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

Learn More →

Sleep Staging Using Noncontact-Measured Vital Signs

Sleep Staging Using Noncontact-Measured Vital Signs Hindawi Journal of Healthcare Engineering Volume 2022, Article ID 2016598, 11 pages https://doi.org/10.1155/2022/2016598 Research Article 1 1 2,3 4 4 Zixia Wang , Shuai Zha , Baoxian Yu , Pengbin Chen , Zhiqiang Pang , 2,3 and Han Zhang Department of Physics and Telecommunication Engineering, South China Normal University, Guangzhou 510006, China Department of Electronic and Information Engineering, South China Normal University, Foshan 528000, China Guangdong Provincial Engineering Technology Research Center of Cardiovascular Individual Medicine & Big Data, South China Normal University, Guangzhou 510006, China Guangzhou SENVIV Technology Co.,Ltd., Guangzhou 510006, China Correspondence should be addressed to Baoxian Yu; yubx@m.scnu.edu.cn and Han Zhang; zhanghan@scnu.edu.cn Received 12 January 2022; Revised 20 May 2022; Accepted 13 June 2022; Published 8 July 2022 Academic Editor: Haihong Zhang Copyright © 2022 Zixia Wang 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. As a physiological phenomenon, sleep takes up approximately 30% of human life and signi†cantly a‡ects people’s quality of life. To assess the quality of night sleep, polysomnography (PSG) has been recognized as the gold standard for sleep staging. �e drawbacks of such a clinical device, however, are obvious, since PSG limits the patient’s mobility during the night, which is inconvenient for in-home monitoring. In this paper, a noncontact vital signs monitoring system using the piezoelectric sensors is deployed. Using the so-designed noncontact sensing system, heartbeat interval (HI), respiratory interval (RI), and body movements (BM) are separated and recorded, from which a new dimension of vital signs, referred to as the coordination of heartbeat interval and respiratory interval (CHR), is obtained. By extracting both the independent features of HI, RI, and BM and the coordinated features of CHR in di‡erent timescales, Wake-REM-NREM sleep staging is performed, and a postprocessing of staging fusion algorithm is proposed to re†ne the accuracy of classi†cation. A total of 17 all-night recordings of noncontact measurement simultaneous with PSG from 10 healthy subjects were examined, and the leave-one-out cross-validation was adopted to assess the performance of Wake-REM-NREM sleep staging. Taking the gold standard of PSG as reference, numerical results show that the proposed sleep staging achieves an averaged accuracy and Cohen’s Kappa index of 82.42% and 0.63, respectively, and performs robust to subjects su‡ering from sleep-disordered breathing. into N1, N2, and N3 stages [4]. REM and NREM alternate 1. Introduction in cycles of about 90 min [5]. NREM, especially deep sleep Sleep plays an important role in body recovery, memory, (N3), is more prominent during the †rst hours of sleep and immunity enhancement, which takes almost one- and is essential toward physical recovery [6, 7]. REM link third of human life [1]. Poor sleep quality usually results to dreaming, more prominently during the last hours of in physical and mental health problems, such as fatigue, sleep, acts toward the recovery of people’s mental state [8]. anxiety, and even death [2]. It has been reported that sleep �erefore, a †ne-grained classi†cation of Wake-REM- duration is closely related to mortality [3]. �erefore, NREM stages is crucial for assessing sleep quality. long-term sleep quality monitoring is of great signi†cance Polysomnography (PSG) is the gold standard in the to protect human health. clinic for evaluating sleep quality and diagnosing sleep- Sleep staging is an important characteristic to qualify related diseases. Physiological signal acquisition during sleep quality. According to the American Academy of sleep is usually conducted in the hospital since the ac- Sleep Medicine (AASM), a complete sleep cycle consists of quisition process of vital signs is cumbersome. To obtain wake, rapid eye movement (REM), and non-REM the physiology signals including electroencephalography (NREM) stages, where the last one can be further divided (EEG), electrooculography (EOG), electromyography 2 Journal of Healthcare Engineering features between heartbeat interval and respiratory in- (EMG), electrocardiography (ECG), and respiratory rate [9], a large number of sensors are attached to the skin of terval in different timescales has not been reported yet. In order to address the above issues, this paper focuses patients directly [10], which is impractical for long-term monitoring at home. on a noncontact sensing-based sleep staging for Wake/ Many studies have attempted to classify sleep stages REM/NREM classification using vital signs recorded by automatically under more natural sleep monitoring piezoelectric sensors. Specially, the contributions of this conditions by using a limited number of sensing channels. paper, in comparison with the existing studied, can be In [11], wrist-actigraphy was used to record body summarized as follows. movements during sleep and achieved a sleep/wake (1) A noncontact vital signs monitoring system using classification performance of 77.8%. In [12], the authors piezoelectric sensors is deployed, by which both the took infants as experimental subjects and used wrist- independent features of heartbeat interval (HI), re- actigraphy to achieve a sleep/wake classification. How- spiratory interval (RI), body movements (BM), and ever, this method has limitations when the subject likes to the coordinated features between HI and RI (a.k.a. be quiet in wake period. Motivated by the fact that the CHR) are extracted from the noncontact-measured autonomic nerves (assessed by heartbeat and respiratory vital signs and employed for Wake/REM/NREM rate) changes in different sleep stages, respiratory in- stage classification. In addition, the effects of features ductive plethysmography signals have been used in three- in different timescales on sleep staging are also stage classification (i.e., REM/NREM/Wake) and achieved analyzed. the classification performance of 80.38% [13]. In [14], the (2) A postprocessing with respect to the fusion of the authors also considered single-lead ECG heartbeat in- classified stages is developed to further improve the terval detection and yielded the sleep/wake classification performance of sleep staging. For validations, 17 all- performance of 76%. In order to improve the classification night experiments were examined simultaneously performance, cardiorespiratory features including both with PSG. Numerical comparisons demonstrate that the features extracted from heartbeat interval and respi- the proposed sleep staging achieves average accuracy ratory are extensively studied [15–19]. Specifically, using and Kappa index of 82.42% and 0.63, respectively, ECG sensors, features of heart rate variability in time and and performs robust to subjects suffering from sleep- frequency domain are fused with the time and effort disordered breathing. (fluctuation) features of respiratory signal to aid sleep staging, and the performance of REM/NREM/Wake stages classification was improved to almost 80% [16]. In 2. Materials and Methods [20, 21], ECG sensors were used to extract features with respect to heart rate, respiratory rate, and body move- 2.1. System Setup and Vital Signs Acquisition. For non- ment. Although the existing single-lead ECG sensor based contact vital signs acquisition, the noninvasive heart rate methods can achieve satisfactory performance of sleep and respiratory rate sensing device developed by staging, the acquisition of signals requires physical con- Guangzhou SENVNV Co. was deployed [29], where the tact with subjects’ skin, which is inappropriate to long- sleep-related physiological signals, including BCG, re- term home monitoring. spiratory signal, and artifact motion can be recorded In order to solve the discomfort between sensors and during night sleep in a noncontact manner. After pre- skin in the process of signal acquisition, many studies processing, heartbeat interval, respiratory interval, and have focused on noncontact monitoring technologies body movements are separated, and sleep stage-related [22]. Using piezoelectric sensors, heart rate was obtained features are extracted from the above physiological signals based on noncontact-measured ballistocardiogram (BCG) and then fed into machine learning classifiers for sleep signals, by which three-stage classification of night sleep is stage classification. Finally, an empirical rule-based performed [23]. In [24], the authors considered PVDF postprocessing is applied to improve the classification sensors to record night-sleep physiological signals, in- performance. Figure 1 depicts the framework of the cluding BCG, respiratory signal, and body movements, noncontact sleep staging system. and then employed long short-term memory (LSTM) *e noncontact vital signs monitoring system neural network [25] to perform end-to-end sleep stage deployed in this study consists of a piezoelectric sensor, classifications. By analogy, in [26, 27], continuous-wave circuit, and processing modules. As shown in Figure 2, the Doppler radar sensing technology was adopted to dis- piezoelectric sensor module is placed under the pillow to tinguish Wake/REM/LightSleep/DeepSleep states and perceive the vibration induced by heartbeat, respiratory achieved an accuracy of 81% and 66.7% in comparison rate, and body movements. *e converted voltage signals with PSG standard. Taking radar sensing as the vital sign are then amplified, sampled at 1 kHz in the circuit module. monitoring system, the authors in [28] evaluated the A 12-bit analog-to-digital conversion (ADC) is conducted performance of Wake/REM/NREM sleep staging, where before transmission of vital signs from circuit to pro- the overall accuracy reached up to 88.4%. Although sleep- cessing module for offline signal processing and sleep related vital signs (i.e., heart rate and respiratory rate) and stage classification. For reference, we simultaneously body movements can be measured by using the existing collect EEG, ECG, EOG, and EMG from PSG (Medcare) as noncontact sensing devices, the effect of coordinated the gold standard to evaluate the staging performance. Journal of Healthcare Engineering 3 Pre-process of Feature extraction vital sings Sleep stage HI BCG Non-contact vital Sleep stage classification Wake sings acquisition RI Classification Stage fusion NREM Respiratory Vital signs CHR REM Body BM movement Figure 1: �e overall architecture of the noncontact sleep monitoring system. (a) (b) Figure 2: �e overview of the vital signs monitoring system. (a) �e subject simultaneously monitored by both PSG and the noncontact vital signs system. (b) �e device of the vital signs monitoring system. 