18 females. All the subjects in the healthy group were age- Age 69.36 ± 7.42 matched. +e details regarding the patients belonging to the UPDRS part III 0 healthy group are shown in Table 2. Disease duration (months) 0 UPDRS and H&Y represent Unified Parkinson’s Disease Rating Scale and Hoehn and Yahr scale, respectively. UPDRS is widely used for checking the severity of the disease [25]. H&Y scale is a clinical rating scale, which is used to define different categories of motor functions in the PD [26]. Tinetti gait scale is widely used for predicting the fall risk of elderly people. +e participants were asked to wear the accelerometer on the left knee as well as the right knee. Two wearable triaxial accelerometers with a sampling frequency of 32 Hz (Fit Meter, Fit. Life, Suwon, Korea) were used. +e triaxial ac- celerometer measures body movements in all directions: anterior-posterior, mediolateral, and vertical. It is small and lightweight (35 mm × 35 mm × 13 mm and 13.7 gm). It is sensitive to acceleration from − 8 g to 8 g, allowing for Figure 1: Location of the accelerometers specified for the proposed monitoring of almost all human physical activities. All the study. participants wore the accelerometers at a distance of 34 cm from the ground, as shown in Figure 1. All the participants were asked to walk along a six-meter track. For validation of acceleration signal. +e filtered signal was integrated for gait the proposed approach, the gait characteristics were also event detection. +e objective was to detect the initial measured by using the 3D motion analysis system (VICON, contacts (ICs) of the leg, which are also termed as the heel- Oxford, UK). +e motion was captured during the walking strike event in a gait cycle. +e locations of ICs were detected process. Five important gait characteristics were measured from the points of minima in the smoothed signal by de- that include step time, stride time, step length, stride length, termining the first-order derivative using the Gaussian and walking speed. For estimating gait status more objec- continuous wavelet transform. +e flowchart of the pro- tively, the Korean version of the Tinetti gait scale [7] was posed algorithmic approach based on the accelerometer data used. +e gait characteristics obtained from the 3D motion is shown in Figure 2. system and the Tinetti gait scale were used for validation of In this study, five gait characteristics such as step time, the proposed approach. stride time, step length, stride length, and walking speed were estimated for the feasibility study of objective assessment of PD using wearable accelerometer data. +ese five charac- 3.2. Estimation of Gait Characteristics. A variant of the teristics have received great attention from the researchers in method proposed by Del Din et al. [12] was used to detect the gait-related study and its effectiveness for the assessment of gait cycle. +e measured acceleration values along X-, Y-, PD. Five major domains of gait study have been proposed by and Z-axes represent linear accelerations along the medial- Hollman et al. using the factor analysis: (1) step time and lateral (ML), anterior-posterior (AP), and vertical (V) di- stride time represented by the rhythm domain; (2) tempor- rections, respectively. +e corrections are needed to over- ophasic domain of gait cycle represented by the phase do- come the effect of gravitational component, error due to main; (3) step variability represented by the variability imprecise position of wearable accelerometer, etc. [27]. +e domain; (4) step length, stride length, and gait speed rep- dynamic tilt correction approach proposed in [27] was used resented by the pace domain; (5) step width represented by to transform the acceleration from ML and AP directions to the base of the support domain [28]. +e aforementioned five characteristics have also been used recently to detect the FoG a global horizontal-vertical coordinate system. +e resulting vertical acceleration signal was used hereafter for gait event [20]. Walking speed, stride length, and stride time have been identification. A low-pass fourth-order Butterworth filter given high importance by Schlachetzki et al. [29] for