Introduction. Polysomnography (PSG) captures multi-modal physiological signals during sleep that may encode early signatures of neurodegeneration. Predicting future cognitive impairment from multi-institutional PSG is difficult due to cross-site heterogeneity in channel availability, sampling rates, and recording equipment, and a large prevalence shift between the artificially balanced training set (~50%) and real-world deployment conditions (5–15%). Methods. We propose a multimodal feature fusion framework combining deep and handcrafted representations. SleepFM, a pretrained, frozen SetTransformer encoder, extracts 128-dimensional temporal embeddings from four physiological streams (EEG, EOG, EMG, and cardiorespiratory signals) independently, yielding a 512-dimensional deep representation per recording. These are concatenated with 100 tabular features comprising patient demographics, CAISR automated annotation statistics (sleep stage distributions, arousal, respiratory, and limb event rates), and handcrafted Hjorth and per-band spectral descriptors. A focal-loss-regularized multilayer perceptron (MLP) trained with AdamW optimization learns from the resulting 612-dimensional input. Leave-one-site-out (LOSO) cross-validation across the three training sites (S0001, I0002, I0006; n=622) provides unbiased cross-site generalization estimates. Results. Team NeruoAI achieved a LOSO cross-validation AUROC of 0.65 on the training set. Our official validation AUROC on the held-out set was 0.61. Conclusion. Combining a pretrained sleep foundation model with automated annotation statistics and handcrafted features yields competitive cross-site generalization for cognitive impairment prediction. Future work will introduce modality-specific importance weighting and a Cox proportional hazards objective for time-to-diagnosis modeling. Most ambitiously, we plan to pretrain PSG-FM, a novel PSG-specific foundation model extending LaBraM's Vector-Quantized Neural Spectrum Prediction (VQ-NSP) framework to all PSG modalities, EEG, EOG, EMG, ECG and respiratory, using modality-aware spatial embeddings pretrained on public sleep databases to improve cross-site robustness.