Alzheimer's disease and related disorders are a major cause of disability and dependency. Diagnosis is often made in late stages, after the clinical symptoms have become strongly manifested and subsequently noticed by close relatives. This delay motivates the development of physiological markers for earlier risk stratification. As part of the 2026 George B. Moody PhysioNet Challenge, our team, OOM_finder, investigated whether transfer sleep-metric outcomes derived from overnight polysomnography (PSG) could directly predict future cognitive impairment.
We generated temporally ordered whole-night probabilistic and binarized sleep-metric representations targeting sleep stages, arousals, leg movements, and respiratory obstructions, using the provided CAISR algorithmic annotations for sleep staging, arousal detection, respiratory-event detection, and limb-movement analysis, together with outputs from a self-trained open-source SleePyPhases model. These representations were then used as input for a hierarchical transformer with epoch-level, higher-resolution arousal, and sparse-event branches. This architecture was designed to jointly model global sleep architecture, sleep fragmentation, and transient nocturnal events across the entire night while preserving uncertainty in upstream predictions.
Performance on the training data was assessed using both 5-fold cross-validation and a leave-one-site-out evaluation. The mean AUROC was 0.67647 ± 0.03383 across folds and 0.71978 ± 0.08389 across held-out sites. During the unofficial phase, our best challenge submission achieved an AUROC of 0.588, placing us 66th on the leaderboard.
Our findings suggest that sleep-derived outcome sequences over the whole night are clinically relevant intermediate representation for estimating future cognitive vulnerability from PSG. An important avenue of future work during the official phase will be the systematic evaluation of additional sleep representations transferred from open-source models, such as SleepFM and Philosopher's Stone.