Redefining Sleep: Data Driven Biomarkers of Sleep Quality

2019 Strategic Research Grant

M. Brandon Westover, MD, PhD

Massachusetts General Hospital

Key Project Outcomes

1. Analysis of effect of chronic HIV infection on sleep brain age: One special population where sleep disturbances and cognitive problems go hand-in-hand is patients with chronic HIV infection. Age-related comorbidities and immune activation have both been suggested as risk factor for accelerated brain aging in people with HIV (PWH). We conducted a study to estimate the effect of HIV infection on one of the key metrics of sleep quality that we are investigating, the “brain age index”, BAI. We found that the average effect of HIV on BAI is 3.35 years. Compared to HIV- controls, HIV+ participants exhibited a reduction in delta band power during deep sleep and rapid eye movement sleep. Our results provide evidence that HIV contributes to advanced brain aging reflected in sleep EEG. Our results also suggest the testable hypothesis that targeting sleep disorders and cardiovascular disease could mitigate the effect of HIV on brain health.

2. Night-to-night stability and test-retest reliability of the brain age index: We analyzed night-to-night stability of the brain age index (BAI) in multi-night sleep EEG recordings. We found that the within-patient night-to-night standard deviation of BAI was 7.5 years. Estimates of BAI derived by averaging over 2, 3, and 4 nights had standard deviations of 4.7, 3.7, and 3.0 years, respectively. Thus, averaging BAI over n nights reduces night-to-night variability of BAI by a factor of √n. This suggests that, when applied at the individual level, it is advantageous to record multiple nights of sleep and average them to obtain a more stable estimate of brain age. The importance of this study is that, with increasing ease of EEG acquisition, and especially home wearable EEG technology, BAI has the potential to track brain health and detect deviations from normal physiologic function. The measure of night-to-night variability and how this is reduced by averaging across multiple nights provides a basis for using BAI in patients’ homes to identify patients who should undergo further investigation or monitoring.

3. “Growth curves” for features of sleep: To help identify additional data-driven biomarkers of sleep quality beyond BAI, we have embarked on a large-scale project to estimate “growth curves” for the central features of the overnight sleep EEG. We developed “age norms”, or “sleep growth curves”, providing reference means and standard deviation values of sleep EEG parameters from over most of the lifespan, estimated from a large number of subjects. These growth curves provide a population-averaged normal trajectory of aging. We used 3,429 PSGs from 3,008 subjects with ages 11 days up to 94 years to estimate a smooth age norm of sleep EEG spectrograms. We also described the age-dependent changes in absolute and relative band powers, spindle parameters, and sex-dependent changes. Since the age norm is a smooth function of age, it can be queried for any age between 11 days to 94 years old. The age norm provides a reference that can be compared to quantify deviation from normal aging for individuals with disorders that potentially affect brain health. Together with the BAI, these growth curves allow not only detection of deviation from normal brain aging, but also insights into which specific aspects of sleep deviate from normal.

4. Data driven sleep depth: We developed a measure of sleep depth, with the intent of using average sleep depth overnight as one data driven measure of sleep quality. We call our measure of sleep depth the “ordinal sleep depth”, OSD. Based on a dataset of overnight sleep electroencephalography (EEG) from 600 patients, we developed a deep learning model to estimate sleep depth in 3-second epochs. The model was trained using a multi-tasking approach, in which the model was optimized to simultaneously to predict sleep stages on an ordinal scale, and conventional categorical stages. Sleep depth was defined as the probability estimated by the model of awakening within the next 30 seconds. We correlated OSD with the arousability index, i.e. the probability of an arousal or awakening within the next 30-second epoch. We found that OSD strongly predicts arousal within the next 30 seconds. Within each conventional sleep stage, the median depth of sleep as measured by OSD is lighter (higher vertical bar) in patients with dementia. These results suggest that patients with neurodegenerative exhibit generally lighter sleep within each of the conventional sleep stages.