2.2. Noncontact Vital Signs Acquisition. �is study involved Table 1: Information of the datasets. a total of 10 healthy subjects (7 males and 3 females), who Subject Weight Height Total are college students, aged from 21 to 25 years old. All Gender Age Valid (night) ID (kg) (m) (night) subjects are healthy and have a regular night sleep with an N1 M 23 65.23 1.70 5 4 average duration of 6 to 9 hours. In addition, subjects with N2 M 22 66.52 1.72 4 3 sleep disorder, smoking habits, intake of medicine, or N3 F 21 48.36 1.60 1 0 drinks infecting sleep will not be included. As shown in N4 M 23 70.44 1.74 4 4 Table 1, a total of 24 night-sleep experiments are carried N5 F 24 52.62 1.64 1 0 out, in which the recorded data in 7 night sleep are invalid N6 M 23 69.87 1.73 2 2 due to the collapse of the electrodes of PSG. N7 M 22 62.15 1.67 3 2 N8 F 25 68.53 1.76 1 0 N9 M 22 60.76 1.66 1 1 2.3. Preprocessing of Vital Signs. Since the vital signs N10 M 23 67.39 1.75 2 1 measured by the noncontact device are mixed with BCG, respiratory signal, body movements, and noise, pre- processing of vital signs is †rstly performed to separate Next, we extract BCG from the resulting signals of BM  0. di‡erent types of signals for heartbeat interval and re- Considering that the spectrum of a typical BCG ranges from spiratory interval detection. �e procedures are shown as 3 Hz to 10 Hz [31], we employ a 2nd-order Butterworth follows. bandpass †lter ranges 2.5–10 (Hz) to remove the undesired First, we remove the power frequency noise by using a signal interference and then detect heartbeat interval by using a band-stop †lter with lower and upper cuto‡ frequencies of forward and backward approach [29]. 49 Hz and 51 Hz, respectively. �en, we identify body move- For respiratory interval detection, we remove the detected ments (BM) from the processed signal. �e reason is that, on BCG signal from the recorded vital signs and then separate the respiratory waves directly by using discrete wavelet transform the one hand, body movements can signi†cantly a‡ect the detection and analysis of vital signs (i.e., BCG and respiratory and Sym8 wavelet package at 8 scales [32]. Finally, respiratory signals). On the other hand, as will be illustrated later, body interval can be obtained by using peak detection [33]. movements signal is an important feature for sleep staging. For validation, a typical example of 20-minute comparison Unlike [30], body movements (BM) are detected in multi-time- between the noncontact-measured HI and RI and that obtained scale procedure. �e detailed procedures are as follows. If the using BIOPAC MP160, which are recognized as the gold peak-to-valley di‡erence (PVD) within a 2 s window is 2.2 standard in related works. As can be observed from Figure 3, times greater than any one of the medians of PVD within a although some errors occur in very limited areas due to the multi-time-scale epochs of 30 s, 60 s, 120 s, and 300 s, the state interference of body motion, the noncontact-measured HI and RI are almost indistinguishable from those obtained using ECG of 2 s window will be considered as body movement; that is, BM  1; otherwise, BM  0. and belt sensors. �e results demonstrate that the measured HI 4 Journal of Healthcare Engineering 0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200 time (s) time (s) RI of piezoelectric sensors HI of piezoelectric sensors HI of PSG RI of PSG (a) (b) Figure 3: An example of comparison between detected HI and RI in a noncontact manner and that obtained by using the gold standard device. and RI in a noncontact manner can be further used as features since the rhythm of heartbeat and respiratory interval vary in for sleep stage classi†cation. di‡erent sleep stages. Speci†cally, features {1, 2} and {10, 11} are the mean and coe¯cient variation of independent HI and RI in 3. Feature Extraction di‡erent timescales in a 60 s epoch. Features {3–7} and {12–16} further describe the trend of °uctuation with Using the non-contact-measured vital signs (i.e., HI, RI, respect to heartbeat interval and respiratory interval and BM signals), we propose feature extraction based on over a 60 s epoch by evaluating the ratio of di‡erence in the above independent and coupled vital signs for sleep terms of heartbeat interval and respiratory interval stage classi†cation. Considering that the relevant features percentiles. Motivated by [18, 26], features {8, 9} and depend on the information of HI and RI, we de†ne the {17, 18} are the mean absolute deviation (MAD) and the heartbeat interval and respiratory interval in di‡erent averaged cumulative di‡erence (ACD) of HI and RI, timescales as respectively, which can also re°ect the variations of α β I h heartbeat interval and respiratory interval in each 60 s (t) n n n n1 HI  ,t  1, 3, 5, 7, 9, 11, 13, 15, (1) epoch. α β n1 n n α β I r (t) n1 n n n 3.2. Coordinated Features between HI and RI. Motivated by RI  ,t  1, 3, 5, 7, 9, 11, 13, 15, (2) α β n1 n n the cardiopulmonary coupling technology [34], we propose to characterize the coordinated features between HI and RI, re- where N is the number of heartbeat intervals within t-second ferred to as CHR features, aiming to compensate for the scale, I h and I r are the duration of n every interval. α is n n th n limitations of independent features of HI and RI. Speci†cally, the proportion of I h or I r in time t-second ( α  1). n n n n1 we de†ne the ratio of HI features over RI features to evaluate β  0 denotes n interval occurs in body movement episode; n th the similarities and di‡erences of the coordinated features in otherwise, β  1. Specially, if the t-second signal is †lled with di‡erent sleep stages. Similar to the independent features class, (t) (t) body movement, HI and RI are de†ned as the invalid CHR features also include the ratio of the mean (feature 19), value. Intuitively, the longer the timescale employed for HI and coe¯cient variation (feature 20), di‡erent percentiles (feature RI detection in (1) and (2), the higher the accuracy of HI and 21–31), MAD (feature 32), and ACD of HI and RI (feature 33), RI, since the relative errors are reduced in the statistical process. as shown in Table 3. By taking advantage of HI and RI in di‡erent timescales, in- dependent features with respect to HI and RI can be char- acterized accordingly. 3.3. BM Features. As reported by [35], body movement usually occurs in wake and light sleep (a.k.a., N1N2) stages due to sleep posture changes every 5–10 minutes, while rarely 3.1. Independent Features of HI and RI. Table 2 shows all appearing in deep sleep (i.e., N3) and REM stages. Motivated by independent features of HI and RI in di‡erent timescales. the above facts, we extracted the BM features for sleep stage �e motivation of independent features extraction with classi†cation, as shown in Table 4. respect to HI and RI is similar to the existing studies [26] HI (ms) RI (ms) Journal of Healthcare Engineering 5 (t) (t) Table 2: Independent features of HI (RI ). Feature Feature name Feature description index (t) (t) 1 (10) Mean Mean value of HI (RI ) (t) (t) 2 (11) CV Coefficient variation of HI (RI ): Standard deviation divided by mean (t) (t) 3–7 Ratio of percentile A and percentile B of HI (RI ): Inter ratio percentiles (12–16) (A, B) ∈ [(100, 0), (90, 10), (80, 20), (70, 30), (60, 40)] (t) (t) 8 (17) MAD Median absolute deviation of HI (RI ) Averaged cumulative difference: the moving average of the absolute difference between the former 30 9 (18) ACD (t) (t) seconds and the latter 30 seconds of HI (RI ) for the range from k − q to k + q minutes (q � 2) (t) (t) Table 3: Coordinated features of HI and RI . Feature index Feature name Feature description (t) (t) 19 Ratio of mean Ratio of mean of HI to mean of RI (t) (t) 20 Ratio of CV Ratio of CV of HI to CV of RI (t) (t) 21–31 Intra ratio percentiles Ratio of percentile A of HI to percentile B of RI : (A, B) ∈ [(100, 0), (90, 10), (80, 20), (70, 30), (60, 40), (50, 50), (40, 60), (30, 70), (20, 80), (10, 90), (0, 100)] (t) (t) 32 Ratio of MAD Ratio of MAD of HI to MAD of RI (t) (t) 33 Ratio of ACD Ratio of ACD of HI to ACD of RI Table 4: Features of BM. Feature Feature name Feature description index 34 Motion ratio *e proportion of body movement in the current epoch 35 Motion nums *e number of periods of successive one-value signal in the current epoch 36 Largest motion ratio Longest one period in epoch divided by 60 37 Average motion ratio Motion ratiodivided byMotion nums *e proportion of body movement of epoch located, respectively, n ([1, 2]) epochs 38–39 Motion ratio of the previous n epochs before the current one *e proportion of body movement of epoch located, respectively, n ([1, 2]) epochs 40–41 Motion ratio of the next n epochs after the current one Using the above extracted features from the non-contact- all Wake labeled periods exceeds 80% of a 5-minute measured vital signs, different classifiers including Random timescale, such a 5-minute timescale is fused as Forest (RF) [36], Support Vector Machine (SVM) [37], De- Wake stage. cision Tree (DT) [38], K-Nearest Neighbor (KNN) [39], and (4) Define P as the proportion of REM stages over REM AdaBoost [40] are employed for sleep stage classification. total sleep time; for two adjacent REM labels in- tervals with other stage labels, we adopt the following 4. Stage Fusion fusion criteria: (a) When P < 5%, two adjacent REM stages less Since the epoch-by-epoch