In a study published in Nature Medicine, researchers have developed SleepFM, a multimodal sleep foundation model that uses overnight sleep data to predict long-term disease risk. The study highlights sleep as a powerful yet underutilised source of health information for early diagnosis and risk stratification.
From Sleep Disorders to Systemic Health Insights
Sleep disorders affect millions worldwide and increasingly serve as both contributors to and early indicators of systemic diseases. Polysomnography (PSG) remains the gold standard for sleep assessment, capturing detailed physiological signals. However, earlier machine learning approaches largely focused on single diseases or limited sleep metrics. Consequently, much of PSG’s complexity remained untapped.
Building SleepFM Using Large-Scale PSG Data
To address this gap, researchers developed SleepFM using PSG data from four major cohorts—BioSerenity, MrOS, MESA, and Stanford Sleep Clinic—covering nearly 65,000 participants and 585,000 hours of sleep recordings. The team pretrained the model using self-supervised contrastive learning, while reserving the Sleep Heart Health Study (SHHS) dataset for external validation.
Strong Performance on Benchmark Tasks
Following pretraining, researchers fine-tuned SleepFM on standard benchmarks. The model estimated chronological age with a mean absolute error of 7.33 years, achieved an AUROC of 0.86 for sex classification, and demonstrated competitive performance in sleep staging and apnea detection. Although specialised models occasionally outperformed SleepFM on specific tasks, the foundation model delivered consistently strong results across datasets.
Predicting Long-Term Disease Risk
As reported by news-medical.net, next, the team linked sleep data with electronic health records to predict over 1,000 disease categories. SleepFM showed robust predictive performance across neurological, cardiovascular, oncological, and mental health conditions. Notably, it achieved an AUROC of 0.93 for Parkinson’s disease, 0.88 for hypertensive heart disease, and strong prediction for cancers such as prostate and breast cancer. Importantly, these results reflect statistical risk stratification rather than causation.
Generalisation, Mortality Prediction, and Clinical Promise
Moreover, SleepFM generalised well across time and clinical sites, outperforming supervised baselines and accurately predicting all-cause mortality (AUROC 0.85). Overall, the study demonstrates that AI-powered sleep models like SleepFM could complement existing risk tools and enable earlier disease detection. Future research may further enhance impact by integrating sleep data with imaging, genomics, and health records.