Predicting cognitive impairment before clinical diagnosis may support earlier assessment, monitoring, and intervention. As team OOM_finder, we investigated as part of the George B. Moody PhysioNet Challenge whether overnight polysomnography (PSG) contains information predictive of cognitive impairment subsequently diagnosed 1--6 years later. We compared three levels of PSG representation: strongly regularized logistic regression using demographic and sleep-summary features, a hierarchical transformer operating on temporally resolved algorithmic sleep annotations, and a pretrained multimodal encoder of raw physiological signals. The raw-signal approach achieved the strongest internal discrimination with an age-conditioned AUROC of 0.804. However, this advantage did not transfer to the official hidden validation set, where the compact sleep-summary model achieved the highest age-conditioned AUROC of 0.703 and a reward of 0.157. The annotation-based and raw-signal models achieved hidden-validation age-conditioned AUROCs of 0.562 and 0.612, respectively. These results suggest that increased representational complexity does not necessarily improve cross-site generalization and that compact, clinically interpretable sleep features may help predict future cognitive impairment.