classification is in a timescale of 60 s, REM the classified sleep stages are sparse in time domain. According than 20 minutes are fused as one REM stage. to the AASM, a rule-based postprocessing is proposed to (b) When P > 10%, two adjacent REM stages less REM than 7 minutes are fused as one REM stage. improve the prediction performance by reasonably fusing the sparsely classified epoch-by-epoch sleep stage in time. (c) When P ∈ [5%, 10%], two adjacent REM REM stages less than 15 minutes are fused as one REM (1) Following the rules of AASM, the sequence of sleep stage. stage is from Wake to NREM and then to REM. Based on this fact, the discrete REM labels directly 5. Experimental Results followed by Wake labels are modified to Wake labels. (2) If the label of a single epoch is different from that of For the evaluation of sleep staging performance, we both the previous and the following epochs, it will be adopt leave-one-out cross-validation. To elaborate a relabeled as that of the previous epoch. little further, classifiers are trained and tested by 16 and 1 (3) For sparsely predicted Wake labels interleaved with samples, respectively, which is repeated until every single either REM or NREM labels, when the proportion of sample is tested. *e 17 samples based on PSG sleep 6 Journal of Healthcare Engineering HI: 23.2% 0.08 0.07 BM: 25.4% 0.06 CHR: 32.6% 0.05 RI: 18.8% 0.04 0.03 0.02 0.01 0.00 5 10 15 20 25 30 35 40 Feature index Figure 4: Feature importance of HI, RI, BM, and CHR. 83 0.64 0.62 0.60 0.58 0.56 0.54 0.52 77 0.50 1357 9 11 13 15 13579 11 13 15 Time–scale Time–scale without CHR features without CHR features with CHR features with CHR features (a) (b) Figure 5: Accuracy and Kappa in di‡erent timescales. staging slices form a total of 14666 epochs, where Wake, 5.1. E“ect of Independent and Coordinated Features on Sleep REM, and NREM stages are 2668, 2562, and 9434 epochs, Staging. In order to examine the e‡ectiveness of the accounting for 18.19%, 17.47%, and 64.34%, respectively. extracted features using independent features of HI, RI, In this study, we adopt di‡erent weights assigned in [24] BM, and coordinated features between HI and RI, we †rst to avoid over†tting. Similar to [41], we quantify the evaluate the feature importance using RF classi†er [42]. performance in terms of accuracy and Cohen’s Kappa Figure 4 shows the contributions of both the independent coe¯cient, respectively, which are given by features (HI, RI, and BM) and the coordinated features (CHR) to sleep stage classi†cation. As shown in Figure 4, TP p − p class 0 e p (accuracy)  , Kappa  , the feature importance of HI class features, RI class T 1 − p class{Wake,NREM,REM} features, CHR class features, and BM class features ac- count for 25.4%, 23.2%, 32.6%, 18.8%, respectively. (3) Among them, ACD extracted by HI is the most important where T and TP denote the total number of epochs class feature, revealing that the cumulative di‡erence during and those correctly classi†ed into the corresponding sleep is particularly important for sleep staging. It is also class and p is the hypothetical probability of chance e noted that CHR features contribute most to sleep stage agreement. classi†cation, which demonstrates that coordination be- tween HI and RI is essential to discriminate sleep stages. TP + FP TP + FN class class class class p   × . Specially, Ratio of ACD to CHR performs better, which is T T class{Wake,NREM,REM} consistent with the excellent performance of ACD in the independent features. (4) Accuracy (%) Feature importance Kappa Journal of Healthcare Engineering 7 1.0 0.8 0.6 0.4 0.2 0.0 NREM REM Wake without sleep stage fusion with sleep stage fusion Figure 6: �e classi†cation performance with and without the proposed sleep stage fusion. Wake Wake REM REM NREM NREM 0 200 400 600 800 1000 0 200 400 600 800 1000 epoch epoch (a) (b) Wake REM NREM 0 200 400 600 800 1000 epoch (c) Figure 7: (a) Reference sleep stages provided by PSG. (b) Classi†ed sleep stages estimated without stage fusion. (c) Classi†ed sleep stages estimated with stage fusion. NREM 8466 531 437 REM 1068 1376 118 Wake 348 86 2234 NREM REM Wake Predicted class Figure 8: Confusion matrix of Wake-REM-NREM sleep stage classi†cation. True class Accuracy of three stages 8 Journal of Healthcare Engineering Table 5: Information of the subjects with sleep-disordered breathing. Subject ID Gender Age AHI (times/hour) Severity Total (night) Valid (night) A1 M 49 10.6 Mild 1 1 A2 F 70 6.3 Mild 2 1 A3 M 70 7.7 Mild 1 1 A4 F 69 8.1 Mild 1 1 A5 M 73 5.9 Mild 2 1 A6 M 50 6.7 Mild 1 1 A7 M 51 7.5 Mild 1 1 Table 6: Performance of the sleep-disordered breathing subjects. Subject ID Accuracy (%) Kappa 1 76.07 0.58 2 75.30 0.48 3 65.11 0.39 4 78.49 0.63 5 68.11 0.41 6 79.81 0.67 7 82.62 0.62 Average 75.07 0.54 Std 5.85 0.11 Wake REM NREM 0 200 400 600 800 epoch (a) Wake REM NREM 0 200 400 600 800 epoch (b) Wake REM NREM 0 200 400 600 800 epoch (c) Figure 9: (a) Reference sleep stages provided by PSG. (b) Classified sleep stages estimated without stage fusion. (c) Classified sleep stages estimated with stage fusion. Journal of Healthcare Engineering 9 with sleep-disordered breathing (most of these subjects *en, we analyze the effect of time resolution with respect to features extraction on sleep staging. As shown in Figure 5, both suffer from mild sleep apnea syndrome), aiming to verify the effectiveness of the proposed design. *e data used in the the accuracy and Kappa tend to increase when t< 10 s and then decrease as the timescale increases. A possible explanation to experiment was jointly recorded by *e First Affiliate this behavior is that non-contact-measured HI and RI suffers Hospital of Guangzhou Medical University and Guangzhou from inevitable errors due to artifact motion and noise, as SENVNV Co., and the experiment has obtained the consent shown in Figure 3. In this case, increasing the timescale yields a of the subjects, and personal private information is kept higher accuracy of HI and RI. Benefiting from the increase of confidential. *e information of the recruited subjects is timescale when t< 10 s, the improvement of feature accuracy listed in Table 5. improved sleep staging performance. As the timescale grows Using the proposed scheme with noncontact-measured larger (i.e., t> 10 s), it simultaneously reduces the sensitivity vital signs, the performance of sleep staging in comparison with PSG is shown in Table 6. It can be seen that the averaged with respect to the variation of HI and RI in different timescales, thus leading to a reduced performance. Furthermore, it can be accuracy and Kappa coefficient with respect to the recruited subjects suffering from sleep-disordered breathing are 75.07% seen that the classification results using CHR features perform better than those without CHR features, demonstrating the and 0.54, respectively. *is demonstrates the robustness of the effectiveness of the CHR features. *e highest accuracy and so-designed features and model. Figure 9 provides a typical Kappa are 82.42% and 0.63 when the timescale is 9 s. sleep stage classification from a subject with sleep-disordered breathing. 5.2. Sleep Stage Fusion. Next, we investigate the effect of 6. Conclusion stage fusion as the postprocessing of classification on sleep staging. Figure 6 shows the classification performance with *is paper studied feature-aided sleep stage classification using and without the proposed sleep stage fusion. It can be seen noncontact-measured vital signs. In addition to the analysis of that the applied stage fusion significantly improves the re- independent features such as HI, RI, and BM, which are sults, especially in Wake and REM. *e accuracy of Wake is characterized from BCG, respiratory rate, and body move- increased by 31.6%, and that of REM is increased by 39.9%. ments signals in different timescales, we validated through *e result is reasonable since the epoch-by-epoch classified experiments that the coordinated features between HI and RI sleep stages over a specific interval are mapped to an play an important role in sleep staging. In order to improve the identical sleep stage. Moreover, the accuracy of NREM still performance of classification, we developed a rule-based maintains a high level of 89.8%, although it may be reduced postprocessing to fuse the classified results of discrete time by 3.2% from 93.0% due to the slight fusion errors in the dimensions. *e experimental results in comparison with PSG fusion of Wake and REM. demonstrate the effectiveness of the proposed design. To further demonstrate the performance improvement of stage fusion, Figure 7 provides a typical example with and without Data Availability the proposed sleep stage fusion. As can be seen from Figure 7, the proposed stage fusion significantly improves the results. No datasets are generated during the current study. *e datasets analyzed during this work are publicly available. 5.3. Wake-REM-NREM Discrimination. We further verify Consent the confusion matrix based on RF classifier, and the result is shown in Figure 8. It can be seen from Figure 8 that the Informed consent was obtained from all subjects involved in prediction accuracy rates of Wake, REM, and NREM are the study. 