Cognitive impairment is often recognized only after symptoms emerge. Overnight polysomnography (PSG) captures brain, cardiovascular, respiratory, oxygenation, and muscle physiology that may provide earlier markers of cognitive decline. However, institutional variation in PSG acquisition and patient populations makes cross-site prediction challenging. Competing as team NeuroAI in the 2026 George B. Moody PhysioNet Challenge, we developed a feature- stable multiscale framework combining whole-night sleep architecture and cardiorespiratory summaries with EEG sleep depth at 3 s, interhemispheric EEG contrasts at 5 s, and event timing at 1 s. Specialized descriptors are measured within records or between homologous channels, so amplifier gain and montage scale cancel before modeling and are retained only when their age-conditioned associations agree in direction across all training hospitals. The framework achieved an age-conditioned AUROC of 0.823 (95% CI, 0.797 to 0.848) in leave-one-site-out evaluation. On the hidden validation set, it achieved an age-conditioned AUROC of 0.838, AUROC of 0.926, and AUPRC of 0.655, ranking second among unique team submissions. These results support multiscale PSG for robust prediction of cognitive-impairment risk 1 to 6 years before diagnosis.