Aims: As part of the 2026 PhysioNet Challenge, team "no_sleep" developed a framework for the early prediction of cognitive impairment using multimodal physiological signals (EEG, EOG, ECG, EMG, and respiration) and demographic factors (age, sex, and BMI). Early identification is essential for initiating medical treatments and behavioral interventions that can delay cognitive decline and ease the socioeconomic burden on patients, families, and the healthcare system. Methods: We evaluated two distinct methodological approaches. The first utilized a LightGBM gradient boosting model trained on hand-crafted statistical and time-domain metrics (Hjorth Activity, Mobility, and Complexity; Zero Crossing Rate; and RMS energy), automated sleep stage annotations and clinical event markers, and patient demographic. The second approach employed multimodal transfer learning via the SleepFM framework. We adopted a Hierarchical Set Transformer as a compact backbone encoder (4.8M parameters) to process raw polysomnography (PSG) signals. This architecture utilizes spatial attention pooling across signal channels and temporal pooling across epochs to generate 128-dimension latent embeddings. To address missing physiological recordings, we incorporated binary modality presence flags, fusing the resulting embeddings with demographic data to train the classifier. Results: In the unofficial challenge phase, the LightGBM approach achieved a Challenge Score of 0.509 on the hidden validation set. The Multimodal Transfer Learning approach outperformed the light gradient boosting model, achieving a macro-AUC of 0.65 and an AUPRC of 0.63 on a held-out test split from the public challenge data, with similar accuracy and F1-score of 0.58. Conclusion: Our results demonstrate that multimodal transfer learning using foundation model architectures shows significant promise for the early detection of cognitive impairment. Future work will focus on scaling the training data through multi-center sources and refining the transformer architecture to enhance predictive performance.