Sleep classification from wrist-worn accelerometer data using random forests

A Machine Learning Model for Predicting Sleep and Wakefulness Based on Accelerometry, Skin Temperature and Contextual Information

Purpose

Body-worn accelerometers are commonly used to estimate sleep duration in population-based studies. However, since accelerometry-based sleep/wake-scoring relies on detecting body movements, the prediction of sleep duration remains a challenge. The aim was to develop and evaluate the performance of a machine learning (ML) model to predict accelerometry-based sleep duration and to explore if this prediction can be improved by adding skin temperature data, circadian rhythm based on the estimated midpoint of sleep, and cyclic time features to the model.

Patients and Methods

Twenty-nine adults (17 females), mean (SD) age 40.2 (15.0) years (range 17-70) participated in the study. Overnight polysomnography (PSG) was recorded in a sleep laboratory or at home along with body movement by two accelerometers with an embedded skin temperature sensor (AX3, Axivity, UK) positioned at the low back and thigh. The PSG scoring of sleep/wake was used as ground truth for training the ML model.

Results

Based on pure accelerometer data input to the ML model, the specificity and sensitivity for predicting sleep/wake was 0.52 (SD 0.24) and 0.95 (SD 0.03), respectively. Adding skin temperature data and contextual information to the ML model improved the specificity to 0.72 (SD 0.20), while sensitivity remained unchanged at 0.95 (SD 0.05). Correspondingly, sleep overestimation was reduced from 54 min (228 min, limits of agreement range [LoAR]) to 19 min (154 min LoAR).

Conclusion

An ML model can predict sleep/wake periods with excellent sensitivity and moderate specificity based on a dual-accelerometer set-up when adding skin temperature data and contextual information to the model.

Self-supervised learning of accelerometer data provides new insights for sleep and its association with mortality

Sleep is essential to life. Accurate measurement and classification of sleep/wake and sleep stages is important in clinical studies for sleep disorder diagnoses and in the interpretation of data from consumer devices for monitoring physical and mental well-being. Existing non-polysomnography sleep classification techniques mainly rely on heuristic methods developed in relatively small cohorts. Thus, we aimed to establish the accuracy of wrist-worn accelerometers for sleep stage classification and subsequently describe the association between sleep duration and efficiency (proportion of total time asleep when in bed) with mortality outcomes. We developed a self-supervised deep neural network for sleep stage classification using concurrent laboratory-based polysomnography and accelerometry. After exclusion, 1448 participant nights of data were used for training. The difference between polysomnography and the model classifications on the external validation was 34.7 min (95% limits of agreement (LoA): -37.8-107.2 min) for total sleep duration, 2.6 min for REM duration (95% LoA: -68.4-73.4 min) and 32.1 min (95% LoA: -54.4-118.5 min) for NREM duration. The sleep classifier was deployed in the UK Biobank with 100,000 participants to study the association of sleep duration and sleep efficiency with all-cause mortality. Among 66,214 UK Biobank participants, 1642 mortality events were observed. Short sleepers (<6 h) had a higher risk of mortality compared to participants with normal sleep duration of 6-7.9 h, regardless of whether they had low sleep efficiency (Hazard ratios (HRs): 1.58; 95% confidence intervals (CIs): 1.19-2.11) or high sleep efficiency (HRs: 1.45; 95% CIs: 1.16-1.81). Deep-learning-based sleep classification using accelerometers has a fair to moderate agreement with polysomnography. Our findings suggest that having short overnight sleep confers mortality risk irrespective of sleep continuity.

State of the science and recommendations for using wearable technology in sleep and circadian research

Wearable sleep-tracking technology is of growing use in the sleep and circadian fields, including for applications across other disciplines, inclusive of a variety of disease states. Patients increasingly present sleep data derived from their wearable devices to their providers and the ever-increasing availability of commercial devices and new-generation research/clinical tools has led to the wide adoption of wearables in research, which has become even more relevant given the discontinuation of the Philips Respironics Actiwatch. Standards for evaluating the performance of wearable sleep-tracking devices have been introduced and the available evidence suggests that consumer-grade devices exceed the performance of traditional actigraphy in assessing sleep as defined by polysomnogram. However, clear limitations exist, for example, the misclassification of wakefulness during the sleep period, problems with sleep tracking outside of the main sleep bout or nighttime period, artifacts, and unclear translation of performance to individuals with certain characteristics or comorbidities. This is of particular relevance when person-specific factors (like skin color or obesity) negatively impact sensor performance with the potential downstream impact of augmenting already existing healthcare disparities. However, wearable sleep-tracking technology holds great promise for our field, given features distinct from traditional actigraphy such as measurement of autonomic parameters, estimation of circadian features, and the potential to integrate other self-reported, objective, and passively recorded health indicators. Scientists face numerous decision points and barriers when incorporating traditional actigraphy, consumer-grade multi-sensor devices, or contemporary research/clinical-grade sleep trackers into their research. Considerations include wearable device capabilities and performance, target population and goals of the study, wearable device outputs and availability of raw and aggregate data, and data extraction, processing, and analysis. Given the difficulties in the implementation and utilization of wearable sleep-tracking technology in real-world research and clinical settings, the following State of the Science review requested by the Sleep Research Society aims to address the following questions. What data can wearable sleep-tracking devices provide? How accurate are these data? What should be taken into account when incorporating wearable sleep-tracking devices into research? These outstanding questions and surrounding considerations motivated this work, outlining practical recommendations for using wearable technology in sleep and circadian research.