67.3%, 53.7%, and 83.7%, respectively. Among them, the dominant error comes from the misclassification between Conflicts of Interest REM and NREM stages. Taking the experimental results into account, the main reason of low accuracy of recognition in *e authors declare that there are no conflicts of interest REM could be summarized to three aspects. (1) Frequent regarding the publication of this paper. body movements occur in both Wake and REM stages, leading to a lower accuracy of feature extraction in terms of Authors’ Contributions HI, RI, and CHR, thereby reducing the accuracy of classi- Z.W. and H.Z. carried out the data analysis and experiment fication. (2) *e proportion of REM epochs is lower than and wrote the manuscript. S.Z., B.Y., P.C., Z.P., and H.Z. that of NREM and Wake. *erefore, the proportion of supervised the methodology and organization of this work. misclassified epochs of REM in REM is higher than that in All authors have read and agreed to the published version of the whole sleep stage, whereas the accuracy of the whole the manuscript. system will be less affected. Acknowledgments 5.4. Validation on Subjects Suffering from Sleep-Disorderd Breathing. As a proof-of-concept with respect to the so- *is work was supported by the Natural Science Foundation designed sleep staging model, we further consider 7 subjects of Guangdong Province (no. 2022A1515010104), Blue Fire 10 Journal of Healthcare Engineering Medicine and Biology Society, pp. 2602–2605, Vancouver, Innovation Project of the Ministry of Education (Huizhou) Canada, October 2008. (no. CXZJHZ201803(, Special Funds for the Cultivation of [15] S. J. Redmond and C. Heneghan, “Cardiorespiratory-based Guangdong College Students’ Scientific and Technological sleep staging in subjects with obstructive sleep apnea,” IEEE Innovation (no. pdjh2021b0133), and also supported by Transactions on Biomedical Engineering, vol. 53, no. 3, Science and Technology Project of Guangzhou (nos. pp. 485–496, 2006. 202002030353 and 202102021114). [16] S. J. Redmond, P. de Chazal, C. O’Brien, S. Ryan, W. T. McNicholas, and C. Heneghan, “Sleep staging using cardiorespiratory signals,” Somnologie - Schlafforschung und References Schlafmedizin, vol. 11, no. 4, pp. 245–256, 2007. [17] S. Devot, R. Dratwa, and E. Naujokat, “Sleep/wake detection [1] Y. Castillo-Escario, H. Kumru, I. Ferrer-Lluis, J. Vidal, and R. Jane, ´ “Detection of sleep-disordered breathing in patients based on cardiorespiratory signals and actigraph,” in Pro- with spinal cord injury using a smartphone,” Sensors, vol. 21, ceedings of the 2010 Annual International Conference of the no. 21, p. 7182, 2021. IEEE Engineering in Medicine and Biology, pp. 5089–5092, [2] I. Ferrer-Lluis, Y. Castillo-Escario, J. M. Montserrat, and Buenos Aires, Argentina, November 2010. R. Jane, “Analysis of smartphone triaxial accelerometry for [18] T. Willemen, D. Van Deun, V. Verhaert et al., “An evaluation monitoring sleep-disordered breathing and sleep position at of cardiorespiratory and movement features with respect to home,” IEEE Access, vol. 8, pp. 71231–71244, 2020. sleep-stage classification,” IEEE Journal of Biomedical and [3] F. P. Cappuccio, L. D’Elia, P. Strazzullo, and M. A. Miller, Health Informatics, vol. 18, no. 2, pp. 661–669, 2014. “Sleep duration and all-cause mortality: a systematic review [19] D. Huysmans, I. Castro, P. Borzee ´ et al., “Capacitively-cou- and meta-analysis of prospective studies,” Sleep, vol. 33, no. 5, pled ECG and respiration for sleep-wake prediction and risk pp. 585–592, 2010. detection in sleep apnea patients,” Sensors, vol. 21, no. 19, [4] R. B. Berry, R. Budhiraja, D. J. Gottlieb et al., “Rules for p. 6409, 2021. scoring respiratory events in sleep: update of the 2007 AASM [20] M. Gaiduk, R. Seepold, T. Penzel, J. A. Ortega, M. Glos, and manual for the scoring of sleep and associated events,” Journal N. M. Madrid, “Recognition of sleep/wake states analyzing of Clinical Sleep Medicine, vol. 08, no. 05, pp. 597–619, 2012. heart rate, breathing and movement signals,” in Proceedings of [5] A. Rechtschaffen, “A manual for standardized terminology, the 2019 41st Annual International Conference of the IEEE techniques and scoring system for sleep stages in human Engineering in Medicine and Biology Society, pp. 5712–5715, subjects,” Brain Inf. Serv., 1968. EMBC), Berlin, Germany, October 2019. [6] B. Haex, Back and Bed: Ergonomic Aspects of Sleeping, CRC [21] M. Gaiduk, T. Penzel, J. A. Ortega, and R. Seepold, “Auto- Press, Boca Raton, FL, US, 2004. matic sleep stages classification using respiratory, heart rate [7] K. Adam and I. Oswald, “Sleep helps healing,” BMJ, vol. 289, and movement signals,” Physiological Measurement, vol. 39, no. 6456, pp. 1400-1401, Article ID 6437572, 1984. no. 12, Article ID 124008, 2018. [8] P. Meerlo, R. E. Mistlberger, B. L. Jacobs, H. Craig HellerCraig [22] J. Siyanbade, B. Abdulrazak, and I. Sadek, “Unobtrusive Heller, and D. McGinty, “New neurons in the adult brain: the monitoring of sleep cycles: a technical review,” Bio- role of sleep and consequences of sleep loss,” Sleep Medicine MedInformatics, vol. 2, no. 1, pp. 204–216, 2022. Reviews, vol. 13, no. 3, pp. 187–194, Article ID 18848476, 2009. [23] M. Migliorini, A. M. Bianchi, D. Nistico` et al., “Automatic [9] Y. R. Chien, C. H. Wu, and H. W. Tsao, “Automatic sleep- sleep staging based on ballistocardiographic signals recorded arousal detection with single-lead EEG using stacking en- through bed sensors,” in Proceedings of the 2010 Annual semble learning,” Sensors, vol. 21, no. 18, p. 6049, Article ID International Conference of the IEEE Engineering in Medicine 34577255, 2021. and Biology, pp. 3273–3276, Buenos Aires, Argentina, No- [10] T. Penzel and R. Conradt, “Computer based sleep recording vember 2010. and analysis,” Sleep Medicine Reviews, vol. 4, no. 2, [24] S. H. Choi, H. B. Kwon, H. W. Jin et al., “Long short-term pp. 131–148, 2000. memory networks for unconstrained sleep stage classification [11] A. Domingues, T. Paiva, and J. M. Sanches, “Sleep and using polyvinylidene fluoride film sensor,” IEEE Journal of wakefulness state detection in nocturnal actigraphy based on Biomedical and Health Informatics, vol. 24, no. 12, movement information,” IEEE Transactions on Biomedical pp. 3606–3615, 2020. Engineering, vol. 61, no. 2, pp. 426–434, Article ID 24013826, [25] J. Drzazga and B. Cyganek, “An LSTM network for apnea and 2014. hypopnea episodes detection in respiratory signals,” Sensors, [12] J. Tilmanne, J. Urbain, M. V. Kothare, A. V. Wouwer, and vol. 21, no. 17, p. 5858, 2021. S. V. Kothare, “Algorithms for sleep-wake identification using [26] H. Hong, L. Zhang, C. Gu, Y. Li, G. Zhou, and X. Zhu, actigraphy: a comparative study and new results,” Journal of “Noncontact sleep stage estimation using a CW Doppler Sleep Research, vol. 18, no. 1, pp. 85–98, Article ID 19250177, radar,” IEEE Journal on Emerging and Selected Topics in 2009. Circuits and Systems, vol. 8, no. 2, pp. 260–270, 2018. [13] A. Tataraidze, L. Anishchenko, L. Korostovtseva, B. J. Kooij, [27] T. Lauteslager, S. Kampakis, A. J. Williams, M. Maslik, and M. Bochkarev, and Y. Sviryaev, “Sleep stage classification F. Siddiqui, “Performance evaluation of the circadia con- based on respiratory signal,” in Proceedings of the 2015 37th tactless breathing monitor and sleep analysis algorithm for Annual International Conference of the IEEE Engineering in sleep stage classification,” in Proceedings of the Annual In- Medicine and Biology Society (EMBC), pp. 358–361, IEEE, ternational Conference of the IEEE Engineering in Medicine Milan, Italy, August 2015. and Biology Society, pp. 5150–5153, Montreal, ´ QC, Canada, [14] S. Canisius, T. Ploch, V. Gross, A. Jerrentrup, T. Penzel, and July 2020. K. Kesper, “Detection of sleep disordered breathing by au- [28] F. Zhang, C. Wu, B. Wang et al., “SMARS: sleep monitoring tomated ECG analysis,” in Proceedings of the 2008 30th via ambient radio signals,” IEEE Transactions on Mobile Annual International Conference of the IEEE Engineering in Computing, vol. 20, no. 1, pp. 217–231, 2021. Journal of Healthcare Engineering 11 [29] J. Liang, J. Huang, L. Mu et al., “An effective algorithm for beat-to-beat heart rate monitoring from ballistocardiograms,” Journal of Medical Imaging and Health Informatics, vol. 10, no. 3, pp. 633–640, 2020. [30] I. Sadek, E. Seet, J. Biswas, B. Abdulrazak, and M. Mokhtari, “Nonintrusive vital signs monitoring for sleep apnea patients: a preliminary study,” IEEE Access, vol. 6, pp. 2506–2514, 2017. [31] S. M. Proll, ¨ S. Hofbauer, C. Kolbitsch, R. Schubert, and K. D. Fritscher, “Ejection wave segmentation for contact-free heart rate estimation from ballistocardiographic signals,” in Proceedings of the 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 3571–3576, EMBC), Berlin, Germany, October 2019. [32] J. Gilles, “Empirical wavelet transform,” IEEE Transactions on Signal Processing, vol. 61, no. 16, pp. 3999–4010, 2013. [33] M. Ali, A. AhmadAhmad, S. SultanaSultana, S. R. MirMir, and M. Sawan, “Chemical constituents from the seed husks of Oryza sativa L,” Natural Product Research, vol. 21, pp. 14569–14586, 2021. [34] R. J. *omas, J. E. Mietus, C. K. Peng, and A. L. Goldberger, “An electrocardiogram-based technique to assess cardiopul- monary coupling during sleep,” Sleep, vol. 28, no. 9, pp. 1151–1161, 2005. [35] M. Kagawa, N. Sasaki, K. Suzumura, and T. Matsui, “Sleep stage classification by body movement index and respiratory interval indices using multiple radar sensors,” in Proceedings of the 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 7606–7609, EMBC), Milan, Italy, November 2015. [36] A. Liaw and M. Wiener, “Classification and regression by randomForest,” R News, vol. 2, pp. 18–22, 2002. [37] C. C. Chang and C. J. Lin, “Libsvm,” ACM Transactions on Intelligent Systems and Technology, vol. 2, no. 3, pp. 1–27, 2011. [38] K. Polat and S. Gune ¨ ¸, s “Classification of epileptiform EEG using a hybrid system based on decision tree classifier and fast Fourier transform,” Applied Mathematics and Computation, vol. 187, no. 2, pp. 1017–1026, 2007. [39] T. K. Ho, “Nearest neighbors in random subspaces,” in Joint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pat- tern Recognition (SSPR), pp. 640–648, Springer, Berlin, Heidelberg, 1998. [40] Z. H. Zhou, Ensemble Methods: Foundations and Algorithms, CRC Press, Boca Raton, FL, US, 2012. [41] A. J. Viera and J. M. Garrett, “Understanding interobserver agreement: the kappa statistic,” Family Medicine, vol. 37, no. 5, pp. 360–363, 2005. [42] B. Xue, B. Deng, H. Hong, Z. Wang, X. Zhu, and D. D. Feng, “Non-contact sleep stage detection using canonical correla- tion analysis of respiratory sound,” IEEE Journal of Bio- medical and Health Informatics, vol. 24, no. 2, pp. 614–625, http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Healthcare Engineering Hindawi Publishing Corporation

Sleep Staging Using Noncontact-Measured Vital Signs

Download PDF
11 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/sleep-staging-using-noncontact-measured-vital-signs-SjV2Dfg0Ul

