A multimodal sleep foundation model for disease prediction.

SleepFM accurately predicts 130 conditions with a C-Index of at least 0.75 from a single night of sleep.

• Bonferroni-corrected P < 0.01 for all 130 conditions
• Predictions include all-cause mortality (C-Index 0.84), dementia (0.85), myocardial infarction (0.81), heart failure (0.80), chronic kidney disease (0.79), stroke (0.78), and atrial fibrillation (0.78)
• The model was trained on over 585,000 hours of PSG recordings from approximately 65,000 participants across several cohorts
• Disease risk predictions are derived from latent sleep representations that capture physiological and temporal structure of sleep

SleepFM was trained using a new contrastive learning approach that accommodates multiple PSG configurations.

• The model is described as a multimodal sleep foundation model
• Training data came from a curated dataset across several cohorts totaling approximately 65,000 participants
• The contrastive learning approach was designed to address challenges in standardization, generalizability, and multimodal integration
• The approach enables scalable, label-efficient analysis of polysomnography data

SleepFM demonstrates strong transfer learning performance on the Sleep Heart Health Study dataset, which was excluded from pretraining.

• The Sleep Heart Health Study dataset was held out entirely from the pretraining process
• Transfer learning performance on this out-of-distribution dataset was described as 'strong'
• This demonstrates the model's generalizability to novel datasets not seen during training
• Transfer learning capability supports the model's potential for broad clinical application

SleepFM performs competitively with specialized sleep-staging models on common sleep analysis tasks.

• Achieved mean F1 scores of 0.70–0.78 for sleep staging
• Achieved accuracy of 0.69 for classifying sleep apnea severity
• Achieved accuracy of 0.87 for classifying sleep apnea presence
• Compared favorably against specialized models U-Sleep and YASA
• These results were achieved despite SleepFM being a general-purpose foundation model rather than task-specific

Polysomnography is underutilized as a clinical tool due to challenges in standardization, generalizability, and multimodal integration.

• PSG is described as 'the gold standard for sleep analysis'
• PSG captures rich physiological signals across multiple modalities
• Prior to SleepFM, these challenges limited the broader application of PSG data for disease prediction
• The complex relationship between sleep and disease remains 'poorly understood' according to the authors

SleepFM produces latent sleep representations that capture the physiological and temporal structure of sleep.

• These representations are described as enabling 'accurate prediction of future disease risk'
• The representations are derived from a single night of PSG recording
• The model integrates multiple PSG signal modalities into unified latent representations
• The authors describe this as the model learning 'the language of sleep from multimodal sleep recordings'