the with a cutoff frequency of 15 Hz was used to filter the vertical discrimination of healthy subjects from the PD subjects. Journal of Healthcare Engineering 5 Collection of acceleration data using triaxial wearable accelerometer at knee of left and right legs Preprocessing of raw acceleration data to eliminate the offsets and misalignments Filtering of acceleration data for the Filtering of acceleration data for the left leg to remove noise right leg to remove noise Detection of peaks (minima) Detection of peaks (minima) representing initial contact (IC) of the representing initial contact (IC) left leg using continuous wavelet of the right leg using continuous transform wavelet transform Calculation of spatiotemporal Calculation of spatiotemporal parameters for the left leg using the parameters for the right leg using above data for left leg the above data for right leg Comparison of result with  3D motion analysis system (gold standard) Data Interpretation and validation of results using statistical analysis Figure 2: Flowchart of the algorithmic approach for the estimation of gait characteristics. Bertoli et al. estimated the spatiotemporal parameters such as the inverted pendulum model [21, 31], as shown in Figure 3. stride time, step time, swing time, stance time, stride length, +e step length and stride length can be computed as and gait velocity for the quantitative assessment of PD, mild follows: 􏽱������������ cognitive impairment patients, and healthy older adults [30]. step length � K ∗ 2 2W H − H 􏼁, I h +e step time can be calculated based on the IC events (3) [9] as follows: stride length � 2 ∗ step length, step time(i) � IC(i + 1) − IC(i). (1) where W represents the distance from the ground to the wearable accelerometer and H represents the change in Similarly, the stride time can also be computed based on height of the wearable sensor between two consecutive IC the IC events [9] as follows: events. +is is computed by finding the difference between the maximum and minimum values of the double integrated stride time(i) � IC(i + 2) − IC(i), (2) vertical acceleration signal between two IC events. +e generic multiplying factor K is used for mapping the center where i denotes the index of the IC event in the signal. In the proposed approach, the step length has been estimated using of mass in an inverted pendulum model with that of the K W 1 h 6 Journal of Healthcare Engineering 3.3.2. Support Vector Machine (SVM) Classifier. SVM is one of the classifiers suitable to deal with binary classification problems. +e classifier tries to maximize the margin ar- ithmetically between two input datasets by defining a surface in an input space, which is multidimensional in nature [34]. In another way, SVM selects the hyperplane with the highest possible margin between two classes while separating them. It is impossible for a hyperplane to separate the data between two classes, but it tries to separate as much data as possible to provide good accuracy [35]. In this study, the radial basis kernel function is used, which provides good accuracy compared to other available kernel functions. 3.3.3. Na¨ıve Bayes (NB) Classifier. NB classifier is one of the simple probabilistic classifiers based on the Bayes’ theorem. +is classifier selects mutually independent variables. +is kind of classifier can be employed in the complex real-life scenario as it can be trained efficiently using the supervised IC le Step length IC right learning technique. +e advantage of this algorithm is that it Figure 3: Extended inverted pendulum model [20] for estimation needs less amount of data for training purposes to perform of step length. the classification task. In this study, the classification task has been performed by using the Bayes’ rule to calculate the probability of class label PD or a healthy group [36]. wearable sensor. +e value of K will change based on the value of W . +erefore, to avoid the time-consuming task of mapping for each participant that requires determining K 3.3.4. Decision Tree Classifier. +e decision tree classifier for each participant, W has been fixed at 34 cm, and cor- works on the basis of conditional statements and its possible