Sleep disruption is increasingly recognized as an early marker of neurodegenerative disease, yet the predictive value of routine polysomnography for future cognitive impairment remains underexplored. We developed a pipeline to predict cognitive impairment 3-7 years after a single overnight polysomnogram using algorithmically derived sleep features and clinical demographics. We analyzed 622 recordings from the Human Sleep Project database spanning three US sites. Features were extracted from the CAISR annotation framework, which provides consistent sleep staging, arousal detection, respiratory event classification, and limb movement analysis across sites. We initially derived 211 candidate features but found large feature sets caused severe cross-site overfitting. Through leave-one-site-out cross-validation, we identified ten site-robust features: age, BMI, sleep efficiency, REM and N3 percentages, apnea-hypopnea index, periodic limb movement index, arousal index, fragmentation index, and wake after sleep onset. All sleep features are rates or percentages, eliminating equipment-dependent variability. Our model uses L2-regularized logistic regression on raw feature values without transformation. Leave-one-site-out cross-validation yielded mean AUROC of 0.64. On the hidden validation set, AUROC was 0.573. The most discriminative features were arousal index, fragmentation index, periodic limb movement index, and reduced slow-wave sleep, consistent with known markers of preclinical cognitive decline. Notably, an initial 82-feature ensemble scored only 0.461 on validation, while reducing to ten features with logistic regression improved to 0.573. This demonstrates that minimal, interpretable features with strong regularization outperform complex models for cross-site cognitive impairment prediction, highlighting site-robust feature engineering over model complexity in multi-center sleep studies.