Stanford AI Model Analyzes Sleep Data to Predict Disease Risk

Stanford Medicine researchers developed an artificial intelligence model, SleepFM, to analyze sleep study data and estimate the risk of various health conditions using physiological signals collected during sleep.

The model was trained on a large dataset, incorporating approximately 585,000 hours of polysomnography data from about 65,000 individuals. SleepFM utilized signals such as brain waves, heart rhythms, muscle activity, pulse, and breathing airflow to process and interpret sleep patterns.

To refine its accuracy, SleepFM was paired with long-term electronic health records from around 35,000 patients who received care at the Stanford Sleep Medicine Center over a 25-year period. This approach enabled the model to link sleep characteristics with subsequent health outcomes.

Researchers reported that SleepFM could predict the likelihood of developing 130 different health conditions based on a single night of sleep data. The model’s predictions included a range of diseases such as Parkinson’s disease, dementia, hypertensive heart disease, heart attack, prostate cancer, breast cancer, and all-cause mortality.

What the numbers show

• SleepFM achieved a concordance index of 0.89 for Parkinson’s disease and prostate cancer
• The model reached 0.85 for dementia and 0.87 for breast cancer
• Predictions for heart attack and death had indices of 0.81 and 0.84, respectively

SleepFM also demonstrated strong performance in classifying sleep stages and determining the severity of sleep apnea, matching or surpassing existing advanced models. Its training involved a method called leave-one-out contrastive learning, which allowed the integration of multiple types of physiological data.

The model’s predictive capabilities extended to conditions such as heart failure, chronic kidney disease, stroke, and atrial fibrillation. These outcomes were identified by analyzing the physiological signals recorded during sleep studies.

By leveraging a large and diverse dataset, SleepFM was able to generalize its findings across a wide range of health conditions. The model’s development involved both advanced machine learning techniques and extensive clinical data collected over several decades.

Stanford Medicine researchers stated that SleepFM represents a new approach in linking sleep patterns to long-term health risks using artificial intelligence. The model’s results are based on comprehensive data and established clinical records.

* This article is based on publicly available information at the time of writing.