respondingly, K � 4 has been chosen for this study. Walking consequences. It is a tree-like model. Nodes and branches speed is calculated as follows [21]: are the primary components to build a decision tree model. +ree steps are followed for building a well-designed de- mean step length walking speed � . (4) cision tree model. +e first step is splitting, followed by mean step time stopping, and then finally pruning. +e continuation of the splitting process stops when the model reaches the desired +ese aforementioned five estimated gait characteristics stopping criteria. +e stopping rule is used to avoid the were used as features for the classification of PD groups and problem of overfitting and underfitting. If the stopping rule healthy older group. does not work well, the pruning method is used to improve the overall classification accuracy [37]. A planned-designed PD detection framework should be 3.3. Machine Learning Classifiers and Its Effectiveness for 8is efficient and quick enough to perform the binary classifi- Study. In this study, comparative performance analysis has cation for the classification of PD patients from the healthy been carried out between four machine learning classifiers that have been employed to perform the classification task older group. Accuracy, sensitivity, and specificity are widely used to measure the effectiveness of the system. +e amount between the PD patients and the healthy control adults. of correctness required for the distinction of PD patients from the healthy older group could be measured using the 3.3.1. 8e k-Nearest Neighbour Classifier (k-NN). +e k-NN term accuracy. +e potential to identify PD is measured by classifier performs the classification process based on the sensitivity, and it is usually expressed as the ratio of true positives to the total number of PD patients [21]. +e po- proximity of a data point to the nearest training data points. It generally measures the Euclidean distance to measure the tential to identify PD when the system identifies the PD can be measured by the term specificity. +e subjects belong to closeness between them. +e local data structure has a strong influence on the k-NN algorithm. +ere is no standardized the PD group, correctly identified as PD subject, and are rule to define the value of k. +e classes are selected based on represented as true positives. +e subjects belong to a the majority rule from among the selected number of k- healthy older group, correctly identified as healthy older nearest neighbors, where k is always greater than zero and an groups, and are represented as true negatives. +e subjects integer. +e instability in the result, as well as an increase in belong to the healthy older group but wrongly identified as the variance, can be seen with the smaller values of k. +e PD subjects are represented as false positives. +e subjects belong to the PD group but wrongly identified as the healthy reduction in sensitivity, as well as increasing bias, can be seen with the higher values of k. In general, the k values are older group are represented as false negatives. In this study, the objective is to reduce the false negatives as it affects the chosen depending on the dataset. In this study, a value of k � 5 is chosen as it provides good accuracy [32, 33]. effectiveness of the system. h Journal of Healthcare Engineering 7 Table 3: Mean value of gait characteristics and average error rate for the left and right legs. Sl. no. Parameters Mean value (3D motion capture) Mean value (algorithm) Mean error rate (%) Left leg 1 Step time (s) 0.57 0.54 6.94 ± 2.82 2 Stride time (s) 1.17 1.13 4.76 ± 3.55 3 Step length (m) 0.37 0.34 6.35 ± 2.85 4 Stride length (m) 0.74 0.71 6.51 ± 2.92 5 Walking speed (m/s) 0.64 0.61 7.12 ± 2.74 Right leg 1 Step time (s) 0.54 0.56 7.14 ± 2.52 2 Stride time (s) 1.18 1.14 5.25 ± 3.62 3 Step length (m) 0.37 0.34 6.15 ± 2.81 4 Stride length (m) 0.74 0.70 6.35 ± 2.71 5 Walking speed (m/s) 0.69 0.66 6.72 ± 3.14 plots between the algorithmic approach as well as a 3D 4. Results motion capture system are shown in Figures 4–8. +e mean +e mean value of five estimated gait characteristics based error rate was calculated based on the formula [21] as on the accelerometer data as well as the mean error rate follows: between the algorithmic approach and the 3D motion capture system are highlighted in Table 3. +e correlation (value estimated from acc) − (value estimated from 3D capture) average error rate(%) � ∗ 100. (5) (value estimated from 3D capture) +e total number of subjects including both the groups is 88. In this paper, Tinetti mobility test (TMT) gait scale is used to assess the spatiotemporal gait characteristics such Out of 88 subjects, 62 subjects belong to the training group, as step time, stride time, step length, stride length, and and the rest 26 belong to the validation group. Out of 26 walking speed and its importance in terms of clinical subjects, which belong to the validation group, 14 subjects practices by comparing the score with the result obtained belong to the PD group (PDG) and 12 subjects belong to the using other methods, in this case, computerized gait healthy older group (HOG). Moreover, a 5 split cross-vali- analysis using accelerometer data and 3D motion capture dation was also performed based on the subject’s data to check system. +e true changes in the gait characteristics can be the generalizability of the model. +e cross-validation was easily understand based on the accuracy of the clinical performed in such a way where the data of 62 random subjects observation measures, and it is an important step in clinical were used to train a classifier and the rest data of 26 subjects practices. So, in this paper, we have used Pearson’s cor- were used for checking the testing accuracy. Test set 1, test set 2, test set 3, and test set 4 consist of 26 subjects each. +e relation coefficient to analyze the relationship between the TMT gait scale score, and spatiotemporal gait character- cross-validation was performed using 4 different classifiers, istics derived objectively used computerized gait analysis namely, KNN, SVM, Naive Bayes, and decision tree. +e using accelerometer data and 3D motion capture system. implementation of four different algorithms was done to +e correlation plots between TMT gait scale and gait perform a comparative analysis between the classifiers. After characteristics measured from the 3D motion capture successful cross-validation, it was found that the decision tree system are shown in Figures 4–8. We have found strong plotted the best set of results by prompting a maximum correlations between them, and the results were mentioned accuracy of 88.46%, sensitivity of 92.86%, and specificity of as follows: step time (0.96, p < 0.01), stride time (0.97, 90.91%, respectively. Table 4 shows the results for the cross- p < 0.01), step length (0.98, p < 0.01), stride length (0.99, validation. p < 0.01), and walking speed (0.99, p < 0.01). Similarly, the +e classifiers’ performance has been evaluated using three parameters such as accuracy, sensitivity, and speci- correlation plots between TMT gait scale and gait char- acteristics obtained from the wearable accelerometer data ficity. +e classification results are shown in Table 5. +e are shown in Figures 9–13. We have found moderate to decision tree classifier could able to provide the highest strong correlations between them, and the results were accuracy of 88.46% with a sensitivity of 0.9286 and specificity mentioned as follows: step time (0.57, p < 0.01), stride time of 0.9091. From 14 subjects belonging to PDG, the proposed (0.54, p < 0.01), step length (0.84, p < 0.01), stride length model correctly identified 13 as PDG. Similarly, from the 12 (0.84, p < 0.01), and walking speed (0.75, p < 0.01). subjects belonging to HOG, the proposed model correctly +is study used the split named as stratified train-vali- identified 10 as HOG. +e confusion matrix is shown in dation [21] with a ratio of 70 : 30 for training and validation. Figure 14. 8 Journal of Healthcare Engineering 1 1.4 r = 0.96 r = 0.99 1.2 0.9 0.8 0.8 0.7 0.6 0.6 0.4 0.5 0.2 0.4 0 0.2 0.4 0.6 0.8 1 1.2 Stride length, accelerometer 0.3 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Figure 7: Stride length correlation plot between the accelerometer- Step time, accelerometer based approach and 3D motion system ( p < 0.01). Figure 4: Step time correlation plot between the accelerometer- based approach and 3D motion system ( p < 0.01). 