Mortality risk assessment using deep learning-based frequency analysis of electroencephalography and...

Area of Science:

• Neuroscience
• Sleep Medicine
• Biostatistics

Background:

• Nocturnal polysomnography (PSG) records electroencephalography (EEG) and electrooculography (EOG) signals during sleep.
• The relationship between the spectral power of these signals and all-cause mortality is not fully understood.
• Investigating sleep microstructure may reveal novel health and mortality predictors.

Purpose of the Study:

• To determine if EEG and EOG frequency content during PSG can predict all-cause mortality.
• To identify specific frequency bands and sleep stages associated with mortality risk.

Main Methods:

• Deep learning-based survival models were applied to PSG power spectra from 8716 participants (MrOS Sleep Study and Sleep Heart Health Study).
• SHapley Additive Explanation (SHAP) was used to define sleep-stage-specific power bands.
• Cox Proportional Hazards models evaluated the predictive performance of these bands for mortality.

Main Results:

• Multiple EEG frequency bands across all sleep stages were significant predictors of all-cause mortality after adjusting for covariates.
• Specific EEG bands showed reduced mortality risk (e.g., 12-15 Hz in N2, 0.75-1.5 Hz in N3, 14.75-33.5 Hz in REM sleep).
• Certain EOG low-frequency bands were associated with increased mortality risk (e.g., 0.25 Hz in N3, 0.75 Hz in N1, 1.25-1.75 Hz in wake).

Conclusions:

• Spectral power features in EEG and EOG during sleep are associated with mortality risk.
• These findings suggest that sleep microstructure, as captured by spectral analysis, contains valuable information about an individual's health and longevity.
• Sleep analysis holds potential for identifying individuals at higher risk of mortality.