References (44)

Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2022 Zixia Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
2040-2295
eISSN
2040-2309
DOI
10.1155/2022/2016598
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Healthcare Engineering Volume 2022, Article ID 2016598, 11 pages https://doi.org/10.1155/2022/2016598 Research Article 1 1 2,3 4 4 Zixia Wang , Shuai Zha , Baoxian Yu , Pengbin Chen , Zhiqiang Pang , 2,3 and Han Zhang Department of Physics and Telecommunication Engineering, South China Normal University, Guangzhou 510006, China Department of Electronic and Information Engineering, South China Normal University, Foshan 528000, China Guangdong Provincial Engineering Technology Research Center of Cardiovascular Individual Medicine & Big Data, South China Normal University, Guangzhou 510006, China Guangzhou SENVIV Technology Co.,Ltd., Guangzhou 510006, China Correspondence should be addressed to Baoxian Yu; yubx@m.scnu.edu.cn and Han Zhang; zhanghan@scnu.edu.cn Received 12 January 2022; Revised 20 May 2022; Accepted 13 June 2022; Published 8 July 2022 Academic Editor: Haihong Zhang Copyright © 2022 Zixia Wang 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. As a physiological phenomenon, sleep takes up approximately 30% of human life and signi†cantly a‡ects people’s quality of life. To assess the quality of night sleep, polysomnography (PSG) has been recognized as the gold standard for sleep staging. �e drawbacks of such a clinical device, however, are obvious, since PSG limits the patient’s mobility during the night, which is inconvenient for in-home monitoring. In this paper, a noncontact vital signs monitoring system using the piezoelectric sensors is deployed. Using the so-designed noncontact sensing system, heartbeat interval (HI), respiratory interval (RI), and body movements (BM) are separated and recorded, from which a new dimension of vital signs, referred to as the coordination of heartbeat interval and respiratory interval (CHR), is obtained. By extracting both the independent features of HI, RI, and BM and the coordinated features of CHR in di‡erent timescales, Wake-REM-NREM sleep staging is performed, and a postprocessing of staging fusion algorithm is proposed to re†ne the accuracy of classi†cation. A total of 17 all-night recordings of noncontact measurement simultaneous with PSG from 10 healthy subjects were examined, and the leave-one-out cross-validation was adopted to assess the performance of Wake-REM-NREM sleep staging. Taking the gold standard of PSG as reference, numerical results show that the proposed sleep staging achieves an averaged accuracy and Cohen’s Kappa index of 82.42% and 0.63, respectively, and performs robust to subjects su‡ering from sleep-disordered breathing. into N1, N2, and N3 stages [4]. REM and NREM alternate 1. Introduction in cycles of about 90 min [5]. NREM, especially deep sleep Sleep plays an important role in body recovery, memory, (N3), is more prominent during the †rst hours of sleep and immunity enhancement, which takes almost one- and is essential toward physical recovery [6, 7]. REM link third of human life [1]. Poor sleep quality usually results to dreaming, more prominently during the last hours of in physical and mental health problems, such as fatigue, sleep, acts toward the recovery of people’s mental state [8]. anxiety, and even death [2]. It has been reported that sleep �erefore, a †ne-grained classi†cation of Wake-REM- duration is closely related to mortality [3]. �erefore, NREM stages is crucial for assessing sleep quality. long-term sleep quality monitoring is of great signi†cance Polysomnography (PSG) is the gold standard in the to protect human health. clinic for evaluating sleep quality and diagnosing sleep- Sleep staging is an important characteristic to qualify related diseases. Physiological signal acquisition during sleep quality. According to the American Academy of sleep is usually conducted in the hospital since the ac- Sleep Medicine (AASM), a complete sleep cycle consists of quisition process of vital signs is cumbersome. To obtain wake, rapid eye movement (REM), and non-REM the physiology signals including electroencephalography (NREM) stages, where the last one can be further divided (EEG), electrooculography (EOG), electromyography 2 Journal of Healthcare Engineering features between heartbeat interval and respiratory in- (EMG), electrocardiography (ECG), and respiratory rate [9], a large number of sensors are attached to the skin of terval in different timescales has not been reported yet. In order to address the above issues, this paper focuses patients directly [10], which is impractical for long-term monitoring at home. on a noncontact sensing-based sleep staging for Wake/ Many studies have attempted to classify sleep stages REM/NREM classification using vital signs recorded by automatically under more natural sleep monitoring piezoelectric sensors. Specially, the contributions of this conditions by using a limited number of sensing channels. paper, in comparison with the existing studied, can be In [11], wrist-actigraphy was used to record body summarized as follows. movements during sleep and achieved a sleep/wake (1) A noncontact vital signs monitoring system using classification performance of 77.8%. In [12], the authors piezoelectric sensors is deployed, by which both the took infants as experimental subjects and used wrist- independent features of heartbeat interval (HI), re- actigraphy to achieve a sleep/wake classification. How- spiratory interval (RI), body movements (BM), and ever, this method has limitations when the subject likes to the coordinated features between HI and RI (a.k.a. be quiet in wake period. Motivated by the fact that the CHR) are extracted from the noncontact-measured autonomic nerves (assessed by heartbeat and respiratory vital signs and employed for Wake/REM/NREM rate) changes in different sleep stages, respiratory in- stage classification. In addition, the effects of features ductive plethysmography signals have been used in three- in different timescales on sleep staging are also stage classification (i.e., REM/NREM/Wake) and achieved analyzed. the classification performance of 80.38% [13]. In [14], the (2) A postprocessing with respect to the fusion of the authors also considered single-lead ECG heartbeat in- classified stages is developed to further improve the terval detection and yielded the sleep/wake classification performance of sleep staging. For validations, 17 all- performance of 76%. In order to improve the classification night experiments were examined simultaneously performance, cardiorespiratory features including both with PSG. Numerical comparisons demonstrate that the features extracted from heartbeat interval and respi- the proposed sleep staging achieves average accuracy ratory are extensively studied [15–19]. Specifically, using and Kappa index of 82.42% and 0.63, respectively, ECG sensors, features of heart rate variability in time and and performs robust to subjects suffering from sleep- frequency domain are fused with the time and effort disordered breathing. (fluctuation) features of respiratory signal to aid sleep staging, and the performance of REM/NREM/Wake stages classification was improved to almost 80% [16]. In 2. Materials and Methods [20, 21], ECG sensors were used to extract features with respect to heart rate, respiratory rate, and body move- 2.1. System Setup and Vital Signs Acquisition. For non- ment. Although the existing single-lead ECG sensor based contact vital signs acquisition, the noninvasive heart rate methods can achieve satisfactory performance of sleep and respiratory rate sensing device developed by staging, the acquisition of signals requires physical con- Guangzhou SENVNV Co. was deployed [29], where the tact with subjects’ skin, which is inappropriate to long- sleep-related physiological signals, including BCG, re- term home monitoring. spiratory signal, and artifact motion can be recorded In order to solve the discomfort between sensors and during night sleep in a noncontact manner. After pre- skin in the process of signal acquisition, many studies processing, heartbeat interval, respiratory interval, and have focused on noncontact monitoring technologies body movements are separated, and sleep stage-related [22]. Using piezoelectric sensors, heart rate was obtained features are extracted from the above physiological signals based on noncontact-measured ballistocardiogram (BCG) and then fed into machine learning classifiers for sleep signals, by which three-stage classification of night sleep is stage classification. Finally, an empirical rule-based performed [23]. In [24], the authors considered PVDF postprocessing is applied to improve the classification sensors to record night-sleep physiological signals, in- performance. Figure 1 depicts the framework of the cluding BCG, respiratory signal, and body movements, noncontact sleep staging system. and then employed long short-term memory (LSTM) *e noncontact vital signs monitoring system neural network [25] to perform end-to-end sleep stage deployed in this study consists of a piezoelectric sensor, classifications. By analogy, in [26, 27], continuous-wave circuit, and processing modules. As shown in Figure 2, the Doppler radar sensing technology was adopted to dis- piezoelectric sensor module is placed under the pillow to tinguish Wake/REM/LightSleep/DeepSleep states and perceive the vibration induced by heartbeat, respiratory achieved an accuracy of 81% and 66.7% in comparison rate, and body movements. *e converted voltage signals with PSG standard. Taking radar sensing as the vital sign are then amplified, sampled at 1 kHz in the circuit module. monitoring system, the authors in [28] evaluated the A 12-bit analog-to-digital conversion (ADC) is conducted performance of Wake/REM/NREM sleep staging, where before transmission of vital signs from circuit to pro- the overall accuracy reached up to 88.4%. Although sleep- cessing module for offline signal processing and sleep related vital signs (i.e., heart rate and respiratory rate) and stage classification. For reference, we simultaneously body movements can be measured by using the existing collect EEG, ECG, EOG, and EMG from PSG (Medcare) as noncontact sensing devices, the effect of coordinated the gold standard to evaluate the staging performance. Journal of Healthcare Engineering 3 Pre-process of Feature extraction vital sings Sleep stage HI BCG Non-contact vital Sleep stage classification Wake sings acquisition RI Classification Stage fusion NREM Respiratory Vital signs CHR REM Body BM movement Figure 1: �e overall architecture of the noncontact sleep monitoring system. (a) (b) Figure 2: �e overview of the vital signs monitoring system. (a) �e subject simultaneously monitored by both PSG and the noncontact vital signs system. (b) �e device of the vital signs monitoring system. 