Detecting cognitive impairment (CI) from sleep polysomnography (PSG) is a complex task requiring the integration of diverse physiological signals. We developed a computational approach based on extreme gradient boosting (XGBoost) and principal component analysis (PCA) to predict future diagnoses of CI. Our team, CMCR UJ CM, developed an intermediate feature fusion pipeline leveraging multimodal foundation models: HuBERT for electrocardiogram (ECG), a BENDR-based convolutional architecture for electroencephalogram (EEG), and SleepFM for combined PSG. To combat overfitting on high-dimensional embeddings, we applied PCA to reduce each signal modality to 24 principal components. These were fused with tabular clinical and demographic data and classified using a heavily regularized XGBoost model. Using a 5-fold cross-validation on the public training set, our approach achieved an area under the receiver operating characteristic curve (AUROC) of 0.820 and an area under the precision-recall curve (AUPRC) of 0.380. On the hidden validation set, the team was unranked due to deployment issues. This article is part of `Screening for Cognitive Impairment During Sleep Studies: The George B.\ Moody PhysioNet Challenge 2026'.