AIMS: We aimed to predict cognitive impairment occurring 3–7 years after an overnight polysomnography (PSG) recording using the PhysioNet/CinC Challenge 2026 dataset, with a focus on evaluating cross-site generalization, METHODS: The provided data comprised 622 patients from three sites with balanced labels, including raw multi-channel PSG signals, demographic information, and CAISR algorithmic annotations. From physiological signals we extracted 49 time-domain features (7 per channel type, including Hjorth parameters), with each channel z-score normalized per patient to remove site-dependent amplitude differences. From CAISR annotations we extracted 24 features covering event indices, sleep architecture, fragmentation metrics, respiratory event subtype breakdowns, sleep depth ratios, and probability trace statistics. We compared two demographic feature sets: a reduced set with age only, and a full set adding sex, race, and BMI (mean-imputed for missing values). We performed grid-search cross-validation across seven model families and soft-voting ensembles, evaluated using both stratified 5-fold and leave-one-site-out cross-validation, comparing the reduced (74 features) and full (83 features) feature sets. RESULTS: Model performance differed substantially between validation strategies, indicating strong site-specific overfitting. Under stratified 5-fold CV, Gradient Boosting on the full feature set achieved 0.854 AUROC, but the same configuration dropped to 0.510 under leave-one-site-out CV. Tree-based models generally showed high training AUROC (0.90-0.99) but poor cross-site performance (0.51-0.63). Using the reduced feature set, L1-regularized Logistic Regression achieved 0.687 training and 0.634 validation AUROC under leave-one-site-out. Our submission achieved a score of 0.553 AUROC on the hidden validation set (team name KBA). CONCLUSION: Removing site-related demographic information improved cross-site generalization. Simpler regularized linear models outperformed more complex tree-based models, which severely overfit to site-specific patterns. For future work, we plan to train deep neural networks directly on raw PSG signals, leveraging large external sleep datasets for self-supervised pretraining to learn site-invariant representations, before fine-tuning on the challenge data.