Associations between actigraphy-measured sleep duration, continuity, and timing with mortality in the UK Biobank

STUDY OBJECTIVES

To examine the associations between sleep duration, continuity, timing, and mortality using actigraphy among adults.

METHODS

Data were from a cohort of 88 282 adults (40-69 years) in UK Biobank that wore a wrist-worn triaxial accelerometer for 7 days. Actigraphy data were processed to generate estimates of sleep duration and other sleep characteristics including wake after sleep onset (WASO), number of 5-minute awakenings, and midpoint for sleep onset/wake-up and the least active 5 hours (L5). Data were linked to mortality outcomes with follow-up to October 31, 2021. We implemented Cox models (hazard ratio, confidence intervals [HR, 95% CI]) to quantify sleep associations with mortality. Models were adjusted for demographics, lifestyle factors, and medical conditions.

RESULTS

Over an average of 6.8 years 2973 deaths occurred (1700 cancer, 586 CVD deaths). Overall sleep duration was significantly associated with risk for all-cause (p < 0.01), cancer (p < 0.01), and CVD (p = 0.03) mortality. For example, when compared to sleep durations of 7.0 hrs/d, durations of 5 hrs/d were associated with a 29% higher risk for all-cause mortality (HR: 1.29 [1.09, 1.52]). WASO and number of awakenings were not associated with mortality. Individuals with L5 early or late midpoints (<2:30 or ≥ 3:30) had a ~20% higher risk for all-cause mortality, compared to those with intermediate L5 midpoints (3:00-3:29; p ≤ 0.01; e.g. HR ≥ 3:30: 1.19 [1.07, 1.32]).

CONCLUSIONS

Shorter sleep duration and both early and late sleep timing were associated with a higher mortality risk. These findings reinforce the importance of public health efforts to promote healthy sleep patterns in adults.

A systematic review of the performance of actigraphy in measuring sleep stages

The accuracy of actigraphy for sleep staging is assumed to be poor, but examination is limited. This systematic review aimed to assess the performance of actigraphy in sleep stage classification of adults. A systematic search was performed using MEDLINE, Web of Science, Google Scholar, and Embase databases. We identified eight studies that compared sleep architecture estimates between wrist-worn actigraphy and polysomnography. Large heterogeneity was found with respect to how sleep stages were grouped, and the choice of metrics used to evaluate performance. Quantitative synthesis was not possible, so we performed a narrative synthesis of the literature. From the limited number of studies, we found that actigraphy-based sleep staging had some ability to classify different sleep stages compared with polysomnography.

Comparison of actigraphy-measured and parent-reported sleep in association with weight status among preschool children

This study compared weekday and weekend actigraphy-measured and parent-reported sleep in relation to weight status among preschool-aged children. Participants were 3-6 years old preschoolers from the cross-sectional DAGIS-study with sleep data for ≥2 weekday and ≥2 weekend nights. Parents-reported sleep onset and wake-up times were gathered alongside 24 h hip-worn actigraphy. An unsupervised Hidden-Markov Model algorithm provided actigraphy-measured night time sleep without the guidance of reported sleep times. Waist-to-height ratio and age-and-sex-specific body mass index characterised weight status. Comparison of methods were assessed with consistency in quintile divisions and Spearman correlations. Associations between sleep and weight status were assessed with adjusted regression models. Participants included 638 children (49% girls) with a mean ± SD age of 4.76 ± 0.89. On weekdays, 98%-99% of actigraphy-measured and parent-reported sleep estimates were classified in the same or adjacent quintile and were strongly correlated (rs  = 0.79-0.85, p < 0.001). On weekends, 84%-98% of actigraphy-measured and parent-reported sleep estimates were respectively classified and correlations were moderate to strong (rs  = 0.62-0.86, p < 0.001). Compared with actigraphy-measured sleep, parent-reported sleep had consistently earlier onset, later wake-up, and greater duration. Earlier actigraphy-measured weekday sleep onset and midpoint were associated with a higher body mass index (respective β-estimates: -0.63, p < 0.01 and -0.75, p < 0.01) and waist-to-height ratio (-0.004, p = 0.03 and -0.01, p = 0.02). Though the sleep estimation methods were consistent and correlated, actigraphy measures should be favoured as they are more objective and sensitive to identifying associations between sleep timing and weight status compared with parent reports.

Four State Sleep Staging From a Multilayered Algorithm Using Electrocardiographic and Actigraphic Data

PURPOSE

Sleep studies are important to evaluate sleep and sleep-related disorders. The standard test for evaluating sleep is polysomnography, during which several physiological signals are recorded separately and simultaneously with specialized equipment that requires a technologist. Simpler recordings that can model the results of a polysomnography would provide the benefit of expanding the possibilities of sleep recordings.

METHODS

Using the publicly available sleep data set from the multiethnic study of atherosclerosis and 1769 nights of sleep, we extracted a distinct data subset with engineered features of the biomarkers collected by actigraphic, oxygenation, and electrocardiographic sensors. We then applied scalable models with recurrent neural network and Extreme Gradient Boosting (XGBoost) with a layered approach to produce an algorithm that we then validated with a separate data set of 177 nights.

