This study reports the GT NeuroSignals Lab submission for predicting future cognitive impairment from overnight polysomnography and demographic variables. The method uses a site aware nine component stacking framework designed to improve transportability across heterogeneous acquisition centers. The ensemble integrates complementary representations derived from selective electroencephalographic and electrooculographic channels to accommodate variable channel availability, microfragmentation and sleep transition summaries from automated annotations, brain age and temporal sequence models, a survival oriented gradient boosting component, and complementary models for cognitive risk based on signals, annotations, and their combination. This design preserves information from physiology, sleep architecture, and progression related structure while avoiding dependence on any single feature family or modeling view. At the stack level, component predictions are residualized with respect to site and then standardized within site before final logistic regression aggregation, with auxiliary meta features retained to stabilize the released default. Model selection emphasizes hidden site robustness using local out of fold evaluation rather than random record level splitting. In this protocol, the packaged default achieved an overall AUROC of 0.686, a mean site AUROC of 0.705, and a robust site AUROC of 0.700. During the unofficial phase, the GT NeuroSignals Lab entry achieved an official validation leaderboard AUROC of 0.693. These results suggest that site aware stacking and complementary physiological representations can improve generalization across centers while preserving prognostic information for longitudinal cognitive risk prediction from sleep studies.