1.4 r = 0.99 1.2 r = 0.97 1.8 1.6 0.8 0.6 1.4 0.4 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Walking speed, accelerometer 0.8 Figure 8: Walking speed correlation plot between the acceler- 0.6 ometer-based approach and 3D motion system ( p < 0.01). 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 Stride time, accelerometer Figure 5: Stride time correlation plot between the accelerometer- 0.9 ∗ ∗ r = 0.57 based approach and 3D motion system ( p < 0.01). 0.8 0.7 r = 0.98 1.4 1.2 0.6 0.5 0.8 0.4 0.6 0.3 0.4 02468 10 12 0.2 Tinetti gait scale Figure 9: Step time correlation plot between the accelerometer- 0 0.2 0.4 0.6 0.8 1 1.2 ∗ based approach and Tinetti gait scale ( p < 0.01). Step length, accelerometer Figure 6: Step length correlation plot between the accelerometer- learning approach can be used for automatic detection and based approach and 3D motion system ( p < 0.01). differentiation of PD patients from the healthy older group. Although wearable sensors have been widely used in many fields, these have not been given enough importance in 5. Discussion PD-related assessment due to distorted gait pattern. Sijobert +is study proposes an algorithmic approach to estimate the et al. [38] have proposed a technique that provides a mean gait characteristics of PD subjects as well as the healthy older error rate of 10.3% for the PD group and 6% for the healthy groups. +e approach is validated using measuring instru- group. +e comparison has been made based on the esti- ments and clinical scale. It is also proposed that the machine mated stride length calculated using wearable sensor data Step length, 3D motion capture Stride time, 3D motion capture Step time, 3D motion capture Step time Walking speed, 3D motion capture Stride length, 3D motion capture Journal of Healthcare Engineering 9 1.4 ∗ 1.4 r = 0.54 1.3 r = 0.75 1.2 1.2 1.1 1 0.8 0.9 0.8 0.6 0.7 0.4 0.6 0.2 0.5 0 2 4 6 8 10 12 14 Tinetti gait scale 02468 10 12 14 Figure 10: Stride time correlation plot between the accelerometer- Tinetti gait scale based approach and Tinetti gait scale ( p < 0.01). Figure 13: Walking speed correlation plot between the acceler- ometer-based approach and Tinetti gait scale ( p < 0.01). Table 4: 5 split cross-validation. Performance (%) KNN SVM NB Decision tree Accuracy test set 1 82.11 81.36 84.52 86.28 0.7 r = 0.84 Sensitivity test set 1 0.8746 0.7801 0.8225 0.9152 0.6 Specificity test set 1 0.8452 0.725 0.8654 0.8833 Accuracy test set 2 83.64 84.25 81.20 84.31 0.5 Sensitivity test set 2 0.8055 0.8139 0.8558 0.8631 0.4 Specificity test set 2 0.8519 0.8687 0.8411 0.8551 Accuracy test set 3 86.32 84.93 85.31 82.28 0.3 Sensitivity test set 3 0.9025 0.8755 0.9032 0.8111 Specificity test set 3 0.8947 0.9054 0.8748 0.8364 0.2 Accuracy test set 4 85.57 87.23 84.41 88.46 0.1 Sensitivity test set 4 0.9125 0.9189 0.8956 0.9286 Specificity test set 4 0.8836 0.8997 0.8735 0.9091 02468 10 12 14 Accuracy test set 5 87.26 84.39 79.32 87.32 Tinetti gait scale Sensitivity test set 5 0.8568 0.8793 0.8178 0.9025 Specificity test set 5 0.9034 0.8998 0.8227 0.9131 Figure 11: Step length correlation plot between the accelerometer- based approach and Tinetti gait scale ( p < 0.01). Table 5: Classification results. Performance k-NN SVM NB Decision tree Accuracy (%) 85.57 87.23 84.41 88.46 Sensitivity (%) 0.9125 0.9189 0.8956 0.9286 Specificity (%) 0.8836 0.8997 0.8735 0.9091 1.4 r = 0.84 1.2 PDG HOG 0.8 0.6 13 1 PDG 0.4 0.2 02468 10 12 14 210 HOG Tinetti gait scale Figure 12: Stride length correlation plot between the accelerom- eter-based approach and Tinetti gait scale ( p < 0.01). Predicted labels Figure 14: Confusion matrix. Step length Stride length Stride time Walking speed True labels 10 Journal of Healthcare Engineering and further validated using the GAITRite-based walkway sys- to get recommended for the clinicians to use in the labo- tem. +e estimated mean error rate for the five gait charac- ratory as well as in the home environment. teristics is found to be less than 8% with our proposed approach, In the future, we will collect gait data