Deep Learning of Multimodal Features Derived from Raw Polysomnography for Cognitive Impairment Prediction

Somesh Saini1, Matheus Araujo2, Sikawat Thanaviratananich2
1Cleveland Clinic, 2Cleveland Clinic Foundation


Abstract

Cognitive impairment is often diagnosed years after symptoms begin, motivating earlier, noninvasive risk assessment. We investigated whether multimodal features derived from overnight polysomnography can predict cognitive impairment 1-6 years after a sleep study. Each 30-second epoch was represented by 105 features, including EEG, EMG, SpO2, heart-rate variability, and automated sleep annotations. We then fused the epoch-level features with whole-night aggregated features and demographics using a temporal convolutional network for binary prediction of cognitive impairment. As part of the George B. Moody PhysioNet Challenge 2026, our team (CLECLINIC) achieved an age-conditioned AUROC of 0.696 and a prevalence-weighted reward of 0.125 on the official validation set. These results support the potential of routine polysomnography for opportunistic cognitive-risk assessment.