Continuous Sleep Depth Index Annotation with Deep Learning Yields Novel Digital Biomarkers for Sleep Health

TL;DR

This paper introduces a deep learning framework to convert discrete sleep staging into a continuous Sleep Depth Index (SDI) from polysomnography, leveraging a pairwise ranking loss to encode ordinal depth across sleep stages and integrating REM classification. Trained on four large cohorts and externally validated, the method yields a robust SDI that correlates with arousal duration and reveals nuanced sleep structures beyond conventional staging. Time-series SDI biomarkers enable Gaussian mixture clustering that identifies a Disturbed sleep subtype linked to higher prevalence of sleep disorders and worse health outcomes, including higher all-cause mortality and fatal cardiovascular risk. The approach offers a scalable, clinically relevant digital biomarker toolkit for sleep health and demonstrates potential for integration into practice via open-source tools and a web annotation app.

Abstract

Traditional sleep staging categorizes sleep and wakefulness into five coarse-grained classes, overlooking subtle variations within each stage. It provides limited information about the duration of arousal and may hinder research on sleep fragmentation and relevant sleep disorders. To address this issue, we propose a deep learning method for automatic and scalable annotation of continuous sleep depth index (SDI) using existing discrete sleep staging labels. Our approach was validated using polysomnography from over 10,000 recordings across four large-scale cohorts. The results showcased a strong correlation between the decrease in sleep depth index and the increase in duration of arousal. Specific case studies indicated that the sleep depth index captured more nuanced sleep structures than conventional sleep staging. Gaussian mixture models based on the digital biomarkers extracted from the sleep depth index identified two subtypes of sleep, where participants in the disturbed sleep group had a higher prevalence of sleep apnea, insomnia, poor subjective sleep quality, hypertension, and cardiovascular disease. The disturbed subtype was associated with a 42% (hazard ratio 1.42, 95% CI 1.24-1.62) increased risk of mortality and a 29% (hazard ratio 1.29, 95% CI 1.00-1.67) increased risk of fatal cardiovascular disease. Our study underscores the utility of the proposed method for continuous sleep depth annotation, which could reveal more detailed information about the sleep structure and yield novel digital biomarkers for routine clinical use in sleep medicine.

Figures (11)

• Figure 1: General overview of the study. a. Four channels of physiological signals from PSG, EEG, EOG, EMG, and ECG were used in this study. The MESA, MROS, and CFS cohorts were used as the training set and interval validation set and the SHHS cohort was used as the external validation set. b. Using a deep learning method, the neural network was able to transform the discrete sleep staging into a continuous sleep depth index. c. There were several interesting pieces of evidence found in the sleep depth index, which added new insights into the understanding of sleep structure. Digital biomarkers extracted from the sleep depth index were used for clustering, resulting in sleep subtypes exhibiting varied health outcomes. A web app was provided for continuous SDI annotation supporting for EDF-format inputs. PSG, Polysomnography; EEG, Electroencephalography; EOG, Electrooculography; EMG, Electromyography; ECG, Electrocardiography; SDI, Sleep depth index; RB, Ratio below a certain threshold; CV, Coefficient of variation; AP, Proportion of area under the sleep depth index curve; SK, Skewness; MDR, Mean depth value of the REM epoch; PR, Proportion of REM to the total sleep duration; APPe, Approximate entropy; DETRf, Detrended fluctuation analysis.
• Figure 2: Cases of varied patterns of the physiological signals across the sleep stages. a. The three stages belonged to the same N2 stage but the sleep depth values differed. The sleep depth index better captured the lower frequency feature of deep sleep and the smaller magnitude of EMG. b. The three stages belonged to the same REM stage but the sleep depth values differed. The sleep depth index better captured the lower frequency feature of deep sleep and the larger magnitude of EOG. c. The first two stages belonged to the N1 stage but the second one was labeled with a slightly larger sleep depth index. The third stage shared similar patterns with the second stage, but it was labeled as the N2 stage, where the sleep depth values were close. SDI, Sleep depth index; EEG, Electroencephalography; EOG, Electrooculography; EMG, Electromyography; ECG, Electrocardiography.
• Figure 3: The correlation between the decreased magnitude of the sleep depth index and the increase in the duration of arousal. a. The SHHS cohort. b. The CFS cohort. c. The MESA cohort. d. The MROS cohort. The duration of arousal was computed as the proportion of arousal duration in a 30-second epoch. We calculated the ten deciles of 0 to 1 and averaged the values in each interval for linear regression fitting. The red dotted line represented the diagonal line. The average relationship was almost perfectly linear.
• Figure 4: Effect size estimates with 95% CIs for demographic, SDI-based features, and several health outcomes. The region above the dashed line with blue data points was about the comparison of continuous variables, using the t-test to compare the between-group differences and Cohen's d as the measure of effect size. The effect size was computed by misusing the value of the normal sleep group from the disturbed sleep group. The lower region was about the comparison of categorical variables, using the Chi-squared test to compare the between-group differences and odds ratio as the measure of effect size. The disturbed sleep group was regarded as value 1 when computing the effect size. BMI, Body Mass Index; SDI, Sleep depth index; RB, Ratio below a certain threshold; CV, Coefficient of variation; AP, Proportion of area under the sleep depth index curve; SK, Skewness; MDR, Mean depth value of the REM epoch; PR, Proportion of REM to the total sleep duration; APPe, Approximate entropy; DETRf, Detrended fluctuation analysis; CVD, Cardiovascular disease.
• Figure 5: Kaplan–Meier curves across the two subtypes for a. all-cause mortality b. fatal cardiovascular disease. HR, hazard ratio; CI, confidence interval.