Early identification of cognitive impairment (CI) is essential for mitigating the progression of neurodegenerative diseases. This work presents REMedy's contribution to the George B. Moody PhysioNet Challenge 2026, a physiology-informed XGBoost framework to predict CI diagnosis 1-6 years after polysomnography (PSG) recordings.
Given the importance of interpretability for understanding physiological mechanisms and enabling clinical translation, a preliminary XAI-based analysis combining neural-network attention and Grad-CAM was used to identify informative PSG modalities. Together with signal availability and physiological plausibility, this analysis guided the selection of 244 raw features, spanning age and sex, respiratory dynamics, sleep-EEG measures, heart-rate variability and fragmentation. To account for age-related confounding, ECG-derived biological age and the biological–chronological age gap were incorporated alongside chronological age. To handle heterogeneous channel availability, seven XGBoost classifiers were trained for all EEG, ECG, and respiration signal availability combinations. Hyperparameters were optimized using age-conditioned AUROC as the search score, in a five-fold stratified nested cross-validation with fold-specific preprocessing. Hyperparameters selected across the outer folds were combined into consensus values and subsequently used to fit the final ensemble on the complete training set. For each recording, the classifier matching the available physiological modalities was selected dynamically. Binary predictions were obtained using a probability threshold optimized by maximizing the F1-score.
The difference between AUROC (0.837) and age-conditioned AUROC (0.636) suggests that that model discrimination may rely on age confounding caused by training data imbalances. During model development, a revised pipeline obtained stronger leave-one-site-out performance (age-conditioned AUROC 0.520 ±0.007 vs 0.626 ± 0.076) across the available training cohorts but lower performance (0.614) on the hidden validation set, indicating that performance on a specific external cohort and robustness to site shift are not equivalent, whereas clinical translation will require broader multicenter validation to establish robustness across unseen acquisition sites and patient populations.