As part of the 2026 Moody PhysioNet Challenge, our team developed a multimodal foundation model to predict cognitive impairment from polysomnography (PSG) data. To address the high dimensionality in sleep recordings, we used state-of-the-art embeddings, including for neurophysiological (EEG) and respiratory signals. To establish baselines, we evaluated performances across using single modality embeddings with cross-validation, with AUROC scores of 0.54 for demographics, 0.58 for breathing embeddings, and 0.65 for EEG embeddings. Using a multimodal fusion approach of masked supervised training, our model attained a substantially higher AUROC of 0.737 on the hidden validation set, ranking our team 5th out of 243 submissions. These results show that our multimodal fusion approach effectively extracts the different information across the signals, leading to a substantial performance gain over individual baselines. For the official phase, we intend to transition from static patient-level embeddings to a full Transformer-based architecture to capture temporal information and enhance cross-modal synchronization across all the PSG modalities.