Sleep Monitoring: A Comparison Between Three Wearable Instruments

Transcription

1 MILITARY MEDICINE, 176, 7:811, 2011 Sleep Monitoring: A Comparison Between Three Wearable Instruments Nelleke C. van Wouwe, PhD * ; Pierre J. L. Valk, MSc * ; Bertil J. Veenstra, MSc ABSTRACT During military operations soldiers often encounter extreme environmental circumstances like heat, cold, prolonged physical exercise, and disturbed sleep, which hamper their performance. Monitoring changes in physiological parameters may assist with adequate interventions to prevent the negative consequences and support recovery. The current study was employed to reduce the number of measurement instruments to monitor physiological variables, especially with respect to adequate sleep prediction. We compared three instruments with respect to their effectiveness in predicting sleep; the Equivital, Sensewear, and Actiwatch. Additionally, we investigated the added value of cardio respiratory to accelerometer signals to estimate sleep duration. The Equivital model (based on acceleration data) and Sensewear predict sleep and wake as accurate as the commonly used Actiwatch model, and the cardio respiratory Equivital data further improve accuracy and specificity. In sum, the current study provides an indication that the Equivital system (or any other chestband that measures 3-dimensional acceleration plus other physiological variables) might be interchanged with an Actiwatch for sleep prediction. INTRODUCTION A military environment and the conditions soldiers are exposed to often result in extreme mental and physical stress that negatively affect the soldiers performance. Monitoring the changes in physiological parameters like heart rate variability (HRV), respiration, and acceleration that correlate with a reduction in performance enables anticipation and intervention to postpone or reduce performance decrements. One of the stressors often encountered in a military environments is sleep deprivation or reduced sleep quality, 1 4 which have been shown to negatively affect cognitive performance, i.e., memory, attention, and alertness 5 8 and physical recovery and resilience. 9,10 Adequate sleep monitoring may facilitate interventions to prevent these negative consequences or provide indications that sleep recovery is necessary. The accepted standard to predict sleep quality and duration is polysomnography. 11 Outside the laboratory, this method is less feasible because it is expensive, uncomfortable, and time consuming. Wrist actigraphy is often used for ambulatory sleep monitoring; it is a relatively easy and inexpensive method to predict sleep duration on the basis of wrist movements 12 and a convenient method to implement in a field study. However, this instrument is limited to the measurement of accelerations and unable to measure additional physiological parameters like HRV or respiration. Moreover, actigraphy is exclusively based on acceleration measurements, which occasionally leads to misclassification for activities with low movement like reading or watching television. * Department of Human Performance, TNO Defence, Security and Safety, Soesterberg, The Netherlands. TGTF, Royal Netherlands Army, Ministry of Defence, Utrecht, The Netherlands. Additionally, different instruments and algorithms have been applied to measure and transform raw acceleration data into meaningful units that can be used for sleep classification that reduces the comparability between these studies and sometimes hampers the transfer of a specific sleep algorithm to other measurement instruments. For example, an actigraph provides output in counts, instead of raw acceleration data (g), and is thereby limited to a sleep algorithm that processes counts. Alternatively, other physiological variables might inform us about changes in the activity of the autonomic nervous system during sleep/wake transitions, 16 like HRV or respiration. Recently, cardio respiratory signals (separately or combined with acceleration data) have been used for sleep/wake classification in ambulatory and controlled lab studies Compared to actigraphy, some of the algorithms including both acceleration and cardio respiratory signals showed relatively high classification accuracy (i.e., 96% vis-à-vis 80% in models using actigraphy 16 ), although the sample size of these studies was small. 17,18 HRV is linked with the autonomic nervous system which varies throughout the night with different sleep stages. 23 However, the way of calculating and interpreting HRV results differentiates24 and, therefore, complicates the use of HRV for sleep classification. HRV is commonly analyzed by a power spectral analysis which is bound to methodological constraints and time consuming; electrocardiogram (ECG) signals have to be offline selected and corrections for respiratory variations are necessary. Recent methodological developments suggest that the root mean square of successive interbeat interval differences (RMSSD) of the ECG is a computationally more efficient measure of parasympathetic activity. 25 The current study was employed to investigate whether we can reduce the number of physiological measurement instruments and at the same time improve sleep wake classification compared to predictions exclusively based on acceleration. MILITARY MEDICINE, Vol. 176, July

