SleepLM: Language Models for Sleep

🔑 Enhanced Key Takeaways

• •SleepFM, a distinct multimodal sleep foundation model from Stanford, was trained on 585,000 hours of PSG data from 65,000 participants and predicts 130 diseases like dementia (C-Index 0.85) and heart failure (0.80) from one night of sleep.[1][3]
• •PFTSleep from Mount Sinai analyzes full-night sleep using transformer architecture on 1 million hours of data, outperforming traditional methods in sleep stage classification via self-supervision.[2]
• •SleepFM employs leave-one-out contrastive learning to harmonize PSG modalities like EEG and ECG, enabling reconstruction of missing channels and strong zero-shot disease forecasting.[3][4]

🛠️ Technical Deep Dive

• •SleepFM divides PSG data into 5-second segments akin to tokens, processing multimodal signals (EEG, ECG, EMG, airflow) with leave-one-out contrastive learning to reconstruct hidden modalities.[3][4]
• •PFTSleep uses a patch foundational transformer to analyze entire 8-hour nights of brain waves, muscle activity, heart rate, and respiration, trained self-supervised on 1,011,192 hours without 30-second epochs.[2]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline