What question did this study set out to answer?

This research aims to evaluate whether continuous sleep depth measurements provide better predictors of sleep fragmentation and apnea events compared to traditional methods.

May 10, 2026

0345 Continuous Sleep Depth and Fragmentation Biomarkers Predict Apnea and Arousal Burden from PSG

Key Points

This research aims to evaluate whether continuous sleep depth measurements provide better predictors of sleep fragmentation and apnea events compared to traditional methods.
Analyzed 151 full-night polysomnograms from the PhysioNet HMC sleep database.
Utilized a multitask neural network to estimate Sleep Depth Index per epoch, suppressing jitter with temporal smoothness.
Correlations with delta power and spindle density were assessed, and associations with apnea-hypnea index and arousal index were examined.
Sleep Depth Index showed strong correlation with delta power (rho = 0.71) and spindle density (rho = 0.44).
SDI fragmentation biomarkers correlated with apnea-hypnea index (rho = 0.47) and arousal index (rho = 0.53).
Biomarker model achieved an AUC of 0.78 for predicting AHI >= 15 events/h and 0.80 for AHI >= 30 events/h.

Abstract

Abstract Introduction The sleep process doesn’t move in discrete boxes of thirty seconds, it slides continuously. We examine the question of whether a sleep-depth signal that continues through the sleep period, instead of discretized into stages, has the potential to provide sleep fragmentation indexes more informative than those based upon the latter. Methods The PhysioNet Haaglanden Medisch Centrum (HMC) sleep database consisted of 151 full-night polysomnograms, and a multitask neural network was applied to estimate a Sleep Depth Index (SDI) per epoch (0 = wake, 1 = deepest NREM), where a separate head handled REM sleep and a temporal smoothness term suppressed jitter. For each night, the SDI series yielded fragmentation indexes (mean SDI, time spent at low depths, total variation/volatility, and number of events where the index drops per hour). Physiologic validity was evaluated through correlations with delta power and density of spindles, while clinical validity was examined through associations with the apnea-hypnea index (AHI) and the arousal index (ArI), and through the ability to distinguish moderate to severe OSA (AHI = 15) and severe OSA (AHI = 30) through cross-validation across subjects. Results SDI tracked expected physiology, relating to delta band power (Spearman’s rho = 0.71) and spindle density (rho = 0.44). SDI fragmentation biomarkers rose with severity and related to the AHI (rho = 0.47) and ArI (rho = 0.53). The biomarker model for the prediction of an AHI of =15 events/h had an AUC of 0.78 and for the prediction of an AHI of =30 events/h the model’s AUC was 0.80, an improvement upon the model based upon sleep stage (delta AUC = 0.04) for the prediction of an AHI of =15 events/h and 0.05 for the prediction of an AHI of =30 events/h. Conclusion The sleep fragmentation signatures derived in this manner are interpretable and correlate with the sleep apnea and arousal events recorded in PSG sleep studies. “SDI-type” metrics might be combined with traditional sleep staging to assess the value of instability in a way that has scaling potential in terms of relating PSG physiologic measures to meaningful risk. Support (if any)

Mark Helpful

Bookmark

Relay