Introduction: Sleep patterns offer a vital physiological window into cognitive health. Deep learning can extract critical information from multimodal Polysomnography (PSG) to predict cognitive impairment. We adapt SleepFM to classify cognitive impairment from seven PSG modalities, exploring both contrastive training from scratch and fine-tuning with publicly available pretrained weights.
Methods: This study trains model from PSG recordings across seven modalities (EEG, EOG, chin EMG, leg EMG, ECG, respiratory, SpO2). Our pipeline has three stages: (1) representation learning, where a SleepFM-inspired encoder tokenizes raw signals into patch embeddings via a CNN and models cross-modal interactions with a Transformer; (2) embedding extraction, where representations are precomputed; and (3) classification, where a temporal head predicts cognitive impairment. In the first approach, the encoder is trained from scratch with a leave-one-out multimodal contrastive objective, and a BiLSTM head combined with demographics (age, sex, BMI). In the second, we use frozen SleepFM pretrained weights and explore several improvements to the classification head: replacing BiLSTM with xLSTM, incorporating CAISR-derived sleep annotations, demographics imputation, focal cross-entropy loss, modality dropout, and annotation dropout. Final predictions are produced via 5-fold cross-validation stratified by site and label with AUROC-based checkpointing, ensembled by averaging fold probabilities.
Results: On the leaderboard, the from-scratch approach achieved 0.621 AUROC, while frozen SleepFM weights reached 0.59. In cross-validation, the first approach yielded 0.63 AUROC (v1 Kaggle data), while the second reached 0.82 (v2 data). Although frozen embeddings with xLSTM and feature engineering outperformed from-scratch methods during cross-validation, these gains did not transfer to the leaderboard, suggesting significant overfitting.
Conclusions: The frozen encoder likely overfits because it was not pretrained on all challenge modalities;training from scratch generalized better. Our approach can be improved by combining from-scratch pretraining with the xLSTM head, adopting leave-one-site-out cross-validation, and other self-supervised pretraining objectives.