2 First, we will test whether the accelerometer data of three wearable instruments, the Equivital system, 26 the Sensewear,27 and the Actiwatch, 28 are equally predictive to distinguish between sleep and wake. Additionally, the added value of cardio respiratory signals to classify between sleep and wake will be tested. We will generate a sleep classification model based on the Equivital acceleration and cardio respiratory data and compare the accuracy rates with other models. Subjective data from a sleep diary, markers given by the subjects on either Sensewear or Actiwatch, and specified positive and negative sleep moments during the night (determined by the sleep protocol) were incorporated to classify a period as asleep or awake. METHOD Participants Eleven subjects participated in this study, mean age 36.7 years (SD = 10.7). All subjects participated voluntarily and gave their written-informed consent before participation, as part of procedures that complied fully with relevant laws and with standards of ethical conduct in human research as regulated by the Medical Ethical Committee of the Netherlands Organization for Applied Research (TNO), Human Factors. Procedure Subjects were asked to wear three recording devices for two nights; the Equivital chestband, 26 the Sensewear Armband, 27 and an Actiwatch wristband. 28 According to the sleep protocol, recording started one hour before bedtime. We used different protocols for each of the two nights. One protocol included a resting period of 15 minutes (negative sleep sample) before the subject intended to sleep (lights off). The second protocol asked subjects to set an alarm clock for three hours after lights off. The order of both protocols was counterbalanced across subjects. Subjects were asked to mark wakeful periods throughout both nights and fill out a sleep diary subsequent to each of the measurement nights. Measurement Instruments Actigraph The Actiwatch28 is a watch-like instrument which is placed around the nondominant wrist. It contains a 3-dimensional (3D) piezoelectric accelerometer that records the number of times that the device accelerates by an amount greater than a certain threshold setting. The maximum sampling frequency is 32 Hz. These counts are accumulated into a 1-minute sampling interval and stored in the actigraph memory. Equivital An Equivital physiological monitoring system 26 was used as a recording device for accelerometer and cardio respiratory signals. It measures a two-lead ECG, respiration effort (by means of impedance and belt), 3D acceleration, and skin temperature. The sensors are integrated in a belt worn around the upper chest area. The current study used acceleration, respiration (belt), and ECG signals. Acceleration and respiration are sampled at 25.6 Hz and ECG at 256 Hz. Sensewear The Bodymedia Sensewear Armband 27 was worn around the upper right arm. It includes a 2-axis accelerometer, heat flux sensor, galvanic skin response sensor, skin temperature sensor, and a near-body ambient temperature sensor. The SenseWear sensors acquire physiological data at 32 Hz. Sleep Diary Subjects filled out a sleep diary consisting of questions about their wake up and bedtime, sleep quality, naps during the day, and wakeful periods throughout the night. Analysis Preprocessing Acceleration We applied a 0.5- to 11-Hz filter on the Equivital accelerometer signal to remove slow gravitational movements and excessive accelerations. Additionally, we calculated the power spectral density estimate of the acceleration signal in the 0.5- to 3-Hz band (power spectral density acceleration = PSD ACC) 29,30 with a segment size of 3 minutes. Acceleration signals measured by the Sensewear Armband and Actiwatch wristband were preprocessed internally and subsequently analyzed by the software or sleep algorithm provided by the manufacturer. Actigraphic counts were classified according to the formula stated by the Actiwatch manual 31 : Activity in the current sample (1 minute) was determined by the sum of the current activity and activity in surrounding samples (up to 3 minutes). Activity within 1 minute of the scored sample was reduced by a factor of 5.Within 2 minutes of the scored sample; activity was reduced by a factor 25. A total score of 40 or higher was scored as awake. The data produced by the Sensewear were processed by the Sensewear software (according to a currently unspecified algorithm, 27 ) and provided a minute-to-minute sleep classifi cation. Heart Rate and Respiratory Signals The RMSSD ECG signal was calculated by means of Matlab algorithms, separately for each 3-minute segment. Deviant interbeat (R-R) intervals (i.e., above 4 seconds or below 0.25 seconds) were excluded from the analysis. Moreover, we calculated the power spectral density estimate of the raw respiratory signals (0.07- to 0.5-Hz frequency band). Subsequently we determined the ratio between high and low respiratory frequency bands (high frequency power/[high frequency power + low frequency power]), which was used as input for the sleep prediction model (power spectral density respiration = PSD RESP). High frequency power was measured within to 0.5-Hz frequency band and low frequency power was measured within to 0.15-Hz frequency band. 812 MILITARY MEDICINE, Vol. 176, July 2011