5. Analysis of correlation between sleep brain age and cognition: We have completed a preliminary analysis of how the sleep brain age index (BAI) correlates with cognitive performance in 150 subjects who underwent cognitive testing after routine sleep testing (polysomnography, PSG), and of how the BAI algorithm can be further refined to provide a more precise and direct biomarker of subjects’ cognitive/brain health. In this analysis we analyzed the degree to which BAI correlates with cognitive performance. We also investigated whether cognitive performance can be directly predicted using brain activity from an overnight sleep EEG, using LASSO regression to propose a series of new model which we call the “brain health index” (BHI). Our results show that BAI is strongly correlated with crystallized cognition, but not less strongly correlated with fluid cognition. The new BHI index that we have developed shows significant correlations with multiple cognitive variables. In total, our results show that cognition is predictable using features derived from an overnight sleep EEG. This supports the overall premise of our work, that overnight sleep EEG can provide easily accessible biomarkers of brain health. A manuscript is under review.

6. Optimization of Sleep Spindle Detection: Features of sleep spindles are one component of the BAI. Alterations in sleep spindles have been linked to cognitive impairment. However, flexibility surrounding spindle definitions and algorithm parameter settings present a methodological challenge. We aimed to characterize how spindle detection parameter settings influence the association between spindle features and cognition and to identify parameters with the strongest association with cognition. We analyzed sleep EEGs and cognitive data from 167 subjects. Subjects ages were 49 ± 18 years. We explored 1000 combinations of parameters in an open-source spindle detector, and used four features of detected spindles (amplitude, density, duration, and peak frequency) to fit linear multiple regression models to predict cognitive scores. Spindle features were associated with the ability to predict raw fluid cognition and age-adjusted fluid cognition scores with the best spindle parameters. Fast spindle features generally showed better performance relative to slow spindle features. Spindle features weakly predicted total cognition and poorly predicted crystallized cognition regardless of parameter settings. Our exploration of spindle detection parameters identified optimal parameters for studies of fluid cognition and revealed the role of parameter interactions for both slow and fast spindles. Our findings support sleep spindles as a sleep-based biomarker of fluid cognition.

7. Robust BAI: In our prior work, the night-to-night variability of BAI was high: 7.5 years when measured by the standard deviation of BAIs. The level of noise leads to low sensitivity in individualized application of BAI, since one needs an even higher BAI to be certain of abnormal brain aging, where false negative rate is high and true positive rate is low. We developed a more robust version of the brain age index (rBAI) to address this gap by restricting to longitudinally stable sleep EEG features. We examined longitudinal stability of 195 sleep EEG features. An optimal sleep EEG feature list was selected to fulfill both high longitudinal precision and high correlation with age. The age prediction model was then trained using these stable sleep EEG features on 1324 healthy participants. The rBAI, i.e., brain age minus chronological age, was then validated by checking its association with 10 diagnoses and 11 future incidences of outcomes including mortality. For external validation, we computed rBAI on Sleep Heart Health Study (SHHS) and Osteoporotic Fractures in Men Study (MrOS) and checked their correlation. For night-to-night variability, we computed rBAI for a home sleep study collected using a home wearable EEG device where participants have an average of 7 nights of recordings. All analyses were compared to the previous version of BAI. The rBAI showed a lower mean absolute deviation (MAD) in three independent cohortts. The rBAI’s discrimination ability for 3 diagnoses (bipolar disorder, hypertension, ischemic stroke), were also higher than those of the previous version BAI. The rBAI’s predicted survival curves of 3 future outcomes (bipolar disorder, type II diabetes, and mortality) were significant. In terms of night-to-night variability, the standard deviation of rBAI was substantially lower than the previous version of BAI. We conclude that, by considering the longitudinal stability of sleep EEG features, it is possible to develop a robust version of BAI, which has lower prediction error on healthy participants, lower night-to-night variability, and stronger associations with outcomes compared to the previous version of BAI. This enhancement greatly improves its sensitivity at the individual level, opening its possibility to home use and further decision making. A manuscript is currently being prepared to submit for publication.

Journal Articles

IEEE Transactions on Biomedical Engineering

Automated Scoring of Respiratory Events in Sleep With a Single Effort Belt and Deep Neural Networks


HIV increases sleep-based brain age despite antiretroviral therapy

Optimal spindle detection parameters for predicting cognitive performance

Sleep and Breathing

Sleep apnea and respiratory anomaly detection from a wearable band and oxygen saturation