RESULTS

The algorithm achieved an overall performance of 0.833 accuracy and 0.736 kappa in classifying into four states: wake, light sleep, deep sleep, and rapid eye movement (REM). Using feature analysis, we demonstrated that heart rate variability is the most salient feature, which is similar to prior reports.

CONCLUSIONS

Our results demonstrate the potential benefit of a multilayered algorithm and achieved higher accuracy and kappa than previously described approaches for staging sleep. The results further the possibility of simple, wearable devices for sleep staging. Code is available at https://github.com/NovelaNeuro/nEureka-SleepStaging.

Mixed-Input Deep Learning Approach to Sleep/Wake State Classification by Using EEG Signals

Sleep stage classification plays a pivotal role in predicting and diagnosing numerous health issues from human sleep data. Manual sleep staging requires human expertise, which is occasionally prone to error and variation. In recent times, availability of polysomnography data has aided progress in automatic sleep-stage classification. In this paper, a hybrid deep learning model is proposed for classifying sleep and wake states based on a single-channel electroencephalogram (EEG) signal. The model combines an artificial neural network (ANN) and a convolutional neural network (CNN) trained using mixed-input features. The ANN makes use of statistical features calculated from EEG epochs, and the CNN operates on Hilbert spectrum images generated during each epoch. The proposed method is assessed using single-channel Pz-Oz EEG signals from the Sleep-EDF database Expanded. The classification performance on four randomly selected individuals shows that the proposed model can achieve accuracy of around 96% in classifying between sleep and wake states from EEG recordings.

Self-supervised learning of accelerometer data provides new insights for sleep and its association with mortality

Background

Sleep is essential to life. Accurate measurement and classification of sleep/wake and sleep stages is important in clinical studies for sleep disorder diagnoses and in the interpretation of data from consumer devices for monitoring physical and mental well-being. Existing non-polysomnography sleep classification techniques mainly rely on heuristic methods developed in relatively small cohorts. Thus, we aimed to establish the accuracy of wrist-worn accelerometers for sleep stage classification and subsequently describe the association between sleep duration and efficiency (proportion of total time asleep when in bed) with mortality outcomes.

Methods

We developed and validated a self-supervised deep neural network for sleep stage classification using concurrent laboratory-based polysomnography and accelerometry data from three countries (Australia, the UK, and the USA). The model was validated within-cohort using subject-wise five-fold cross-validation for sleep-wake classification and in a three-class setting for sleep stage classification wake, rapid-eye-movement sleep (REM), non-rapid-eye-movement sleep (NREM) and by external validation. We assessed the face validity of our model for population inference by applying the model to the UK Biobank with 100,000 participants, each of whom wore a wristband for up to seven days. The derived sleep parameters were used in a Cox regression model to study the association of sleep duration and sleep efficiency with all-cause mortality.

Findings

After exclusion, 1,448 participant nights of data were used to train the sleep classifier. The difference between polysomnography and the model classifications on the external validation was 34.7 minutes (95% limits of agreement (LoA): -37.8 to 107.2 minutes) for total sleep duration, 2.6 minutes for REM duration (95% LoA: -68.4 to 73.4 minutes) and 32.1 minutes (95% LoA: -54.4 to 118.5 minutes) for NREM duration. The derived sleep architecture estimate in the UK Biobank sample showed good face validity. Among 66,214 UK Biobank participants, 1,642 mortality events were observed. Short sleepers (<6 hours) had a higher risk of mortality compared to participants with normal sleep duration (6 to 7.9 hours), regardless of whether they had low sleep efficiency (Hazard ratios (HRs): 1.69; 95% confidence intervals (CIs): 1.28 to 2.24 ) or high sleep efficiency (HRs: 1.42; 95% CIs: 1.14 to 1.77).

Interpretation

Deep-learning-based sleep classification using accelerometers has a fair to moderate agreement with polysomnography. Our findings suggest that having short overnight sleep confers mortality risk irrespective of sleep continuity.

40 years of actigraphy in sleep medicine and current state of the art algorithms

For the last 40 years, actigraphy or wearable accelerometry has provided an objective, low-burden and ecologically valid approach to assess real-world sleep and circadian patterns, contributing valuable data to epidemiological and clinical insights on sleep and sleep disorders. The proper use of wearable technology in sleep research requires validated algorithms that can derive sleep outcomes from the sensor data. Since the publication of the first automated scoring algorithm by Webster in 1982, a variety of sleep algorithms have been developed and contributed to sleep research, including many recent ones that leverage machine learning and / or deep learning approaches. However, it remains unclear how these algorithms compare to each other on the same data set and if these modern data science approaches improve the analytical validity of sleep outcomes based on wrist-worn acceleration data. This work provides a systematic evaluation across 8 state-of-the-art sleep algorithms on a common sleep data set with polysomnography (PSG) as ground truth. Despite the inclusion of recently published complex algorithms, simple regression-based and heuristic algorithms demonstrated slightly superior performance in sleep-wake classification and sleep outcome estimation. The performance of complex machine learning and deep learning models seem to suffer from poor generalization. This independent and systematic analytical validation of sleep algorithms provides key evidence on the use of wearable digital health technologies for sleep research and care.