from a large which used the wearable accelerometer to collect the data. +e number of PD patients to summarize the gait characteristics results of our proposed approach provide the feasibility of our in a better way so that it could be promoted for clinical ap- approach when compared with the previous study. +e pro- plications. We would also like to combine brain EEG signals posed study provides some new ideas that are as follows: with the gait data to understand more about the relation and detect the symptoms like freezing of gait before it happens. We (i) +e results obtained in this study include various would like to combine the MRI image with the gait data for phenotypes and severity of PD due to the large more accurate diagnosis and early detection of PD. sample size. (ii) +is study uses the Tinetti gait scale [7] and the 3D Data Availability motion capture system [39] for validation of gait status. +e previous report [38] has demonstrated +e data used to support the experiments and the finding of the validation using GAITRite that can assess the study have been duly included in Section 3 of the paper. spatiotemporal data by using pressure parameters. Section 3.1 clearly describes about the data. (iii) +e gait characteristics estimated using our pro- posed approach have been compared with the Conflicts of Interest clinical scale, and the result shows a good level of agreement, which makes the method feasible to be +e authors declare that they have no conflicts of interest. implemented in the real-life environment. (iii) It is observed from the study that only acceler- Acknowledgments ometer data can provide enough information for performing the gait analysis, which leads to re- +is research was supported by the National Research dundancy of gyroscopic data, which indirectly saves Foundation (NRF) of Korea grant funded by the Korea battery power, time, and cost. government (MSIT) (Grant number 2019R1C1C1011197) and also funded by Ministry of Trade, Industry and Energy +e strength and possibility of the wearable sensor-based (MOTIE), Korea, through the Education program for Cre- PD assessment are aimed at long-term monitoring of gait. ative and Industrial Convergence (Grant number N0000717). Gait disturbance usually gets aggravated in specific cases such as starting time, meeting narrow space, or obstacle [40]. +e hospital has limited space, and therefore, it is difficult to References replicate gait disturbance observed in a real-world scenario. [1] C. C. Walton, J. M. Shine, J. M. Hall et al., “+e major impact +e battery mounted in our device lasts up to 72 hours. +e of freezing of gait on quality of life in Parkinson’s disease,” limitation of this study is that the patients in the relatively Journal of Neurology, vol. 262, no. 1, pp. 108–115, 2014. early stage of PD were not enrolled, and therefore, this study [2] J. M. Hausdorff, “Gait dynamics in Parkinson’s disease: is not intended at assessing various balance status in PD such common and distinct behavior among stride length, gait as a freeze of gait (FOG) or hypokinetic and short-step gait. variability, and fractal-like scaling,” Chaos: An Interdisci- +e classifiers used in this study have shown good accuracy. plinary Journal of Nonlinear Science, vol. 19, no. 2, Article ID All the classifiers showed the acceptable results in terms of 026113, 2009. performance parameters such as accuracy, sensitivity, and [3] N. Giladi and Y. Balash, “+e clinical approach to gait dis- turbances in Parkinson’s disease; maintaining independent specificity. mobility,” Journal of Neural Transmission, vol. 70, pp. 327– 332, 2006. 