The heart signal hidden inside a night of sleep
An externally tested AI model finds predictive information in ECG recordings already collected during sleep studies.
A sleep study records more than breathing. It often contains a single-lead electrocardiogram—a stream of cardiac information collected while someone is asleep. A new modeling study asks whether that already-recorded signal can reveal more about future cardiovascular risk.
This is an appealing idea because it begins with data already obtained for another clinical purpose. But using an existing measurement more efficiently is not the same as proving that an extra risk score improves care.
How the model was tested
Researchers fine-tuned a deep neural network on overnight ECG and sleep-stage data from 15,809 Massachusetts General Hospital patients. They then evaluated it in separate cohorts at Emory, with 9,810 patients, and Beth Israel Deaconess, with 12,576.
The model’s output was associated with later atrial fibrillation, stroke, heart failure and mortality after adjustment for several conventional risk factors and sleep characteristics. For example, a one-standard-deviation increase in the output was associated with atrial-fibrillation hazard ratios of 2.03 and 2.72 in the two external cohorts.
Those ratios describe relationships between a score and later outcomes. They are not reductions in risk caused by using AI. Not every endpoint was equally consistent: the myocardial-infarction estimate in one validation cohort had a confidence interval crossing one.
Prediction is only part of the job
External validation is valuable, but clinical usefulness also requires reliable calibration, appropriate thresholds and a plan for the person receiving a high-risk result. A model can identify a statistical pattern without telling a clinician which action would improve that person’s health.
The study used outcomes derived from health records, including diagnostic codes. Differences in coding, referral patterns and the populations sent for sleep testing can affect generalizability. This is not evidence that every person needs a sleep study to screen for heart disease.
What would make the next study stronger
The next useful comparison would test what changes when clinicians receive the model’s result: appropriate evaluation, treatment decisions, patient experience, unnecessary testing and ultimately clinical outcomes. The burden of a false alarm belongs beside the possible benefit of earlier recognition.
The opportunity is not merely to collect more data. It is to discover whether information already present can answer a better question. That is a promising role for AI—as long as the pathway from prediction to useful care remains visible.
Sources and evidence
Prediction of cardiovascular outcomes using electrocardiography from overnight polysomnography
October 9, 2026 · Retrospective prediction-model development and external validation