A novel method to increase specificity of sleep-wake classifiers based on wrist-worn actigraphy

The knowledge of the distribution of sleep and wake over a 24-h day is essential for a comprehensive image of sleep-wake rhythms. Current sleep-wake scoring algorithms for wrist-worn actigraphy suffer from low specificities, which leads to an underestimation of the time staying awake. The goal of this study (ClinicalTrials.gov Identifier: NCT03356938) was to develop a sleep-wake classifier with increased specificity. By artificially balancing the training dataset to contain as much wake as sleep epochs from day- and nighttime measurements from 12 subjects, we optimized the classification parameters to an optimal trade-off between sensitivity and specificity. The resulting sleep-wake classifier achieved high specificity of 80.4% and sensitivity of 88.6% on the balanced dataset containing 3079.9 h of actimeter data. In the validation on night sleep of separate adaptation recordings from 19 healthy subjects, the sleep-wake classifier achieved 89.4% sensitivity and 64.6% specificity and estimated accurately total sleep time and sleep efficiency with a mean difference of 12.16 min and 2.83%, respectively. This new, device-independent method allows to rid sleep-wake classifiers from their bias towards sleep detection and lay a foundation for more accurate assessments in everyday life, which could be applied to monitor patients with fragmented sleep-wake rhythms.

Generalizability and performance of methods to detect non-wear with free-living accelerometer recordings

Wearable physical activity sensors are widely used in research and practice as they provide objective measures of human behavior at a low cost. An important challenge for accurate assessment of physical activity behavior in free-living is the detection non-wear. Traditionally, heuristic algorithms that rely on specific interval lengths have been employed to detect non-wear time; however, machine learned models are emerging. We explore the potential of detecting non-wear using decision trees that combine raw acceleration and skin temperature, and we investigate the generalizability of our models, traditional heuristic algorithms, and recently developed machine learned models by external validation. The Decision tree models were trained using one week of data from thigh- and hip-worn accelerometers from 64 children. External validation was performed using data from wrist-worn accelerometers of 42 adolescents. For non-wear episodes longer than 60 min, the heuristic algorithms performed the best with F1-scores above 0.96. However, regarding episodes shorter than 60 min, the best performing method was the decision tree model including the six most important predictors with F1 scores above 0.74 for all sensor locations. We conclude that for classifying non-wear time, researchers should carefully select an appropriate method and we encourage the use of external validation when reporting on machine learned non-wear models.

Contact-free radar recordings of body movement can reflect ultradian dynamics of sleep

This work aimed to evaluate if a contact-free radar sensor can be used to observe ultradian patterns in sleep physiology, by way of a data processing tool known as Locomotor Inactivity During Sleep (LIDS). LIDS was designed as a simple transformation of actigraphy recordings of wrist movement, meant to emphasise and enhance the contrast between movement and non-movement and to reveal patterns of low residual activity during sleep that correlate with ultradian REM/NREM cycles. We adapted the LIDS transformation for a radar that detects body movements without direct contact with the subject and applied it to a dataset of simultaneous recordings with polysomnography, actigraphy, and radar from healthy young adults (n = 12, four nights of polysomnography per participant). Radar and actigraphy-derived LIDS signals were highly correlated with each other (r > 0.84), and the LIDS signals were highly correlated with reduced-resolution polysomnographic hypnograms (r_radars  >0.80, r_actigraph  >0.76). Single-harmonic cosine models were fitted to LIDS signals and hypnograms; significant differences were not found between their amplitude, period, and phase parameters. Mixed model analysis revealed similar slopes of decline per cycle for radar-LIDS, actigraphy-LIDS, and hypnograms. Our results indicate that the LIDS technique can be adapted to work with contact-free radar measurements of body movement; it may also be generalisable to data from other body movement sensors. This novel metric could aid in improving sleep monitoring in clinical and real-life settings, by providing a simple and transparent way to study ultradian dynamics of sleep using nothing more than easily obtainable movement data.

Sleep classification using Consumer Sleep Technologies and AI: A review of the current landscape

Classifying sleep stages in real-time represents considerable potential, for instance in enabling interactive noise masking in noisy environments when persons are in a state of light sleep or to support clinical staff in analyzing sleep patterns etc. However, the current gold standard for classifying sleep stages, Polysomnography (PSG), is too cumbersome to apply outside controlled hospital settings and requires manual as well as highly specialized knowledge to classify sleep stages. Using data from Consumer Sleep Technologies (CSTs) to inform machine learning algorithms represent a promising opportunity for automating the process of classifying sleep stages, also in settings outside the confinements of clinical expert settings. This study reviews 27 papers that use CSTs in combination with Artificial Intelligence (AI) models to classify sleep stages. AI models and their performance are described and compared to synthesize current state of the art in sleep stage classification with CSTs. Furthermore, gaps in the current approaches are shown and how these AI models could be improved in the near-future. Lastly, the challenges of designing interactions for users that are asleep are highlighted pointing towards avenues of more interactive sleep interventions based on AI-infused CSTs solutions.

Detecting Sleep and Nonwear in 24-h Wrist Accelerometer Data from the National Health and Nutrition Examination Survey

INTRODUCTION

Estimating physical activity, sedentary behavior, and sleep from wrist-worn accelerometer data requires reliable detection of sensor nonwear and sensor wear during both sleep and wake.

PURPOSE

This study aimed to develop an algorithm that simultaneously identifies sensor wake-wear, sleep-wear, and nonwear in 24-h wrist accelerometer data collected with or without filtering.

