A new artificial intelligence (AI) model, named SleepFM, has demonstrated the ability to predict the onset of various diseases using sleep data. Researchers at the Stanford Sleep Medicine Center and collaborators developed and tested the model.### Model Development and Initial TestingSleepFM was trained using polysomnography data from approximately 35,000 patients, aged 2 to 96, recorded at the Stanford Sleep Medicine Center between 1999 and 2024. Following its training phase, the model was fine-tuned for specific tasks.Initially, SleepFM was evaluated on standard sleep analysis tasks, including the classification of sleep stages and the diagnosis of sleep apnea severity. In these tests, the model achieved performance comparable to or exceeding current state-of-the-art models.### Disease Prediction CapabilitiesResearchers then applied SleepFM to predict future disease onset. This involved pairing the polysomnography data with up to 25 years of electronic health records from the same patient cohort. The model analyzed over 1,000 disease categories within these health records.SleepFM identified 130 disease categories that could be predicted from patient sleep data with a C-index (concordance index) of reasonable accuracy. The C-index measures a model's predictive performance, indicating its ability to predict which of two individuals will experience an event first. A C-index of 0.8 signifies that the model's prediction was concordant with actual outcomes 80% of the time.The model exhibited strong predictive performance for several categories, including cancers, pregnancy complications, circulatory conditions, and mental disorders, achieving C-indices above 0.8. Specific predictions included:* Parkinson's disease (C-index 0.89)* Dementia (C-index 0.85)* Hypertensive heart disease (C-index 0.84)* Heart attack (C-index 0.81)* Prostate cancer (C-index 0.89)* Breast cancer (C-index 0.87)* Death (C-index 0.84)Models with C-indices around 0.7 have previously been utilized in clinical settings for applications such as predicting patient responses to cancer treatments.### Future ResearchThe research team is exploring methods to enhance SleepFM's predictive capabilities, potentially through the integration of data from wearable devices. Efforts are also underway to interpret the specific patterns the model identifies when making disease predictions. Researchers noted that while heart signals were more prominent in heart disease predictions and brain signals in mental health predictions, the combination of all data modalities yielded the most accurate results. Discrepancies between different physiological channels, such as a brain showing sleep patterns while the heart exhibits awake patterns, were identified as significant for disease prediction.The study was supported by funding from the National Institutes of Health, Knight-Hennessy Scholars, and Chan-Zuckerberg Biohub.