Predicting cognitive impairment (CI) from multi-center polysomnography (PSG) suffers from severe site-specific confounding and two-fold data scarcity: sparse clinical annotations and severe inter-site class imbalances. We analyzed clinical PSG data to predict CI onset within 3 to 7 years. Initially, we hand-crafted features capturing spectral, autonomic, and neurodegenerative markers. A baseline gradient boosting model utilizing the top 30 features achieved 0.891 stratified cross-validation AUROC, but Leave-One-Site-Out (LOSO) validation dropped to 0.519, revealing critical site-overfitting. Adversarial feature ablation exposed massive demographic leakage; strictly removing body-mass index (BMI) improved a regularized neural network's LOSO AUROC to 0.623. Transitioning to raw signals, an end-to-end spatiotemporal network utilizing Domain Adversarial Neural Networks (DANN) achieved a validated LOSO AUROC of 0.595. Using hand-crafted features, team ThoroughbredHerd achieved an unofficial phase score of 0.428. Our ongoing research addresses structural data leakage and scarcity via a two-fold deep learning framework. First, to counteract global data scarcity, we integrate self-supervised foundation models—specifically leveraging a pre-trained(EEG, EOG) models utilizing a 4D Fourier positional encoding that natively adapts to arbitrary temporal lengths and unseen electrode configurations without requiring fixed montages. Because varied biological signals embed specific patterns at distinct frequencies, custom frequency-aware extraction methods are applied per modality. These models will be fine-tuned utilizing high-fidelity expert human annotations, specifically targeting epochs of strict human-algorithmic concordance as high-purity training signals. Second, to resolve local site scarcity, we actively augment data from underrepresented clinical sites. We capture both micro-temporal events and macro-sequence dependencies by integrating convolutional and attention-based networks. This robust, data-augmented spatiotemporal approach systematically bypasses confounding site-artifacts, ensuring true signals are learned for CI prediction.