3 Sleep Wake Scoring Positive examples of sleep and wake were chosen by means of the subjective diary and the sleep protocol. We applied the following criteria to score sleep or wake and to select a balanced amount of sleep and wake samples: (1) If subjects marked a minute as awake or when the alarm was set, 4 minutes subsequent to the marker were scored as awake. Additionally, 15 minutes before lights off and 15 minutes resting in bed were scored as awake. (2) A subject was considered to be asleep when neither the markers nor the information in the sleep diary indicated that the subject was awake. To ensure that different stages of sleep would be represented in our sample, we selected four periods of sleep for every subject: 60 to 75 minutes after lights off, 30 to 45 minutes after the alarm was set, 180 to 195 minutes after lights off, and 120 to 135 minutes before lights on. This selection adds up to 60% of sleep and 40% wake samples for all subjects. Sleep Algorithm Development and Statistical Analysis We first used the acceleration data to model sleep (measured in mg on x, y, z axes). This provided us with an indication whether the accelerometer of the Equivital system, which measures chest instead of wrist movements and contains position information from each axis, might be a useful sleep classifier. With more generic acceleration data (measured in g) applicable for sleep classification, future studies are not bound to the Actiwatch as a measurement instrument. Additionally, we tested whether cardio respiratory variables (RMSSD for ECG signals and logarithm of the power spectral density for respiration signals) would increase predictability of the model. These variables might be especially useful to distinguish between low activity wake patterns and sleep. A sleep scoring model using the Equivital output (PSD ACC, RMSSD ECG, and PSD RESP) was generated by a generalized estimating equations regression model, 32 based on a data set of six participants (training data), which contained positive and negative examples of sleep (i.e., reading or watching TV in bed). A generalized estimating equations regression analysis considers the measurements within participants as repeated measurements and accounts for this dependency. Explanatory variables were added stepwise into the model and subsequently compared with respect to significant beta coefficients, explained variance (R 2 ), sensitivity and specificity. Explanatory variables included the natural logarithm of: RMSSD ECG, PSD RESP, and PSD ACC. Sleep (wake, sleep) was included as a dependent variable. To begin with, data from the five remaining participants were used to test the model. To evaluate the results of the model we determined sensitivity, specificity, and accuracy relative to the subjective markers and sleep diary. (1) Sensitivity = true positives/(true positives + false negatives). (2) Specificity = true negatives/(true negatives + false positives). (3) Accuracy = true positives + true negatives/(true and false positives+ true and false negatives). Next to that, the outcomes of the model with the highest explanatory variance were compared with the Actiwatch and Sensewear predictions. Sensitivity, specificity, and accuracy were calculated for the sleep and wake classifications generated by the algorithms from the Actiwatch Sleep Analysis software and Sensewear software. RESULTS Table I shows the added value of the cardio respiratory variables next to acceleration, in terms of explained variance. Model development was based on a training set of six subjects wearing the devices for two nights. Figure 1 shows the sensitivity and specificity for each combination of predictors as in Table I. Table II presents sensitivity, specificity, and accuracy for each of the regression models based on the Equivital data. Table III shows sensitivity, specificity, and accuracy for each instrument and sleep prediction model. Formula 1 predicts sleep based on acceleration (PSD ACC) and HRV (RMSSD ECG). Sleep values 1.5 indicate sleep, whereas values >1.5 indicate wake. Acceleration and HRV turned out to be significant predictors of sleep, whereas respiration (PSD RESP) was not a significant predictor, see Table I and Formula 1. The model based on acceleration and ECG also shows a large area under the receiver operating characteristic (ROC) curve (0.89), see Figure 1, although the other models show almost similar results; i.e., 0.88 (PSD ACC), 0.89 (combined model of PSD ACC, RMSSD ECG, and PSD RESP), and 0.89 (PSD ACC, PSD RESP) TABLE I. Beta Coefficients and R 2 With Acceleration, ECG and Respiration Included Stepwise in the Model as Sleep Predictors Model Dependent Variable Constant PSD ACC b 1 RMSSD ECG b 2 PSD RESP b 3 R 2 1 Sleep Sleep Sleep Sleep Significant b-coefficients (p < 0.01) are presented in bold. Model based on training data (6 subjects measured 2 nights). MILITARY MEDICINE, Vol. 176, July