METHODS

Using sensor data labeled with polysomnography ( n = 21) and directly observed wake-wear data ( n = 31) from healthy adults, and nonwear data from sensors left at various locations in a home ( n = 20), we developed an algorithm to detect nonwear, sleep-wear, and wake-wear for "idle sleep mode" (ISM) filtered data collected in the 2011-2014 National Health and Nutrition Examination Survey. The algorithm was then extended to process original raw data collected from devices without ISM filtering. Both algorithms were further validated using a polysomnography-based sleep and wake-wear data set ( n = 22) and diary-based wake-wear and nonwear labels from healthy adults ( n = 23). Classification performance (F1 scores) was compared with four alternative approaches.

RESULTS

The F1 score of the ISM-based algorithm on the training data set using leave-one-subject-out cross-validation was 0.95 ± 0.13. Validation on the two independent data sets yielded F1 scores of 0.84 ± 0.60 for the data set with sleep-wear and wake-wear and 0.94 ± 0.04 for the data set with wake-wear and nonwear. The F1 score when using original, raw data was 0.96 ± 0.08 for the training data sets and 0.86 ± 0.18 and 0.97 ± 0.04 for the two independent validation data sets. The algorithm performed comparably or better than the alternative approaches on the data sets.

CONCLUSIONS

A novel machine-learning algorithm was designed to recognize wake-wear, sleep-wear, and nonwear in 24-h wrist-worn accelerometer data that are applicable for ISM-filtered data or original raw data.

Development and testing of methods for detecting off-wrist in actimetry recordings

Study Objectives

In field studies using wrist-actimetry, not identifying/handling off-wrist intervals may result in their misclassification as immobility/sleep and biased estimations of rhythmic patterns. By comparing different solutions for detecting off-wrist, our goal was to ascertain how accurately they detect nonwear in different contexts and identify variables that are useful in the process.

Methods

We developed algorithms using heuristic (HA) and machine learning (ML) approaches. Both were tested using data from a protocol followed by 10 subjects, which was devised to mimic contexts of actimeter wear/nonwear in real-life. Self-reported data on usage according to the protocol were considered the gold standard. Additionally, the performance of our algorithms was compared to that of visual inspection (by 2 experienced investigators) and Choi algorithm. Data previously collected in field studies were used for proof-of-concept analyses.

Results

All methods showed similarly good performances. Accuracy was marginally higher for one of the raters (visual inspection) than for heuristically developed algorithms (HA, Choi). Short intervals (especially &lt; 2 h) were either not or only poorly identified. Consecutive stretches of zeros in activity were considered important indicators of off-wrist (for both HA and ML). It took hours for raters to complete the task as opposed to the seconds or few minutes taken by the automated methods.

Conclusions

Automated strategies of off-wrist detection are similarly effective to visual inspection, but have the important advantage of being faster, less costly, and independent of raters’ attention/experience. In our study, detecting short intervals was a limitation across methods.

Multi-Scale Evaluation of Sleep Quality Based on Motion Signal from Unobtrusive Device

Sleep disorders are a growing threat nowadays as they are linked to neurological, cardiovascular and metabolic diseases. The gold standard methodology for sleep study is polysomnography (PSG), an intrusive and onerous technique that can disrupt normal routines. In this perspective, m-Health technologies offer an unobtrusive and rapid solution for home monitoring. We developed a multi-scale method based on motion signal extracted from an unobtrusive device to evaluate sleep behavior. Data used in this study were collected during two different acquisition campaigns by using a Pressure Bed Sensor (PBS). The first one was carried out with 22 subjects for sleep problems, and the second one comprises 11 healthy shift workers. All underwent full PSG and PBS recordings. The algorithm consists of extracting sleep quality and fragmentation indexes correlating to clinical metrics. In particular, the method classifies sleep windows of 1-s of the motion signal into: displacement (DI), quiet sleep (QS), disrupted sleep (DS) and absence from the bed (ABS). QS proved to be positively correlated (0.72±0.014) to Sleep Efficiency (SE) and DS/DI positively correlated (0.85±0.007) to the Apnea-Hypopnea Index (AHI). The work proved to be potentially helpful in the early investigation of sleep in the home environment. The minimized intrusiveness of the device together with a low complexity and good performance might provide valuable indications for the home monitoring of sleep disorders and for subjects’ awareness.

Sleep Measurement Using Wrist-Worn Accelerometer Data Compared with Polysomnography

This study determined if using alternative sleep onset (SO) definitions impacted accelerometer-derived sleep estimates compared with polysomnography (PSG). Nineteen participants (48%F) completed a 48 h visit in a home simulation laboratory. Sleep characteristics were calculated from the second night by PSG and a wrist-worn ActiGraph GT3X+ (AG). Criterion sleep measures included PSG-derived Total Sleep Time (TST), Sleep Onset Latency (SOL), Wake After Sleep Onset (WASO), Sleep Efficiency (SE), and Efficiency Once Asleep (SE_ASLEEP). Analogous variables were derived from temporally aligned AG data using the Cole-Kripke algorithm. For PSG, SO was defined as the first score of 'sleep'. For AG, SO was defined three ways: 1-, 5-, and 10-consecutive minutes of 'sleep'. Agreement statistics and linear mixed effects regression models were used to analyze 'Device' and 'Sleep Onset Rule' main effects and interactions. Sleep-wake agreement and sensitivity for all AG methods were high (89.0-89.5% and 97.2%, respectively); specificity was low (23.6-25.1%). There were no significant interactions or main effects of 'Sleep Onset Rule' for any variable. The AG underestimated SOL (19.7 min) and WASO (6.5 min), and overestimated TST (26.2 min), SE (6.5%), and SE_ASLEEP (1.9%). Future research should focus on developing sleep-wake detection algorithms and incorporating biometric signals (e.g., heart rate).