6. Conclusion [4] M. Morris, R. Iansek, J. Mcginley, T. Matyas, and F. Huxham, “+ree-dimensional gait biomechanics in Parkinson’s disease: +e proposed study highlights the feasibility of wearable evidence for a centrally mediated amplitude regulation dis- accelerometers for gait analysis of PD patients. +e algo- order,” Movement Disorders, vol. 20, no. 1, pp. 40–50, 2004. rithmic approach used in this study is able to estimate the [5] T. R. Beijer, S. R. Lord, and M. A. D. Brodie, “Comparison of handheld video camera and GAITRite measurement of gait gait characteristics with an acceptable mean error rate. +e impairment in people with early stage Parkinson’s disease: a validation study was performed to compare the estimated pilot study,” Journal of Parkinson’s Disease, vol. 3, no. 2, values from the algorithmic approach with those obtained pp. 199–203, 2013. from the 3D motion capture system and TMT. +e proposed [6] T. G. Hampton and I. Amende, “Treadmill gait analysis approach is a low-cost approach for the detection of PD as characterizes gait alterations in Parkinson’s disease and well as able to distinguish PD subjects from the healthy older amyotrophic lateral sclerosis mouse models,” Journal of Motor group. It is also observed that the proposed classification Behavior, vol. 42, no. 1, pp. 1–4, 2009. model could able to achieve an accuracy of 88.46% with a [7] M. E. Tinetti, T. Franklin Williams, and R. Mayewski, “Fall sensitivity of 0.9286 and a specificity of 0.9091. +e objective risk index for elderly patients based on number of chronic of reducing the false negatives as much as possible could be disabilities,” 8e American Journal of Medicine, vol. 80, no. 3, achieved. +e proposed approach showed enough potential pp. 429–434, 1986. Journal of Healthcare Engineering 11 [8] J. Park, S.-B. Koh, H. J. Kim et al., “Validity and reliability [24] A. Sama, ` C. Perez-L ´ opez, ´ D. Rodr´ıguez-Mart´ın et al., “A double closed loop to enhance the quality of life of Parkinson’s study of the Korean Tinetti mobility test for Parkinson’s disease,” Journal of Movement Disorders, vol. 11, no. 1, Disease patients: REMPARK system,” Studies in Health Technology and Informatics, vol. 207, pp. 115–124, 2015. pp. 24–29, 2018. [9] W. Maetzler, J. Klucken, and M. Horne, “A clinical view on [25] S. Fahn, “Members of the UPDRS development committee. Unified Parkinson’s disease rating scale,” Recent Develop- the development of technology-based tools in managing ments in Parkinson’s Disease, vol. 2, pp. 293–304, 1987. Parkinson’s disease,” Movement Disorders, vol. 31, no. 9, [26] R. Bhidayasiri and D. Tarsy, “Parkinson’s disease: Hoehn and pp. 1263–1271, 2016. Yahr scale,” in Movement Disorders: A Video Atlas, pp. 4-5, [10] J. Taborri, E. Palermo, S. Rossi, and P. Cappa, “Gait parti- Humana Press, Totowa, NJ, USA, 2012. tioning methods: a systematic review,” Sensors, vol. 16, no. 1, [27] A. Millecamps, K. A. Lowry, J. S. Brach, S. Perera, p. 66, 2016. M. S. Redfern, and E. Sejdic, ´ “Understanding the effects of [11] W. Zhao, M. A. Reinthal, D. D. Espy, and X. Luo, “Rule-based pre-processing on extracted signal features from gait accel- human motion tracking for rehabilitation exercises: realtime erometry signals,” Computers in Biology and Medicine, vol. 62, assessment, feedback, and guidance,” IEEE Access, vol. 5, pp. 164–174, 2015. pp. 21382–21394, 2017. [28] J. H. Hollman, E. M. Mcdade, and R. C. Petersen, “Normative [12] W. Tao, T. Liu, R. Zheng, and H. Feng, “Gait analysis using spatiotemporal gait parameters in older adults,” Gait & wearable sensors,” Sensors, vol. 12, no. 2, pp. 2255–2283, 2012. Posture, vol. 34, no. 1, pp. 111–118, 2011. [13] Y. Beck, T. Herman, M. Brozgol, N. Giladi, A. Mirelman, and [29] J. C. M. Schlachetzki, J. Barth, F. Marxreiter et al., “Wearable J. M. Hausdorff, “SPARC: a new approach to quantifying gait sensors objectively measure gait parameters in Parkinson’s smoothness in patients with Parkinson’s disease,” Journal of disease,” PLoS One, vol. 12, no. 10, Article ID e0183989, 2017. Neuroengineering and Rehabilitation, vol. 15, no. 1, p. 49, 2018. [30] M. Bertoli, A. Cereatti, D. Trojaniello et al., “Estimation of [14] J. M. Hausdorff, N. Ganz, E. Gazit, A. Hadad, A. S. Buchman, spatio-temporal parameters of gait from magneto-inertial and A. Mirelman, “Automatic quantification of tandem measurement units: multicenter validation among Parkinson, walking using a wearable device: validity of the instrumented mildly cognitively impaired and healthy older adults,” Bio- tandem walk,” Innovation in Aging, vol. 3, no. Supplement_1, medical Engineering Online, vol. 17, no. 1, p. 58, 2018. p. S335, 2019. [31] W. Zijlstra and A. L. Hof, “Displacement of the pelvis during [15] E. Gazit, A. S. Buchman, R. Dawe et al., “What happens before human walking: experimental data and model predictions,” the first step? A new approach to quantifying gait initiation Gait & Posture, vol. 6, no. 3, pp. 249–262, 1997. using a wearable sensor,” Gait & Posture, vol. 76, pp. 128–135, [32] L.