4 FIGURE 1. Receiver operation characteristics curves describing the sensitivity and 1-specificity of each combination of predictors; acceleration, ECG, and respiration (training data). TABLE II. Predictions of Both Models Applied on Test Data (Remaining 5 Subjects Measured 2 Nights) Model Variables Sensitivity Specificity Accuracy 1 ACC ACC, ECG TABLE III. Predictions of Both Equivital Models a and Other Measurement Instruments Calculated for 11 Subjects Model Variables Sensitivity Specificity Accuracy 1 Equivital ACC Equivital ACC, ECG Actiwatch Sensewear a Model based on 6 subjects, see Table I. Formula 1. Sleep = (015 PSD ACC ) + (-0.08 RMSSD ECG ) The highest accuracy, sensitivity, and specificity was obtained when both acceleration and HRV were included in the model, see Table II. Sensitivity obtained from the Actiwatch predictions was higher compared with sensitivity of the other models, whereas specificity and accuracy were better with Equivital and Sensewear. DISCUSSION During military operations soldiers often encounter extreme environmental circumstances like heat, cold, prolonged physical exercise, and disturbed sleep, which hamper their physical and cognitive performance Monitoring changes in physiological parameters may assist with adequate interventions to prevent the negative consequences or provide an indication for recovery. The current study was employed to reduce the number of measurement instruments to monitor physiological variables, especially with respect to adequate sleep prediction. We compared three wearable instruments with respect to their effectiveness in predicting sleep. Additionally, we investigated the additional value of cardio respiratory signals compared to accelerometer signals to estimate sleep. On the basis of the acceleration signals from the current study, the Equivital model detects sleep and wake and the commonly used Actiwatch model. The ECG signals (RMSSD) 814 MILITARY MEDICINE, Vol. 176, July 2011

5 further enhanced the specificity of the predictions (correct wake detection) to 63%. Actigraphy methods like Sadeh and Cole show specificity rates of 44% and 34%. 33 Therefore, in future ambulatory studies interested in sleep estimation as well as predictions based on other physiological variables, the current study provides an indication that the Equivital system (or any other chestband that measures 3D acceleration plus other physiological variables) might be interchanged with an Actiwatch for sleep prediction. This would reduce the number of measurement instruments in ambulatory studies and the number of different data preprocessing steps. However, the current study was based on a small sample set; therefore, similar studies with a larger sample size should be employed to strengthen our conclusions. Next to that, the reliability of the sleep algorithm based on acceleration and ECG has to be established in future studies as well. Additionally, the external validity of the current algorithm remains to be tested in order to generalize the results of this study (which was based on a healthy population with regular sleep schedules) to a larger population, for example, in groups with diminished sleep efficiency, groups who have an irregular sleep schedule or patient groups with different ECG or behavioral patterns. Sleep predictions by the Sensewear Software revealed lower sensitivity than the Actiwatch and Equivital models, but better specificity and comparable accuracy. A practical disadvantage of the Sensewear is its battery life which does not support data collection at a high sample rate throughout the night (i.e., with longer sampling periods, the Sensewear only saves 1 minute averages of the original 32-Hz sampling rate). This limits the analysis of psychophysiological signals. The Equivital acceleration-ecg model built in the current study was based on sleep and wake samples selected from a subjective sleep diary and provided accuracy rates comparable with other algorithms and studies using actigraphy, 15 i.e., around 80% accuracy. However, controlled laboratory studies 16,17 recently generated sleep/wake prediction algorithms based on cardio respiratory and acceleration with accuracy rates from 80% to 96%. Thus, a controlled laboratory study with EEG or electrooculography (EOG) eye movement measurements would enable more specific differentiation between sleep or wake and improve the accuracy of the current model. The model developed in this study enables sleep prediction based on acceleration signals or on acceleration combined with ECG signals and, thereby, facilitates flexible use of available data. Incorporation of computational methods like Neural Networks into measurement systems also seem promising; Neural Networks might facilitate individually based sleep prediction algorithms because feedback loops in the network enable a continuously