As part of the George B. Moody PhysioNet Challenge 2026, our team, insomniac111, developed an open-source machine learning pipeline to predict future cognitive impairment from polysomnography (PSG) data. Our primary model uses TabPFN V2, a pretrained transformer-based tabular foundation model, which we adapt to the challenge training set using transfer learning. The training data include 622 PSG recordings from three sources, with physiological signals, algorithmic annotations, human annotations, demographic metadata, and diagnostic labels. From these data, we extract 243 engineered features spanning demographics, per-channel physiological summary measures, EEG spectral power, ECG heart rate variability, enhanced SpO2 descriptors, algorithmic and human annotation features, and stage-conditional cross-modal features. We apply site-aware cross-validation to evaluate generalization across sources and select the final model based on out-of-fold AUROC. In preliminary testing, our approach achieved a challenge score of 0.660 (AUROC between diagnosed and healthy patients). We also explored alternative approaches, including tree-based XGBoost models and 1D ConvNeXt models trained on raw EEG signals. These results suggest that pretrained tabular foundation models, combined with multimodal PSG feature engineering, are a promising approach for early prediction of cognitive impairment.