-Y. Hu, M.-W. Huang, S.-W. Ke, and C.-F. Tsai, “+e distance function effect on k-nearest neighbour classification [16] A. R. Anwary, H. Yu, and M. Vassallo, “Optimal foot location for medical datasets,” Springer Plus, vol. 5, no. 1, 2016. for placing wearable IMU sensors and automatic feature [33] A. Ozdemir and B. Barshan, “Detecting falls with wearable extraction for gait analysis,” IEEE Sensors Journal, vol. 18, sensors using machine learning techniques,” Sensors, vol. 14, no. 6, pp. 2555–2567, 2018. no. 6, pp. 10691–10708, 2014. [17] S. Qiu, Z. Wang, H. Zhao, L. Liu, and Y. Jiang, “Using body- [34] X.-Y. Wang, Q.-Y. Wang, H.-Y. Yang, and J. Bu, “Color image worn sensors for preliminary rehabilitation assessment in segmentation using automatic pixel classification with sup- stroke victims with gait impairment,” IEEE Access, vol. 6, port vector machine,” Neurocomputing, vol. 74, no. 18, pp. 31249–31258, 2018. pp. 3898–3911, 2011. [18] S. Byun, J. W. Han, T. H. Kim, and K. W. Kim, “Test-Retest [35] S. Akbarzadeh, A. Paap, S. Ahderom, B. Apopei, and reliability and concurrent validity of a single tri-axial accel- K. Alameh, “Plant discrimination by Support Vector Machine erometer-based gait analysis in older adults with normal classifier based on spectral reflectance,” Computers and cognition,” PLoS One, vol. 11, no. 7, Article ID e0158956, 2016. Electronics in Agriculture, vol. 148, pp. 250–258, 2018. [19] M. H. Pham, M. Elshehabi, L. Haertner et al., “Validation of a [36] B. M. Gayathri and C. P. Sumathi, “An automated technique step detection algorithm during straight walking and turning using Gaussian na¨ıve Bayes classifier to classify breast cancer,” in patients with Parkinson’s disease and older adults using an International Journal of Computer Applications, vol. 148, inertial measurement unit at the lower back,” Frontiers in no. 6, pp. 16–21, 2016. Neurology, vol. 8, 2017. [37] Y. Y. Song and L. U. Ying, “Decision tree methods: appli- [20] S. D. Din, A. Godfrey, and L. Rochester, “Validation of an cations for classification and prediction,” Shanghai Archives of accelerometer to quantify a comprehensive battery of gait Psychiatry, vol. 27, pp. 130–135, 2015. characteristics in healthy older adults and Parkinson’s disease: [38] B. Sijobert, M. Benoussaad, J. Denys, R. Pissard-Gibollet, toward clinical and at home use,” IEEE Journal of Biomedical C. Geny, and C. A. Coste, “Implementation and validation of a and Health Informatics, vol. 20, no. 3, pp. 838–847, 2016. stride length estimation algorithm, using a single basic inertial [21] S. Aich, P. Pradhan, J. Park, N. Sethi, V. Vathsa, and sensor on healthy subjects and patients suffering from Par- H.-C. Kim, “A validation study of freezing of gait (FoG) kinson’s disease,” Health, vol. 7, no. 6, pp. 704–714, 2015. detection and machine-learning-based FoG prediction using [39] D. Meldrum, C. Shouldice, R. Conroy, K. Jones, and estimated gait characteristics with a wearable accelerometer,” M. Forward, “Test-retest reliability of three dimensional gait Sensors, vol. 18, no. 10, p. 3287, 2018. analysis: including a novel approach to visualising agreement [22] V. Mikos, C.-H. Heng, A. Tay et al., “A wearable, patient- of gait cycle waveforms with Bland and Altman plots,” Gait & adaptive freezing of gait detection system for biofeedback Posture, vol. 39, no. 1, pp. 265–271, 2014. cueing in Parkinson’s disease,” IEEE Transactions on Bio- [40] Y. Okuma, “Freezing of gait in Parkinson’s disease,” Journal of medical Circuits and Systems, vol. 13, no. 3, pp. 503–515, 2019. Neurology, vol. 253, pp. 27–32, 2006. [23] H. Jeon, W. Lee, H. Park et al., “Automatic classification of tremor severity in Parkinson’s disease using a wearable de- vice,” Sensors, vol. 17, no. 9, p. 2067, 2017.

Journal

Journal of Healthcare Engineering – Hindawi Publishing Corporation

Published: Feb 18, 2020

There are no references for this article.