Polysomnography (PSG) provides a non-invasive view of neurophysiological activity during sleep, with growing evidence linking disrupted sleep architecture to neurodegeneration. This study addresses the PhysioNet Challenge 2026, predicting cognitive impairment (CI) diagnosis within 3 to 7 years from overnight PSG recordings across multiple hospital sites, where cross-site generalization is the central challenge.
First, 370 features were extracted from EEG, EOG, respiratory, and SpO2 signals alongside CAISR-derived algorithmic annotations and sleep architecture measures. EEG spectral power was expressed as relative band power per sleep stage, normalized within each recording to mitigate site-specific amplitude scaling. Hjorth parameters, inter-hemispheric asymmetry, and phase-lag index were computed per sleep stage. Sleep architecture features captured macrostructural dynamics including REM and N3 duration characteristics, fragmentation indices, and cycle regularity. Algorithmic annotation features included sleep stage probabilities, arousal indices, and respiratory event rates. Second, features were selected via permutation importance: 47 features for a Random Forest and 49 features for a LightGBM classifier. To improve robustness to sites lacking frontal EEG electrodes, training data were augmented by zeroing F3 and F4 related features. Third, the base classifiers were calibrated using sigmoid calibration and combined through a logistic regression meta-learner trained on out-of-fold predictions, forming a stacking ensemble.
Our team, HEAT_ETRI, achieved a leave-one-site-out (LOSO) AUROC of 0.726 in internal validation and a challenge AUROC of 0.618 on held-out unseen sites. The gap suggests remaining challenges in cross-site generalization. Ongoing work targets three directions: site-consistent feature selection, retaining only features important across all LOSO folds; robust scaling of signals to reduce inter-site amplitude differences; and replacing stratified out-of-fold splits in the meta-learner with site-based splits, so the stacking layer explicitly learns cross-site generalization.