For the George B. Moody PhysioNet Challenge 2026, Team Blue and Gold developed two approaches using polysomnography (PSG) to predict a subsequent cognitive impairment diagnosis obtained one to six years after the initial PSG. The PSG data presented several modeling challenges: the signals are heterogeneous, channels may be missing, sampling rates vary, and inputs include both time-series and participant-level data. The Frozen SleepFM Cox Model adapts a pretrained encoder to a single diagnosis-based outcome using a bidirectional recurrent risk head and a minibatch approximation to the Cox partial likelihood. A second approach uses XGBoost with 13 selected signal- and automated-annotation-derived features and a reward-tuned decision threshold. On the official hidden validation set, Frozen SleepFM achieved an age-conditioned area under the receiver operating characteristic curve (AUROC) of 0.698 and reward of 0.017; XGBoost achieved 0.674 and 0.049, respectively.