Sleep physiology contains signals associated with long-term neurological health, including risk of cognitive impairment. In this study, we present a fully tabular machine learning pipeline that leverages polysomnography (PSG)-derived features to predict future cognitive impairment.
Each PSG record is transformed into a structured feature representation combining demographic variables and automated annotations derived from the CAISR system. The selected feature configuration emphasizes demographics, sleep architecture, respiratory events, and oxygen desaturation metrics while excluding EEG, heart rate, race, ethnicity, and data availability flags. Engineered features include sleep-stage proportions, sleep efficiency, sleep fragmentation, sleep onset latency, REM latency, wake after sleep onset, bout statistics, stage-probability confidence and entropy, arousal burden and index, respiratory-event subtype indices, limb movement statistics, and oxygen desaturation summaries such as CT90, CT88, ODI3, ODI4, and SpO2 drop-area measures.
Model development is performed using LightGBM with leave-one-site-out cross-validation to ensure generalization across acquisition centers. To mitigate site imbalance, sample weighting is applied, and robustness is improved through five-seed ensembling. Final classification thresholds are selected using out-of-fold predictions.
The resulting model achieves an AUROC of 0.588 and an AUPRC of 0.048, with accuracy of 0.041 and F-measure of 0.074. These results indicate that while tabular PSG-derived features capture relevant signals, predictive performance remains limited when temporal dynamics are not explicitly modeled.
This work highlights both the potential and limitations of feature-engineered PSG representations for cognitive risk prediction, motivating the incorporation of temporal modeling approaches to better capture sleep dynamics.