Automatic sleep scoring with LSTM networks: impact of time granularity and input signals

Supervised automatic sleep scoring algorithms are usually trained using sleep stage labels manually annotated on 30 s epochs of PSG data. In this study, we investigate the impact of using shorter epochs with various PSG input signals for training and testing a Long Short Term Memory (LSTM) neural network. An LSTM model is evaluated on the provided 30 s epoch sleep stage labels from a publicly available dataset, as well as on 10 s subdivisions. Additionally, three independent scorers re-labeled a subset of the dataset on shorter time windows. The automatic sleep scoring experiments were repeated on the re-annotated subset.The highest performance is achieved on features extracted from 30 s epochs of a single channel frontal EEG. The resulting accuracy, precision and recall were of 92.22%, 67.58% and 66.00% respectively. When using a shorter epoch as input, the performance decreased by approximately 20%. Re-annotating a subset of the dataset on shorter time epochs did not improve the results and further altered the sleep stage detection performance. Our results show that our feature-based LSTM classification algorithm performs better on 30 s PSG epochs when compared to 10 s epochs used as input. Future work could be oriented to determining whether varying the epoch size improves classification outcomes for different types of classification algorithms.

Deep phenotyping for precision medicine in Parkinson's disease

A major challenge in medical genomics is to understand why individuals with the same disorder have different clinical symptoms and why those who carry the same mutation may be affected by different disorders. In every complex disorder, identifying the contribution of different genetic and non-genetic risk factors is a key obstacle to understanding disease mechanisms. Genetic studies rely on precise phenotypes and are unable to uncover the genetic contributions to a disorder when phenotypes are imprecise. To address this challenge, deeply phenotyped cohorts have been developed for which detailed, fine-grained data have been collected. These cohorts help us to investigate the underlying biological pathways and risk factors to identify treatment targets, and thus to advance precision medicine. The neurodegenerative disorder Parkinson's disease has a diverse phenotypical presentation and modest heritability, and its underlying disease mechanisms are still being debated. As such, considerable efforts have been made to develop deeply phenotyped cohorts for this disorder. Here, we focus on Parkinson's disease and explore how deep phenotyping can help address the challenges raised by genetic and phenotypic heterogeneity. We also discuss recent methods for data collection and computation, as well as methodological challenges that have to be overcome.

Detecting sleep outside the clinic using wearable heart rate devices

The adoption of multisensor wearables presents the opportunity of longitudinal monitoring of sleep in large populations. Personalized yet device-agnostic algorithms can sidestep laborious human annotations and objectify cross-cohort comparisons. We developed and tested a heart rate-based algorithm that captures inter- and intra-individual sleep differences in free-living conditions and does not require human input. We evaluated it on four study cohorts using different research- and consumer-grade devices for over 2000 nights. Recording periods included both 24 h free-living and conventional lab-based night-only data. We compared our optimized method against polysomnography, sleep diaries and sleep periods produced through a state-of-the-art acceleration based method. Against sleep diaries, the algorithm yielded a mean squared error of 0.04-0.06 and a total sleep time (TST) deviation of [Formula: see text]2.70 (± 5.74) and 12.80 (± 3.89) minutes, respectively. When evaluated with PSG lab studies, the MSE ranged between 0.06 and 0.11 yielding a time deviation between [Formula: see text]29.07 and [Formula: see text]55.04 minutes. These results showcase the value of this open-source, device-agnostic algorithm for the reliable inference of sleep in free-living conditions and in the absence of annotations.

Manual Annotation of Time in Bed Using Free-Living Recordings of Accelerometry Data

With the emergence of machine learning for the classification of sleep and other human behaviors from accelerometer data, the need for correctly annotated data is higher than ever. We present and evaluate a novel method for the manual annotation of in-bed periods in accelerometer data using the open-source software Audacity^®, and we compare the method to the EEG-based sleep monitoring device Zmachine^® Insight+ and self-reported sleep diaries. For evaluating the manual annotation method, we calculated the inter- and intra-rater agreement and agreement with Zmachine and sleep diaries using interclass correlation coefficients and Bland–Altman analysis. Our results showed excellent inter- and intra-rater agreement and excellent agreement with Zmachine and sleep diaries. The Bland–Altman limits of agreement were generally around ±30 min for the comparison between the manual annotation and the Zmachine timestamps for the in-bed period. Moreover, the mean bias was minuscule. We conclude that the manual annotation method presented is a viable option for annotating in-bed periods in accelerometer data, which will further qualify datasets without labeling or sleep records.