2.2. Noncontact Vital Signs Acquisition. �is study involved Table 1: Information of the datasets. a total of 10 healthy subjects (7 males and 3 females), who Subject Weight Height Total are college students, aged from 21 to 25 years old. All Gender Age Valid (night) ID (kg) (m) (night) subjects are healthy and have a regular night sleep with an N1 M 23 65.23 1.70 5 4 average duration of 6 to 9 hours. In addition, subjects with N2 M 22 66.52 1.72 4 3 sleep disorder, smoking habits, intake of medicine, or N3 F 21 48.36 1.60 1 0 drinks infecting sleep will not be included. As shown in N4 M 23 70.44 1.74 4 4 Table 1, a total of 24 night-sleep experiments are carried N5 F 24 52.62 1.64 1 0 out, in which the recorded data in 7 night sleep are invalid N6 M 23 69.87 1.73 2 2 due to the collapse of the electrodes of PSG. N7 M 22 62.15 1.67 3 2 N8 F 25 68.53 1.76 1 0 N9 M 22 60.76 1.66 1 1 2.3. Preprocessing of Vital Signs. Since the vital signs N10 M 23 67.39 1.75 2 1 measured by the noncontact device are mixed with BCG, respiratory signal, body movements, and noise, pre- processing of vital signs is †rstly performed to separate Next, we extract BCG from the resulting signals of BM  0. di‡erent types of signals for heartbeat interval and re- Considering that the spectrum of a typical BCG ranges from spiratory interval detection. �e procedures are shown as 3 Hz to 10 Hz [31], we employ a 2nd-order Butterworth follows. bandpass †lter ranges 2.5–10 (Hz) to remove the undesired First, we remove the power frequency noise by using a signal interference and then detect heartbeat interval by using a band-stop †lter with lower and upper cuto‡ frequencies of forward and backward approach [29]. 49 Hz and 51 Hz, respectively. �en, we identify body move- For respiratory interval detection, we remove the detected ments (BM) from the processed signal. �e reason is that, on BCG signal from the recorded vital signs and then separate the respiratory waves directly by using discrete wavelet transform the one hand, body movements can signi†cantly a‡ect the detection and analysis of vital signs (i.e., BCG and respiratory and Sym8 wavelet package at 8 scales [32]. Finally, respiratory signals). On the other hand, as will be illustrated later, body interval can be obtained by using peak detection [33]. movements signal is an important feature for sleep staging. For validation, a typical example of 20-minute comparison Unlike [30], body movements (BM) are detected in multi-time- between the noncontact-measured HI and RI and that obtained scale procedure. �e detailed procedures are as follows. If the using BIOPAC MP160, which are recognized as the gold peak-to-valley di‡erence (PVD) within a 2 s window is 2.2 standard in related works. As can be observed from Figure 3, times greater than any one of the medians of PVD within a although some errors occur in very limited areas due to the multi-time-scale epochs of 30 s, 60 s, 120 s, and 300 s, the state interference of body motion, the noncontact-measured HI and RI are almost indistinguishable from those obtained using ECG of 2 s window will be considered as body movement; that is, BM  1; otherwise, BM  0. and belt sensors. �e results demonstrate that the measured HI 4 Journal of Healthcare Engineering 0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200 time (s) time (s) RI of piezoelectric sensors HI of piezoelectric sensors HI of PSG RI of PSG (a) (b) Figure 3: An example of comparison between detected HI and RI in a noncontact manner and that obtained by using the gold standard device. and RI in a noncontact manner can be further used as features since the rhythm of heartbeat and respiratory interval vary in for sleep stage classi†cation. di‡erent sleep stages. Speci†cally, features {1, 2} and {10, 11} are the mean and coe¯cient variation of independent HI and RI in 3. Feature Extraction di‡erent timescales in a 60 s epoch. Features {3–7} and {12–16} further describe the trend of °uctuation with Using the non-contact-measured vital signs (i.e., HI, RI, respect to heartbeat interval and respiratory interval and BM signals), we propose feature extraction based on over a 60 s epoch by evaluating the ratio of di‡erence in the above independent and coupled vital signs for sleep terms of heartbeat interval and respiratory interval stage classi†cation. Considering that the relevant features percentiles. Motivated by [18, 26], features {8, 9} and depend on the information of HI and RI, we de†ne the {17, 18} are the mean absolute deviation (MAD) and the heartbeat interval and respiratory interval in di‡erent averaged cumulative di‡erence (ACD) of HI and RI, timescales as respectively, which can also re°ect the variations of α β I h heartbeat interval and respiratory interval in each 60 s (t) n n n n1 HI  ,t  1, 3, 5, 7, 9, 11, 13, 15, (1) epoch. α β n1 n n α β I r (t) n1 n n n 3.2. Coordinated Features between HI and RI. Motivated by RI  ,t  1, 3, 5, 7, 9, 11, 13, 15, (2) α β n1 n n the cardiopulmonary coupling technology [34], we propose to characterize the coordinated features between HI and RI, re- where N is the number of heartbeat intervals within t-second ferred to as CHR features, aiming to compensate for the scale, I h and I r are the duration of n every interval. α is n n th n limitations of independent features of HI and RI. Speci†cally, the proportion of I h or I r in time t-second ( α  1). n n n n1 we de†ne the ratio of HI features over RI features to evaluate β  0 denotes n interval occurs in body movement episode; n th the similarities and di‡erences of the coordinated features in otherwise, β  1. Specially, if the t-second signal is †lled with di‡erent sleep stages. Similar to the independent features class, (t) (t) body movement, HI and RI are de†ned as the invalid CHR features also include the ratio of the mean (feature 19), value. Intuitively, the longer the timescale employed for HI and coe¯cient variation (feature 20), di‡erent percentiles (feature RI detection in (1) and (2), the higher the accuracy of HI and 21–31), MAD (feature 32), and ACD of HI and RI (feature 33), RI, since the relative errors are reduced in the statistical process. as shown in Table 3. By taking advantage of HI and RI in di‡erent timescales, in- dependent features with respect to HI and RI can be char- acterized accordingly. 3.3. BM Features. As reported by [35], body movement usually occurs in wake and light sleep (a.k.a., N1N2) stages due to sleep posture changes every 5–10 minutes, while rarely 3.1. Independent Features of HI and RI. Table 2 shows all appearing in deep sleep (i.e., N3) and REM stages. Motivated by independent features of HI and RI in di‡erent timescales. the above facts, we extracted the BM features for sleep stage �e motivation of independent features extraction with classi†cation, as shown in Table 4. respect to HI and RI is similar to the existing studies [26] HI (ms) RI (ms) Journal of Healthcare Engineering 5 (t) (t) Table 2: Independent features of HI (RI ). Feature Feature name Feature description index (t) (t) 1 (10) Mean Mean value of HI (RI ) (t) (t) 2 (11) CV Coefficient variation of HI (RI ): Standard deviation divided by mean (t) (t) 3–7 Ratio of percentile A and percentile B of HI (RI ): Inter ratio percentiles (12–16) (A, B) ∈ [(100, 0), (90, 10), (80, 20), (70, 30), (60, 40)] (t) (t) 8 (17) MAD Median absolute deviation of HI (RI ) Averaged cumulative difference: the moving average of the absolute difference between the former 30 9 (18) ACD (t) (t) seconds and the latter 30 seconds of HI (RI ) for the range from k − q to k + q minutes (q � 2) (t) (t) Table 3: Coordinated features of HI and RI . Feature index Feature name Feature description (t) (t) 19 Ratio of mean Ratio of mean of HI to mean of RI (t) (t) 20 Ratio of CV Ratio of CV of HI to CV of RI (t) (t) 21–31 Intra ratio percentiles Ratio of percentile A of HI to percentile B of RI : (A, B) ∈ [(100, 0), (90, 10), (80, 20), (70, 30), (60, 40), (50, 50), (40, 60), (30, 70), (20, 80), (10, 90), (0, 100)] (t) (t) 32 Ratio of MAD Ratio of MAD of HI to MAD of RI (t) (t) 33 Ratio of ACD Ratio of ACD of HI to ACD of RI Table 4: Features of BM. Feature Feature name Feature description index 34 Motion ratio *e proportion of body movement in the current epoch 35 Motion nums *e number of periods of successive one-value signal in the current epoch 36 Largest motion ratio Longest one period in epoch divided by 60 37 Average motion ratio Motion ratiodivided byMotion nums *e proportion of body movement of epoch located, respectively, n ([1, 2]) epochs 38–39 Motion ratio of the previous n epochs before the current one *e proportion of body movement of epoch located, respectively, n ([1, 2]) epochs 40–41 Motion ratio of the next n epochs after the current one Using the above extracted features from the non-contact- all Wake labeled periods exceeds 80% of a 5-minute measured vital signs, different classifiers including Random timescale, such a 5-minute timescale is fused as Forest (RF) [36], Support Vector Machine (SVM) [37], De- Wake stage. cision Tree (DT) [38], K-Nearest Neighbor (KNN) [39], and (4) Define P as the proportion of REM stages over REM AdaBoost [40] are employed for sleep stage classification. total sleep time; for two adjacent REM labels in- tervals with other stage labels, we adopt the following 