A sleep study already records the heart. AI has found a second use for the signal

Among the signals collected during an overnight sleep study is an electrical record of the heart. It accompanies the measurements used to investigate sleep disorders, but often receives little analysis. Researchers have now extracted another kind of information from it: clues to a patient’s cardiovascular future.[2]

The study, published in Sleep on 9 October, developed the model using 15,809 Massachusetts General Hospital patients and evaluated it in independent groups of 9,810 patients at Emory University Hospital and 12,576 at Beth Israel Deaconess Medical Center. It combined a single ECG lead with expert-labelled sleep stages.[2]

Adding its output to established clinical and sleep-related risk factors improved prediction of atrial fibrillation, heart failure and death in both external groups. Maximum follow-up was ten years at Emory and seven at Beth Israel. Stroke and heart-attack predictions gained no significant improvement in the reported ROC analysis.[1]

One channel in a much richer recording

The Human Sleep Project’s documentation shows why a sleep study is unusually useful for this work. Its recordings include brain activity, eye movement, muscle activity, breathing, oxygen saturation and an ECG channel. Certified technologists label sleep stages in consecutive 30-second intervals. That provides a time-aligned account of the state the patient was in while the heart signal was being collected.[3]

The model therefore receives more than a long succession of heartbeats. It receives context for interpreting a night’s changing patterns. For a reader accustomed to thinking of medical AI as a new test, this is a different proposition: seek additional information in a measurement already collected for another purpose.

The authors’ public code illustrates the ingredients. The inference pipeline takes an ECG signal, its sampling frequency and sleep-stage annotations, and returns separate outputs for the five outcomes studied. The code requires at least an hour of ECG data. That is a research interface for extended recordings, a substantial difference from interpreting a brief spot check.[4]

What improved—and by how much

Harrell’s concordance index measures how well a model ranks comparable patients by event timing. Values nearer one indicate better ranking. The paper reports these point estimates:[1]

The comparison matters more than a model’s ability to split people into broadly higher- and lower-risk groups. Age, hypertension and other familiar factors already do that. The researchers adjusted for those factors, alongside measures such as sleep-apnoea severity and sleep efficiency. NIH’s account reports that the ECG-based output retained predictive information after those adjustments.[2]

The gains also differ by outcome and hospital. For mortality, the concordance increase is 0.04 at Emory and 0.01 at Beth Israel; for atrial fibrillation, it is 0.05 and 0.04. Those simple differences, calculated from the table, are a more faithful description than assigning the system one headline accuracy figure.

A ranking improvement has a specific use: it could help identify whom to assess more closely. It leaves a separate clinical question to answer—what should happen after a high score? A useful screening pathway needs a sensible follow-up action and evidence that the resulting decisions help patients.

The opportunity is in the existing test

The analyses used each patient’s latest sleep study, excluding prior diagnoses of the condition being predicted. For atrial fibrillation, they also excluded diagnoses within one month afterwards. Cardiovascular diagnoses came from health-record codes in hospital-referred sleep patients.[1]

The public repository accepts missing sleep-stage annotations by assigning an unscored value. That software convenience should not be confused with evidence that an unstaged home recording reproduces the published results. A patch-based application would need its own assessment of recording quality, available sleep context and performance.[4]

The near-term opportunity is easier to picture inside the sleep service itself. If further clinical work establishes a useful pathway, one night’s existing recording could inform two conversations: what disturbed the patient’s sleep, and whether their heart deserves closer attention. The second conversation begins with a signal the first test already captured.

Report an error or suggest a correction