Transfer Learning from a Multimodal Sleep Foundation Model for Cognitive Impairment Prediction Using Polysomnography and Clinical Features

Jingming Zhang1, Zijie Zhu1, Chaoyu Wei1, Zhibin Deng1, Ziqin Yuan1, Xue Luo1, Lei Yang1, Ziling Lan1, Yuqi Wang2, Pan Xia1
1Faculty of Information Engineering and Automation, Kunming University of Science and Technology, 2Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences


Abstract

The George B. Moody PhysioNet Challenge 2026 focused on predicting future cognitive impairment from overnight polysomnography (PSG). Our team, Kust_MeAI, proposed a transfer-learning framework combining a pretrained SleepFM foundation model, patient-level temporal modeling of multimodal PSG embeddings, and automated clinical features. PSG recordings were processed through a SleepFM-compatible pipeline with dynamic channel selection, normalization, and token-based segmentation, and CAISR-derived annotations were used to derive patient-level clinical features. A publicly released SleepFM model was used as a frozen multimodal encoder to extract representations from EEG, respiratory, ECG, and EMG signals. Multimodal embeddings were aggregated through spatial attention, bidirectional temporal modeling, and temporal attention pooling, then combined with demographic and clinical features for cognitive impairment prediction. We trained and evaluated the model using five-fold cross-validation. The proposed approach achieved an age-conditioned AUROC of 0.682 during cross-validation and 0.655 on the official hidden validation set, ranked 46th out of 102 teams.