4. Stage Fusion fusion criteria: (a) When P < 5%, two adjacent REM stages less Since the epoch-by-epoch classification is in a timescale of 60 s, REM the classified sleep stages are sparse in time domain. According than 20 minutes are fused as one REM stage. to the AASM, a rule-based postprocessing is proposed to (b) When P > 10%, two adjacent REM stages less REM than 7 minutes are fused as one REM stage. improve the prediction performance by reasonably fusing the sparsely classified epoch-by-epoch sleep stage in time. (c) When P ∈ [5%, 10%], two adjacent REM REM stages less than 15 minutes are fused as one REM (1) Following the rules of AASM, the sequence of sleep stage. stage is from Wake to NREM and then to REM. Based on this fact, the discrete REM labels directly 5. Experimental Results followed by Wake labels are modified to Wake labels. (2) If the label of a single epoch is different from that of For the evaluation of sleep staging performance, we both the previous and the following epochs, it will be adopt leave-one-out cross-validation. To elaborate a relabeled as that of the previous epoch. little further, classifiers are trained and tested by 16 and 1 (3) For sparsely predicted Wake labels interleaved with samples, respectively, which is repeated until every single either REM or NREM labels, when the proportion of sample is tested. *e 17 samples based on PSG sleep 6 Journal of Healthcare Engineering HI: 23.2% 0.08 0.07 BM: 25.4% 0.06 CHR: 32.6% 0.05 RI: 18.8% 0.04 0.03 0.02 0.01 0.00 5 10 15 20 25 30 35 40 Feature index Figure 4: Feature importance of HI, RI, BM, and CHR. 83 0.64 0.62 0.60 0.58 0.56 0.54 0.52 77 0.50 1357 9 11 13 15 13579 11 13 15 Time–scale Time–scale without CHR features without CHR features with CHR features with CHR features (a) (b) Figure 5: Accuracy and Kappa in di‡erent timescales. staging slices form a total of 14666 epochs, where Wake, 5.1. E“ect of Independent and Coordinated Features on Sleep REM, and NREM stages are 2668, 2562, and 9434 epochs, Staging. In order to examine the e‡ectiveness of the accounting for 18.19%, 17.47%, and 64.34%, respectively. extracted features using independent features of HI, RI, In this study, we adopt di‡erent weights assigned in [24] BM, and coordinated features between HI and RI, we †rst to avoid over†tting. Similar to [41], we quantify the evaluate the feature importance using RF classi†er [42]. performance in terms of accuracy and Cohen’s Kappa Figure 4 shows the contributions of both the independent coe¯cient, respectively, which are given by features (HI, RI, and BM) and the coordinated features (CHR) to sleep stage classi†cation. As shown in Figure 4, TP p − p class 0 e p (accuracy)  , Kappa  , the feature importance of HI class features, RI class T 1 − p class{Wake,NREM,REM} features, CHR class features, and BM class features ac- count for 25.4%, 23.2%, 32.6%, 18.8%, respectively. (3) Among them, ACD extracted by HI is the most important where T and TP denote the total number of epochs class feature, revealing that the cumulative di‡erence during and those correctly classi†ed into the corresponding sleep is particularly important for sleep staging. It is also class and p is the hypothetical probability of chance e noted that CHR features contribute most to sleep stage agreement. classi†cation, which demonstrates that coordination be- tween HI and RI is essential to discriminate sleep stages. TP + FP TP + FN class class class class p   × . Specially, Ratio of ACD to CHR performs better, which is T T class{Wake,NREM,REM} consistent with the excellent performance of ACD in the independent features. (4) Accuracy (%) Feature importance Kappa Journal of Healthcare Engineering 7 1.0 0.8 0.6 0.4 0.2 0.0 NREM REM Wake without sleep stage fusion with sleep stage fusion Figure 6: �e classi†cation performance with and without the proposed sleep stage fusion. Wake Wake REM REM NREM NREM 0 200 400 600 800 1000 0 200 400 600 800 1000 epoch epoch (a) (b) Wake REM NREM 0 200 400 600 800 1000 epoch (c) Figure 7: (a) Reference sleep stages provided by PSG. (b) Classi†ed sleep stages estimated without stage fusion. (c) Classi†ed sleep stages estimated with stage fusion. NREM 8466 531 437 REM 1068 1376 118 Wake 348 86 2234 NREM REM Wake Predicted class Figure 8: Confusion matrix of Wake-REM-NREM sleep stage classi†cation. True class Accuracy of three stages 8 Journal of Healthcare Engineering Table 5: Information of the subjects with sleep-disordered breathing. Subject ID Gender Age AHI (times/hour) Severity Total (night) Valid (night) A1 M 49 10.6 Mild 1 1 A2 F 70 6.3 Mild 2 1 A3 M 70 7.7 Mild 1 1 A4 F 69 8.1 Mild 1 1 A5 M 73 5.9 Mild 2 1 A6 M 50 6.7 Mild 1 1 A7 M 51 7.5 Mild 1 1 Table 6: Performance of the sleep-disordered breathing subjects. Subject ID Accuracy (%) Kappa 1 76.07 0.58 2 75.30 0.48 3 65.11 0.39 4 78.49 0.63 5 68.11 0.41 6 79.81 0.67 7 82.62 0.62 Average 75.07 0.54 Std 5.85 0.11 Wake REM NREM 0 200 400 600 800 epoch (a) Wake REM NREM 0 200 400 600 800 epoch (b) Wake REM NREM 0 200 400 600 800 epoch (c) Figure 9: (a) Reference sleep stages provided by PSG. (b) Classified sleep stages estimated without stage fusion. (c) Classified sleep stages estimated with stage fusion. Journal of Healthcare Engineering 9 with sleep-disordered breathing (most of these subjects *en, we analyze the effect of time resolution with respect to features extraction on sleep staging. As shown in Figure 5, both suffer from mild sleep apnea syndrome), aiming to verify the effectiveness of the proposed design. *e data used in the the accuracy and Kappa tend to increase when t< 10 s and then decrease as the timescale increases. A possible explanation to experiment was jointly recorded by *e First Affiliate this behavior is that non-contact-measured HI and RI suffers Hospital of Guangzhou Medical University and Guangzhou from inevitable errors due to artifact motion and noise, as SENVNV Co., and the experiment has obtained the consent shown in Figure 3. In this case, increasing the timescale yields a of the subjects, and personal private information is kept higher accuracy of HI and RI. Benefiting from the increase of confidential. *e information of the recruited subjects is timescale when t< 10 s, the improvement of feature accuracy listed in Table 5. improved sleep staging performance. As the timescale grows Using the proposed scheme with noncontact-measured larger (i.e., t> 10 s), it simultaneously reduces the sensitivity vital signs, the performance of sleep staging in comparison with PSG is shown in Table 6. It can be seen that the averaged with respect to the variation of HI and RI in different timescales, thus leading to a reduced performance. Furthermore, it can be accuracy and Kappa coefficient with respect to the recruited subjects suffering from sleep-disordered breathing are 75.07% seen that the classification results using CHR features perform better than those without CHR features, demonstrating the and 0.54, respectively. *is demonstrates the robustness of the effectiveness of the CHR features. *e highest accuracy and so-designed features and model. Figure 9 provides a typical Kappa are 82.42% and 0.63 when the timescale is 9 s. sleep stage classification from a subject with sleep-disordered breathing. 5.2. Sleep Stage Fusion. Next, we investigate the effect of 6. Conclusion stage fusion as the postprocessing of classification on sleep staging. Figure 6 shows the classification performance with *is paper studied feature-aided sleep stage classification using and without the proposed sleep stage fusion. It can be seen noncontact-measured vital signs. In addition to the analysis of that the applied stage fusion significantly improves the re- independent features such as HI, RI, and BM, which are sults, especially in Wake and REM. *e accuracy of Wake is characterized from BCG, respiratory rate, and body move- increased by 31.6%, and that of REM is increased by 39.9%. ments signals in different timescales, we validated through *e result is reasonable since the epoch-by-epoch classified experiments that the coordinated features between HI and RI sleep stages over a specific interval are mapped to an play an important role in sleep staging. In order to improve the identical sleep stage. Moreover, the accuracy of NREM still performance of classification, we developed a rule-based maintains a high level of 89.8%, although it may be reduced postprocessing to fuse the classified results of discrete time by 3.2% from 93.0% due to the slight fusion errors in the dimensions. *e experimental results in comparison with PSG fusion of Wake and REM. demonstrate the effectiveness of the proposed design. To further demonstrate the performance improvement of stage fusion, Figure 7 provides a typical example with and without Data Availability the proposed sleep stage fusion. As can be seen from Figure 7, the proposed stage fusion significantly improves the results. No datasets are generated during the current study. *e datasets analyzed during this work are publicly available. 5.3. Wake-REM-NREM Discrimination. We further verify Consent the confusion matrix based on RF classifier, and the result is shown in Figure 8. It can be seen from Figure 8 that the Informed consent was obtained from all subjects involved in prediction accuracy rates of Wake, REM, and NREM are the study. 