Human chronotype: Comparison of questionnaires and wrist-worn actigraphy

In this work, we investigated the accuracy of chronotype estimation from actigraphy while evaluating the required recording length and stability over time. Chronotypes have an important role in chronobiological and sleep research. In outpatient studies, chronotypes are typically evaluated by questionnaires. Alternatively, actigraphy provides potential means for measuring chronotype characteristics objectively, which opens many applications in chronobiology research. However, studies providing objective, critical evaluation of agreement between questionnaire-based and actigraphy-based chronotypes are lacking. We recorded 3-months of actigraphy and collected Morningness-Eveningness Questionnaire (MEQ), and Munich Chronotype Questionnaire (MCTQ) results from 122 women. Regression models were applied to evaluate the questionnaire-based chronotypes scores using selected actigraphy features. Changes in predictive strength were evaluated based on actigraphy recordings of different duration. The actigraphy was significantly associated with the questionnaire-based chronotype, and the best single-feature-based models explained 37% of the variability (R²) for MEQ (p < .001), 47% for mid-sleep time MCTQ-MSFsc (p < .001), and 19% for social jetlag MCTQ-SJLrel (p < .001). Concerning stability in time, the Mid-sleep and Acrophase features showed high levels of stability (test-retest R ~ 0.8), and actigraphy-based MSFsc_acti and SJLrel_acti showed high temporal variability (test-retest R ~ 0.45). Concerning required recording length, features estimated from recordings with 3-week and longer observation periods had sufficient predictive power on unseen data. Additionally, our data showed that the subjectively reported extremes of the MEQ, MCTQ-MSFsc, and MCTQ-SJLrel are commonly overestimated compared to objective activity peak and middle of sleep differences measured by actigraphy. Such difference may be associated with chronotype time-variation. As actigraphy is considered accurate in sleep-wake cycle detection, we conclude that actigraphy-based chronotyping is appropriate for large-scale studies, especially where higher temporal variability in chronotype is expected.

Influence of Study Composition on the Efficacy of Sleep Detection Using Actigraphy

Wearable actigraphy sensors have been useful tools for unobtrusive monitoring of sleep. The influence of the composition and characteristics of study groups such as normal sleep versus sleep disorders affecting the efficacy of sleep assessment using actigraphy has not been fully examined. In this study, we present multi-variate sleep models using actigraphy features obtained from wrist-worn sensors and evaluate the efficacy of sleep detection compared to the overnight polysomnography from two unique datasets: overnight actigraphy recordings in a control population of young healthy individuals (n=31) and 24-hour actigraphy recordings in a more heterogeneous population (n=27) comprised of normal and abnormal sleepers. We evaluate the performance of actigraphy derived logistic regression (LR) and random forest (RF) sleep models for both intra-dataset and inter-dataset training and cross-validation. Both the LR and RF sleep models for the healthy sleep dataset show an area under the receiver operating characteristic (AUROC) of 0.85±0.02 in the control sleep dataset among 50 random splits of training and testing evaluations. We find the AUROC performance from the heterogeneous sleep dataset involving sleep disorders to be relatively lower as 0.74±0.05 and 0.80±0.03 for LR and RF sleep models, respectively. Optimal sleep models trained using heterogeneous datasets perform very well when tested with the normal sleep dataset producing accuracy of ∼92%. Our study supports that using a more diverse training set benefits the sleep classifier model to be more generalizable for both healthy and abnormal sleepers.

pyActigraphy: Open-source python package for actigraphy data visualization and analysis

Over the past 40 years, actigraphy has been used to study rest-activity patterns in circadian rhythm and sleep research. Furthermore, considering its simplicity of use, there is a growing interest in the analysis of large population-based samples, using actigraphy. Here, we introduce pyActigraphy, a comprehensive toolbox for data visualization and analysis including multiple sleep detection algorithms and rest-activity rhythm variables. This open-source python package implements methods to read multiple data formats, quantify various properties of rest-activity rhythms, visualize sleep agendas, automatically detect rest periods and perform more advanced signal processing analyses. The development of this package aims to pave the way towards the establishment of a comprehensive open-source software suite, supported by a community of both developers and researchers, that would provide all the necessary tools for in-depth and large scale actigraphy data analyses.

The possibility to continuously record locomotor movements using accelerometers (actigraphy) has allowed field studies of sleep and rest-activity patterns. It has also enabled large-scale data collections, opening new avenues for research. However, each brand of actigraph devices encodes recordings in its own format and closed-source proprietary softwares are typically used to read and analyse actigraphy data. In order to provide an alternative to these softwares, we developed a comprehensive open-source toolbox for actigraphy data analysis, pyActigraphy. It allows researchers to read actigraphy data from 7 different file formats and gives access to a variety of rest-activity rhythm variables, automatic sleep detection algorithms and more advanced signal processing techniques. Besides, in order to empower researchers and clinicians with respect to their analyses, we created a series of interactive tutorials that illustrate how to implement the key steps of typical actigraphy data analyses. As an open-source project, all kind of user’s contributions to our toolbox are welcome. As increasing evidence points to the predicting value of rest-activity patterns derived from actigraphy for brain integrity, we believe that the development of the pyActigraphy package will not only benefit the sleep and chronobiology research, but also the neuroscientific community at large.

Metrics of sleep apnea severity: beyond the apnea-hypopnea index

Obstructive sleep apnea (OSA) is thought to affect almost 1 billion people worldwide. OSA has well established cardiovascular and neurocognitive sequelae, although the optimal metric to assess its severity and/or potential response to therapy remains unclear. The apnea-hypopnea index (AHI) is well established; thus, we review its history and predictive value in various different clinical contexts. Although the AHI is often criticized for its limitations, it remains the best studied metric of OSA severity, albeit imperfect. We further review the potential value of alternative metrics including hypoxic burden, arousal intensity, odds ratio product, and cardiopulmonary coupling. We conclude with possible future directions to capture clinically meaningful OSA endophenotypes including the use of genetics, blood biomarkers, machine/deep learning and wearable technologies. Further research in OSA should be directed towards providing diagnostic and prognostic information to make the OSA diagnosis more accessible and to improving prognostic information regarding OSA consequences, in order to guide patient care and to help in the design of future clinical trials.

