Effects of sleep restriction and sleep deprivation on behaviorally relevant electrophysiological signal patterns revealed by interpretable machine learning

Acute and accumulated sleep loss differently impair alertness and cognition, yet context-aware biomarkers remain elusive. We applied interpretable machine learning to multimodal physiological data collected under controlled conditions to identify reliable biomarkers. In a within-participant, cross-over design, 47 healthy men (mean age: 23.7 ± 2.8 years) underwent 5-min waking electroencephalographic (EEG) recordings with eyes open after normal sleep (8 h sleep opportunity), sleep restriction (four nights, 6 h sleep opportunity), and acute total sleep deprivation (0 h sleep opportunity). Recordings included six EEG derivations, and simultaneous heart rate and eye-movement data. A subsample (n = 20) completed subjective sleepiness assessments, psychomotor vigilance tests, and a 22-min immersive driving simulation with EEG monitoring.,Classification models using participant-stratified cross-validation achieved 75% accuracy in distinguishing normal sleep from total deprivation despite varied timepoints and cognitive loads. Key EEG features included frontal theta, alpha and beta power, fronto-central mutual information, and a novel biomarker, frontal Hilbert envelope root mean square of successive differences, indicative of altered cortical excitability. Consistent with differential cognitive impairments, sleep restriction showed a distinct physiological signature compared to control and sleep deprivation, particularly in eye movements and heart rate variability. Importantly, the top-10 features showed subjective and behavioral relevance, explaining up to 37% of the variance in subjective sleepiness, vigilant attention and driving performance.,In conclusion, findings demonstrate the feasibility of interpretable machine learning for identifying electrophysiology-based biomarkers of sleep-related cognitive impairment and underscore the need for diverse samples and tasks to enhance generalizability.