learning system. Integration with wireless monitoring systems further improves online monitoring of physiological changes. ACKNOWLEDGMENT The work was funded by the Dutch Ministry of Defence. REFERENCES 1. Fry RW, Grove JR, Morton AR, Zeroni PM, Gaudieri S, Keast D : Psychological and immunological correlates of acute overtraining. Br J Sports Med 1994 ; 28 (4) : Jürimäe J, Mäestu J, Purge P, Jürimäe T : Changes in stress and recovery after heavy training in rowers. J Sci Med Sport 2004 ; 7 (3) : Lieberman HR, Bathalon GP, Falco CM, Morgan CA, Niro PJ, Tharion WJ : The fog of war: decrements in cognitive performance and mood associated with combat-like stress. Aviat Space Environ Med 2005 ; 76 (7): C7 C Booth CK, Probert B, Forbes-Ewan C, Coad RA : Australian army recruits in training display symptoms of overtraining. Mil Med 2006 ; 171 (11) : Chee M, Chuah L : Functional neuroimaging insights into how sleep and sleep deprivation affect memory and cognition. Curr Opin Neurol 2008 ; 21 (4): Dinges DF, Pack F, Williams K, et al : Cumulative sleepiness, mood disturbance, and psychomotor vigilance performance decrements during a week of sleep restricted to 4 5 hours per night. Sleep 1997 ; 20 (4) : Killgore WDS, Grugle NL, Reichardt RM, Killgore DB, Balkin TJ : Executive functions and the ability to sustain vigilance during sleep loss. Aviat Space Environ Med 2009 ; 80 (2): Valk PJL, Simons M : Pros and cons of strategic napping on long haul flights. AGARD-CP-599; NATO-AGARD, Neuilly-sur-Seine, France, pp 5/1 5/5. 9. Belenky G, Wesensten NJ, Thorne DR, et al : Patterns of performance degradation and restoration during sleep restriction and subsequent recovery: a sleep dose-response study. J Sleep Res 2003 ; 12 (1) : Mignot E : Why we sleep: the temporal organization of recovery. PLoS Biol 2008, 6 (4) : e Rechtschaffen A, Kales A (editors): A manual of standardized terminology, techniques and scoring system for sleep stages in human subjects. Washington, DC, U.S. Government Printing Office, Littner M, Hirshkowitz M, Kramer M, et al : Practice parameters for using polysomnography to evaluate insomnia: an update. Sleep 2003 ; 26 (6): Ancoli-Israel S, Cole R, Alessi C, Chambers M, Moorcroft W, Pollak C : The role of actigraphy in the study of sleep and circadian rhythms. Sleep 2003 ; 26 (3): Gorny SW, Spiro JR : Comparing different methodologies used in wrist actigraphy. Sleep Rev 2001 ; Tilmanne J, Urbain J, Kothare MV, Wouwer AV, Kothare SV : Algorithms for sleep-wake identification using actigraphy: a comparative study and new results. J Sleep Res 2009 ; 18 (1) : Ogilvie RD : The process of falling asleep. Sleep Med Rev 2001 ; 5 (3) : Karlen W, Mattiussi C, Floreano D : Adaptive sleep/wake classification based on cardiorespiratory signals for wearable devices. In Proc. IEEE Biomedical Circuits and Systems Conference BIOCAS 2007 ; Karlen W, Mattiussi C, Floreano D : Improving actigraph sleep/wake classification with cardio-respiratory signals. In 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2008 ; Trinder J, Beveren JAV, Smith P, Kleiman J, Kay A : Correlation between ventilation and EEG-defined arousal during sleep onset in young subjects. J Appl Physiol 1997 ; 83 (6) : Burr RL : Interpretation of normalized spectral heart rate variability indices in sleep research: a critical review. Sleep 2007 ; 30 (7) : Okamoto-Mizuno K, Yamashiro Y, Tanaka H, et al : Heart rate variability and body temperature during the sleep onset period. Sleep Biol Rhythms 2008 ; 6 (1) : MILITARY MEDICINE, Vol. 176, July

6 22. Redmond S, Heneghan C : Cardiorespiratory-based sleep staging in subjects with obstructive sleep apnea. IEEE Trans Biomed Eng 2006 ; 53 (3) : Bonnet MH, Arand DL : Heart rate variability: sleep stage, time of night, and arousal influences. Electroencephalogr Clin Neurophysiol 1997 ; 102 (5) : Malik M : Heart rate variability: standards of measurement, physiological interpretation, and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing Electrophysiology Circulation 1996 ; 93 (5), Goedhart AD, Van der Sluis S, Houtveen JH, Willemsen G, De Geus EJC : Comparison of time and frequency domain measures of RSA in ambulatory recordings. Psychophysiology 2007 ; 44 (2) : Equivital physiological monitoring system. Available at equivital.co.uk ; accessed March 16, Sensewear Bodymedia. Available at ; accessed March 16, ActiWatch (Mini Mitter, Sunriver, Oregon). Available at imitter.com/; accessed March 16, Karlen W, Mattiussi C, Floreano D. Sleep and wake classification with ECG and respiratory effort signals. IEEE Transactions on Biomedical Circuits and Systems 2009 ; 3 (2) : Karlen W : Adaptive wake and sleep detection for wearable systems. Ecole Polytechnique Fédérale de Lausanne (EPFL), Thèse no. 4391, CamNtech, Cambridge. Available at ; accessed March 16, Liang KY, Zeger SL : Regression analysis for correlated data. Annu Rev Public Health 1993 ; 14: De Souza L, Benedito-Silva AA, Pires ML, Poyares D, Tufik S, Calil HM : Further validation of actigraphy for sleep studies. Sleep 2003 ; 26 (1) : MILITARY MEDICINE, Vol. 176, July 2011