The Promise of Sleep: A Multi-Sensor Approach for Accurate Sleep Stage Detection Using the Oura Ring

Consumer-grade sleep trackers represent a promising tool for large scale studies and health management. However, the potential and limitations of these devices remain less well quantified. Addressing this issue, we aim at providing a comprehensive analysis of the impact of accelerometer, autonomic nervous system (ANS)-mediated peripheral signals, and circadian features for sleep stage detection on a large dataset. Four hundred and forty nights from 106 individuals, for a total of 3444 h of combined polysomnography (PSG) and physiological data from a wearable ring, were acquired. Features were extracted to investigate the relative impact of different data streams on 2-stage (sleep and wake) and 4-stage classification accuracy (light NREM sleep, deep NREM sleep, REM sleep, and wake). Machine learning models were evaluated using a 5-fold cross-validation and a standardized framework for sleep stage classification assessment. Accuracy for 2-stage detection (sleep, wake) was 94% for a simple accelerometer-based model and 96% for a full model that included ANS-derived and circadian features. Accuracy for 4-stage detection was 57% for the accelerometer-based model and 79% when including ANS-derived and circadian features. Combining the compact form factor of a finger ring, multidimensional biometric sensory streams, and machine learning, high accuracy wake-sleep detection and sleep staging can be accomplished.

Complexity of Body Movements during Sleep in Children with Autism Spectrum Disorder

Recently, measuring the complexity of body movements during sleep has been proven as an objective biomarker of various psychiatric disorders. Although sleep problems are common in children with autism spectrum disorder (ASD) and might exacerbate ASD symptoms, their objectivity as a biomarker remains to be established. Therefore, details of body movement complexity during sleep as estimated by actigraphy were investigated in typically developing (TD) children and in children with ASD. Several complexity analyses were applied to raw and thresholded data of actigraphy from 17 TD children and 17 children with ASD. Determinism, irregularity and unpredictability, and long-range temporal correlation were examined respectively using the false nearest neighbor (FNN) algorithm, information-theoretic analyses, and detrended fluctuation analysis (DFA). Although the FNN algorithm did not reveal determinism in body movements, surrogate analyses identified the influence of nonlinear processes on the irregularity and long-range temporal correlation of body movements. Additionally, the irregularity and unpredictability of body movements measured by expanded sample entropy were significantly lower in ASD than in TD children up to two hours after sleep onset and at approximately six hours after sleep onset. This difference was found especially for the high-irregularity period. Through this study, we characterized details of the complexity of body movements during sleep and demonstrated the group difference of body movement complexity across TD children and children with ASD. Complexity analyses of body movements during sleep have provided valuable insights into sleep profiles. Body movement complexity might be useful as a biomarker for ASD.

Sleep classification from wrist-worn accelerometer data using random forests

Accurate and low-cost sleep measurement tools are needed in both clinical and epidemiological research. To this end, wearable accelerometers are widely used as they are both low in price and provide reasonably accurate estimates of movement. Techniques to classify sleep from the high-resolution accelerometer data primarily rely on heuristic algorithms. In this paper, we explore the potential of detecting sleep using Random forests. Models were trained using data from three different studies where 134 adult participants (70 with sleep disorder and 64 good healthy sleepers) wore an accelerometer on their wrist during a one-night polysomnography recording in the clinic. The Random forests were able to distinguish sleep-wake states with an F1 score of 73.93% on a previously unseen test set of 24 participants. Detecting when the accelerometer is not worn was also successful using machine learning ([Formula: see text]), and when combined with our sleep detection models on day-time data provide a sleep estimate that is correlated with self-reported habitual nap behaviour ([Formula: see text]). These Random forest models have been made open-source to aid further research. In line with literature, sleep stage classification turned out to be difficult using only accelerometer data.

Exploration of Sleep as a Specific Risk Factor for Poor Metabolic and Mental Health: A UK Biobank Study of 84,404 Participants

PURPOSE

Short and long sleep durations have adverse effects on physical and mental health. However, most studies are based on self-reported sleep duration and health status. Therefore, this longitudinal study aims to investigate objectively measured sleep duration and subsequent primary health care records in older adults to investigate the impact of sleep duration and fragmentation on physical and mental health.

METHODS

Data on objective sleep duration were measured using accelerometry. Primary care health records were then obtained from the UK Biobank (n=84,404). Participants (mean age, 62.4 years) were divided into five groups according to their sleep duration derived from the accelerometry data: <5 hours, 5-6 hours, 6-7 hours, 7-8 hours and >8 hours. ICD-10 codes were used for the analysis of primary care data. Wake after sleep onset, activity level during the least active 5 hours and episodes of movement during sleep were analysed as an indication for sleep fragmentation. Binary regression models were adjusted for age, gender and Townsend deprivation score.

RESULTS

A "U-shaped" relationship was found between sleep duration and diseases including diabetes, hypertension and heart disease and depression. Short and long sleep durations and fragmented sleep were associated with increased odds of disease.

CONCLUSION

Six to eight hours of sleep, as well as less fragmented sleep, predicted better long-term metabolic and mental health.