Overnight polysomnography may contain physiological signatures of cognitive impairment diagnosed years later. Team NLICA developed a hybrid sleep-EEG system to predict a future cognitive impairment diagnosis 1-6 years after polysomnography. The large training set comprised 6,600 studies from three sources, and performance was assessed on the official hidden validation set from a source not represented in training.
The system combined an engineered-feature branch with a raw-EEG convolutional branch. The engineered branch represented each study using 281 candidate features spanning demographics, sleep-stage-specific EEG spectral characteristics, spindle activity, alpha peak frequency, slow oscillations, slow-oscillation-spindle coupling, and temporal descriptors from automated sleep annotations. Site-wise location-scale harmonization reduced inter-site distribution differences. Chronological age was removed as a direct predictor, and linear age trends in continuous features were estimated exclusively from negative-class training records and removed from all records to limit direct and linear age-related contributions to the engineered branch. Random-forest importance retained 200 features, which were standardized and used in a 30-model ensemble of XGBoost, random forests, and extremely randomized trees across 10 seeds. In parallel, a four-block one-dimensional convolutional network analyzed artifact-screened 30-s C3-M2, C4-M1, O1-M2, and O2-M1 EEG epochs from N2, N3, and rapid eye movement sleep. Epoch probabilities were averaged per patient. The two branches were combined by equal-weight late fusion, with engineered-only fallback when a CNN prediction was unavailable. A prespecified cohort-level top-decile rule generated binary predictions.
Across three leave-one-site-out folds on the training data, the mean and worst-site age-conditioned AUROCs were 0.694 and 0.654, respectively. On the official hidden validation set, the primary age-conditioned AUROC was 0.736; age-weighted AUROC, overall AUROC, AUPRC, and reward were 0.715, 0.835, 0.302, and 0.172, respectively. These findings support the cross-source predictive value of combining site-harmonized, age-residualized engineered sleep features with raw EEG for future cognitive impairment prediction.