67.3%, 53.7%, and 83.7%, respectively. Among them, the dominant error comes from the misclassification between Conflicts of Interest REM and NREM stages. Taking the experimental results into account, the main reason of low accuracy of recognition in *e authors declare that there are no conflicts of interest REM could be summarized to three aspects. (1) Frequent regarding the publication of this paper. body movements occur in both Wake and REM stages, leading to a lower accuracy of feature extraction in terms of Authors’ Contributions HI, RI, and CHR, thereby reducing the accuracy of classi- Z.W. and H.Z. carried out the data analysis and experiment fication. (2) *e proportion of REM epochs is lower than and wrote the manuscript. S.Z., B.Y., P.C., Z.P., and H.Z. that of NREM and Wake. *erefore, the proportion of supervised the methodology and organization of this work. misclassified epochs of REM in REM is higher than that in All authors have read and agreed to the published version of the whole sleep stage, whereas the accuracy of the whole the manuscript. system will be less affected. Acknowledgments 5.4. Validation on Subjects Suffering from Sleep-Disorderd Breathing. As a proof-of-concept with respect to the so- *is work was supported by the Natural Science Foundation designed sleep staging model, we further consider 7 subjects of Guangdong Province (no. 2022A1515010104), Blue Fire 10 Journal of Healthcare Engineering Medicine and Biology Society, pp. 2602–2605, Vancouver, Innovation Project of the Ministry of Education (Huizhou) Canada, October 2008. (no. CXZJHZ201803(, Special Funds for the Cultivation of [15] S. J. Redmond and C. Heneghan, “Cardiorespiratory-based Guangdong College Students’ Scientific and Technological sleep staging in subjects with obstructive sleep apnea,” IEEE Innovation (no. pdjh2021b0133), and also supported by Transactions on Biomedical Engineering, vol. 53, no. 3, Science and Technology Project of Guangzhou (nos. pp. 485–496, 2006. 202002030353 and 202102021114). [16] S. J. Redmond, P. de Chazal, C. O’Brien, S. Ryan, W. T. McNicholas, and C. Heneghan, “Sleep staging using cardiorespiratory signals,” Somnologie - Schlafforschung und References Schlafmedizin, vol. 11, no. 4, pp. 245–256, 2007. [17] S. Devot, R. Dratwa, and E. Naujokat, “Sleep/wake detection [1] Y. Castillo-Escario, H. Kumru, I. Ferrer-Lluis, J. Vidal, and R. Jane, ´ “Detection of sleep-disordered breathing in patients based on cardiorespiratory signals and actigraph,” in Pro- with spinal cord injury using a smartphone,” Sensors, vol. 21, ceedings of the 2010 Annual International Conference of the no. 21, p. 7182, 2021. IEEE Engineering in Medicine and Biology, pp. 5089–5092, [2] I. Ferrer-Lluis, Y. Castillo-Escario, J. M. Montserrat, and Buenos Aires, Argentina, November 2010. R. Jane, “Analysis of smartphone triaxial accelerometry for [18] T. Willemen, D. Van Deun, V. Verhaert et al., “An evaluation monitoring sleep-disordered breathing and sleep position at of cardiorespiratory and movement features with respect to home,” IEEE Access, vol. 8, pp. 71231–71244, 2020. sleep-stage classification,” IEEE Journal of Biomedical and [3] F. P. Cappuccio, L. D’Elia, P. Strazzullo, and M. A. Miller, Health Informatics, vol. 18, no. 2, pp. 661–669, 2014. “Sleep duration and all-cause mortality: a systematic review [19] D. Huysmans, I. Castro, P. Borzee ´ et al., “Capacitively-cou- and meta-analysis of prospective studies,” Sleep, vol. 33, no. 5, pled ECG and respiration for sleep-wake prediction and risk pp. 585–592, 2010. detection in sleep apnea patients,” Sensors, vol. 21, no. 19, [4] R. B. Berry, R. Budhiraja, D. J. Gottlieb et al., “Rules for p. 6409, 2021. scoring respiratory events in sleep: update of the 2007 AASM [20] M. Gaiduk, R. Seepold, T. Penzel, J. A. Ortega, M. Glos, and manual for the scoring of sleep and associated events,” Journal N. M. Madrid, “Recognition of sleep/wake states analyzing of Clinical Sleep Medicine, vol. 08, no. 05, pp. 597–619, 2012. heart rate, breathing and movement signals,” in Proceedings of [5] A. Rechtschaffen, “A manual for standardized terminology, the 2019 41st Annual International Conference of the IEEE techniques and scoring system for sleep stages in human Engineering in Medicine and Biology Society, pp. 5712–5715, subjects,” Brain Inf. Serv., 1968. EMBC), Berlin, Germany, October 2019. [6] B. Haex, Back and Bed: Ergonomic Aspects of Sleeping, CRC [21] M. Gaiduk, T. Penzel, J. A. Ortega, and R. Seepold, “Auto- Press, Boca Raton, FL, US, 2004. matic sleep stages classification using respiratory, heart rate [7] K. Adam and I. Oswald, “Sleep helps healing,” BMJ, vol. 289, and movement signals,” Physiological Measurement, vol. 39, no. 6456, pp. 1400-1401, Article ID 6437572, 1984. no. 12, Article ID 124008, 2018. [8] P. Meerlo, R. E. Mistlberger, B. L. Jacobs, H. Craig HellerCraig [22] J. Siyanbade, B. Abdulrazak, and I. Sadek, “Unobtrusive Heller, and D. McGinty, “New neurons in the adult brain: the monitoring of sleep cycles: a technical review,” Bio- role of sleep and consequences of sleep loss,” Sleep Medicine MedInformatics, vol. 2, no. 1, pp. 204–216, 2022. Reviews, vol. 13, no. 3, pp. 187–194, Article ID 18848476, 2009. [23] M. Migliorini, A. M. Bianchi, D. Nistico` et al., “Automatic [9] Y. R. Chien, C. H. Wu, and H. W. Tsao, “Automatic sleep- sleep staging based on ballistocardiographic signals recorded arousal detection with single-lead EEG using stacking en- through bed sensors,” in Proceedings of the 2010 Annual semble learning,” Sensors, vol. 21, no. 18, p. 6049, Article ID International Conference of the IEEE Engineering in Medicine 34577255, 2021. and Biology, pp. 3273–3276, Buenos Aires, Argentina, No- [10] T. Penzel and R. Conradt, “Computer based sleep recording vember 2010. and analysis,” Sleep Medicine Reviews, vol. 4, no. 2, [24] S. H. Choi, H. B. Kwon, H. W. Jin et al., “Long short-term pp. 131–148, 2000. memory networks for unconstrained sleep stage classification [11] A. Domingues, T. Paiva, and J. M. Sanches, “Sleep and using polyvinylidene fluoride film sensor,” IEEE Journal of wakefulness state detection in nocturnal actigraphy based on Biomedical and Health Informatics, vol. 24, no. 12, movement information,” IEEE Transactions on Biomedical pp. 3606–3615, 2020. Engineering, vol. 61, no. 2, pp. 426–434, Article ID 24013826, [25] J. Drzazga and B. Cyganek, “An LSTM network for apnea and 2014. hypopnea episodes detection in respiratory signals,” Sensors, [12] J. Tilmanne, J. Urbain, M. V. Kothare, A. V. Wouwer, and vol. 21, no. 17, p. 5858, 2021. S. V. Kothare, “Algorithms for sleep-wake identification using [26] H. Hong, L. Zhang, C. Gu, Y. Li, G. Zhou, and X. Zhu, actigraphy: a comparative study and new results,” Journal of “Noncontact sleep stage estimation using a CW Doppler Sleep Research, vol. 18, no. 1, pp. 85–98, Article ID 19250177, radar,” IEEE Journal on Emerging and Selected Topics in 2009. Circuits and Systems, vol. 8, no. 2, pp. 260–270, 2018. [13] A. Tataraidze, L. Anishchenko, L. Korostovtseva, B. J. Kooij, [27] T. Lauteslager, S. Kampakis, A. J. Williams, M. Maslik, and M. Bochkarev, and Y. Sviryaev, “Sleep stage classification F. Siddiqui, “Performance evaluation of the circadia con- based on respiratory signal,” in Proceedings of the 2015 37th tactless breathing monitor and sleep analysis algorithm for Annual International Conference of the IEEE Engineering in sleep stage classification,” in Proceedings of the Annual In- Medicine and Biology Society (EMBC), pp. 358–361, IEEE, ternational Conference of the IEEE Engineering in Medicine Milan, Italy, August 2015. and Biology Society, pp. 5150–5153, Montreal, ´ QC, Canada, [14] S. Canisius, T. Ploch, V. Gross, A. Jerrentrup, T. Penzel, and July 2020. K. Kesper, “Detection of sleep disordered breathing by au- [28] F. Zhang, C. Wu, B. Wang et al., “SMARS: sleep monitoring tomated ECG analysis,” in Proceedings of the 2008 30th via ambient radio signals,” IEEE Transactions on Mobile Annual International Conference of the IEEE Engineering in Computing, vol. 20, no. 1, pp. 217–231, 2021. Journal of Healthcare Engineering 11 [29] J. Liang, J. Huang, L. Mu et al., “An effective algorithm for beat-to-beat heart rate monitoring from ballistocardiograms,” Journal of Medical Imaging and Health Informatics, vol. 10, no. 3, pp. 633–640, 2020. [30] I. Sadek, E. Seet, J. Biswas, B. Abdulrazak, and M. Mokhtari, “Nonintrusive vital signs monitoring for sleep apnea patients: a preliminary study,” IEEE Access, vol. 6, pp. 2506–2514, 2017. [31] S. M. Proll, ¨ S. Hofbauer, C. Kolbitsch, R. Schubert, and K. D. Fritscher, “Ejection wave segmentation for contact-free heart rate estimation from ballistocardiographic signals,” in Proceedings of the 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 3571–3576, EMBC), Berlin, Germany, October 2019. [32] J. Gilles, “Empirical wavelet transform,” IEEE Transactions on Signal Processing, vol. 61, no. 16, pp. 3999–4010, 2013. [33] M. Ali, A. AhmadAhmad, S. SultanaSultana, S. R. MirMir, and M. Sawan, “Chemical constituents from the seed husks of Oryza sativa L,” Natural Product Research, vol. 21, pp. 14569–14586, 2021. [34] R. J. *omas, J. E. Mietus, C. K. Peng, and A. L. Goldberger, “An electrocardiogram-based technique to assess cardiopul- monary coupling during sleep,” Sleep, vol. 28, no. 9, pp. 1151–1161, 2005. [35] M. Kagawa, N. Sasaki, K. Suzumura, and T. Matsui, “Sleep stage classification by body movement index and respiratory interval indices using multiple radar sensors,” in Proceedings of the 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 7606–7609, EMBC), Milan, Italy, November 2015. [36] A. Liaw and M. Wiener, “Classification and regression by randomForest,” R News, vol. 2, pp. 18–22, 2002. [37] C. C. Chang and C. J. Lin, “Libsvm,” ACM Transactions on Intelligent Systems and Technology, vol. 2, no. 3, pp. 1–27, 2011. [38] K. Polat and S. Gune ¨ ¸, s “Classification of epileptiform EEG using a hybrid system based on decision tree classifier and fast Fourier transform,” Applied Mathematics and Computation, vol. 187, no. 2, pp. 1017–1026, 2007. [39] T. K. Ho, “Nearest neighbors in random subspaces,” in Joint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pat- tern Recognition (SSPR), pp. 640–648, Springer, Berlin, Heidelberg, 1998. [40] Z. H. Zhou, Ensemble Methods: Foundations and Algorithms, CRC Press, Boca Raton, FL, US, 2012. [41] A. J. Viera and J. M. Garrett, “Understanding interobserver agreement: the kappa statistic,” Family Medicine, vol. 37, no. 5, pp. 360–363, 2005. [42] B. Xue, B. Deng, H. Hong, Z. Wang, X. Zhu, and D. D. Feng, “Non-contact sleep stage detection using canonical correla- tion analysis of respiratory sound,” IEEE Journal of Bio- medical and Health Informatics, vol. 24, no. 2, pp. 614–625,

Journal

Journal of Healthcare Engineering – Hindawi Publishing Corporation

Published: Jul 8, 2022

There are no references for this article.