Polysomnography (PSG) captures brain, cardiac, respiratory, and motor activity during sleep and may reveal early signs of future cognitive impairment (CI) years before clinical diagnosis. In the framework of the George B. Moody PhysioNet Challenge 2026, we developed a screening pipeline built around a specific physiological hypothesis: CI is preceded by accelerated, decoupled aging of the brain and the heart during sleep. From each recording we derived 139 stage-resolved EEG, EOG, EMG, ECG, respiratory, limb-movement, sleep-architecture, and annotation-confidence features, which we site-standardized and residualized on age and sex before deriving out-of-fold
brain-age'' andheart-age'' gaps from two dedicated gradient-boosting regressors. A calibrated histogram-based gradient boosting classifier, thresholded to maximize the Challenge's prevalence-based reward on out-of-fold (OOF) predictions, produced the final CI probability. On the hidden official-phase validation set our submission achieved an age-conditioned AUROC of 0.600 and a prevalence-based reward of -0.026. Post-hoc leave-one-site-out cross-validation reproduces the AUROC almost exactly (0.634), showing the shortfall reflects site-level distribution shift, not a failure to learn real signal.