0341 A Large-Scale EEG Foundation Model Enables Label-Efficient, Multi-Device Sleep Analytics

Abstract Introduction Contrastive self-supervised learning has advanced pretraining in computer vision and, more recently, EEG. However, most prior EEG work has focused on BCI or high-density recordings. Here, we develop a sleep-focused foundation model trained on a large heterogeneous corpus spanning PSG, 10-20 EEG and dry-electrode EEG (Waveband) recordings to improve label efficiency across diverse sleep analytics tasks. Methods We pretrained an ~5M-parameter encoder on >128,000 overnight sleep studies (PSG and dry-EEG) totalling ~1.8 million hours of EEG from 14 different datasets. A contrastive objective encouraged temporally and/or contextually similar windows to be closer in latent space than dissimilar ones. The pretrained model was fine-tuned on four downstream tasks: (1) sleep staging on PSG, (2) sleep staging on dry-EEG headband data, (3) EEG artifact detection on dry-EEG, and (4) arousal detection on PSG. For each task, we varied the number of fine-tuning recordings from 10 to the full labelled set. Performance was compared against a competitive fully-supervised baseline (RobustSleepNet or SimpleSleepNet) trained from scratch on the same number of recordings to quantify label-efficiency gains from pretraining. Results Across tasks, the pretrained model matched or approached state-of-the-art performance while requiring roughly 5–10× fewer labelled recordings. With only 10 fine-tuning recordings, it achieved Cohen’s κ = 0.716 for PSG sleep staging, κ = 0.730 for dry-EEG sleep staging, F1 = 0.968 for artifact detection, and F1 = 0.71 for arousal detection. On the same amount of data, state-of-the-art baselines reached Cohen’s κ = 0.542 for PSG sleep staging, κ = 0.541 for dry-EEG sleep staging, F1 = 0.948 for artifact detection, and F1 = 0.65 for arousal detection. Conclusion Contrastive self-supervised pretraining on a large, heterogeneous sleep EEG corpus yields substantial label efficiency and generalization across downstream tasks and devices, supporting foundation models as a practical and scalable approach for algorithm training on data-limited cohorts and tasks. Support (if any)