A Physiology-Informed Deep Learning Approach for Sleep-EEG Brain Age Estimation

Sleep electroencephalography (EEG) contains rich signatures of biological aging, but deep-learning brain-age models often provide limited physiologic readouts. This study presents a physiology-informed deep learning framework that estimates brain age from whole-night polysomnography (PSG) and provides spectral-band-level and sleep-stage-level readouts of the features associated with age predictions. Trained and evaluated on approximately 10,000 PSG recordings across four cohorts (SHHS, MrOS, STAGES, and CFS), the model achieved a mean absolute error (MAE) of 5.89 $\pm$ 5.01 years. External validation in an independent cohort (HomePAP) yielded an MAE of 5.41 $\pm$ 4.57 years. The brain age gap (BAG; predicted minus chronological age) was associated with mortality risk. Specifically, each 10-year increase in BAG was associated with a 19% increase in all-cause mortality risk (hazard ratio (HR)=1.19 [95% CI: 1.11-1.28]; p<0.001) and a 25% increase in cardiovascular mortality risk (HR=1.25 [1.07-1.46]; p=0.006) after adjustment for covariates. A trainable Gabor filter bank provides explicit center-frequency and bandwidth parameters for the learned spectral representation. Spectral features extracted using the learned filter bank explained additional variance in age beyond features computed from canonical EEG bands (incremental $\triangle \rm{R}^{2}$=4.7%; p<0.001). Attribution analysis with SHapley Additive exPlanations (SHAP) revealed age-dependent shifts in spectral importance, including decreased delta and increased beta contributions with age. The model also reflected known age-related changes in sleep stages (increased wake, reduced N3/REM). Collectively, these findings provide brain-age estimates with explicit spectral-band-level and sleep-stage-level readouts that align with established sleep neurophysiology.