Can a sleep-study ECG predict future heart risk?

The answer: promising for some outcomes, but not yet a care decision

A neural network applied to the single-lead electrocardiogram recorded during an overnight sleep study identified patients at higher long-term risk of atrial fibrillation, heart failure and death in two external US hospital cohorts. Its extra value was less convincing for stroke and myocardial infarction. The result suggests that an ECG already collected during polysomnography may contain useful risk information beyond routine clinical variables and standard sleep measurements.

It does not show that the system diagnoses hidden disease, selects a treatment or improves patient outcomes. The study looked backwards through health records, and the researchers did not put the score into a live pathway. For a clinician, the model is therefore a candidate for further validation and prospective risk-stratification trials, not a reason to start medication or label a patient with cardiovascular disease.[1]

The denominator: three sleep-lab cohorts and 38,195 patients

The investigators used three datasets assembled through the Human Sleep Project. Massachusetts General Hospital supplied the development data. The paper reports 17,328 MGH subjects with eligible ECG recordings; after outcome eligibility and splitting, 12,381 patients were used for training and 3,428 for internal validation, giving the 15,809-person MGH outcome cohort described in the abstract. External testing used 9,810 patients from Emory University Hospital and 12,576 from Beth Israel Deaconess Medical Center. Across the outcome analyses, the denominator was 38,195 patients.

Event counts were substantial but varied by endpoint. In the MGH training group there were 1,208 incident atrial-fibrillation events, 987 strokes, 341 myocardial infarctions, 1,099 heart-failure events and 827 deaths. The MGH validation group recorded 368, 312, 82, 313 and 237 respectively. Emory recorded 605 atrial-fibrillation events, 781 strokes, 464 myocardial infarctions, 832 heart-failure events and 498 deaths; Beth Israel recorded 409, 180, 130 and 520 events for the four cardiovascular outcomes.

These were sleep-clinic populations, not a random sample of the public. Referral for polysomnography often reflects suspected sleep apnoea, insomnia, unusual sleepiness or other clinical concerns. That makes the study relevant to sleep services but limits any claim that the same risk relationships or calibration will hold in primary care, wearable-device users or people without a reason for specialist sleep testing.[1]

What the model saw and what it was compared with

The team trained a deep residual neural network with an attention mechanism on the overnight single-lead ECG and expert-labelled sleep stages. The network had previously been trained for arrhythmia detection and was then fine-tuned to predict future outcomes. Follow-up extended to as much as ten years at Emory and seven years at Beth Israel. People with a prior record of the outcome were excluded, and atrial-fibrillation diagnoses made within one month of the sleep study were excluded from that incident analysis.

The comparison was not simply model versus nothing. Cox models included age, sex, body-mass index, smoking, hypertension, diabetes and sleep-study measures including the apnoea-hypopnoea index, arousal index, limb movements, sleep stages and sleep efficiency. The researchers then tested whether adding the neural-network output improved prediction beyond those variables. That is a more meaningful question than reporting the network in isolation.

Outcomes were identified from electronic records using ICD-9 and ICD-10 diagnosis codes. That makes a study of this size practical, but coding is an imperfect proxy for clinical confirmation and may miss disease treated elsewhere. The model also may be learning signals correlated with existing but undocumented illness rather than detecting a modifiable pathway years before disease begins.[1]

The strongest gains were for atrial fibrillation, heart failure and death

For each standard-deviation increase in the neural-network score, adjusted hazard ratios at Emory and Beth Israel were 2.03 and 2.72 for atrial fibrillation, 1.69 and 2.33 for heart failure, and 1.82 and 1.65 for all-cause mortality. Associations were smaller for stroke, at 1.19 and 1.40. For myocardial infarction they were 1.38 at Emory and 1.10 at Beth Israel; the Beth Israel confidence interval of 0.86 to 1.40 included no association.

Discrimination tells the same qualified story. At Emory, adding the neural score changed the concordance index from 0.70 to 0.75 for atrial fibrillation, 0.72 to 0.75 for heart failure and 0.74 to 0.78 for death. At Beth Israel the changes were 0.77 to 0.81, 0.74 to 0.78 and 0.80 to 0.81. For stroke, the changes were 0.71 to 0.71 and 0.75 to 0.76; for myocardial infarction, 0.71 to 0.72 and 0.79 to 0.79.

A hazard ratio is not an individual probability, and a higher concordance index does not show that treatment based on the score helps anyone. The modest or absent incremental gains for stroke and myocardial infarction are also important. A single headline saying the model predicts cardiovascular disease would hide meaningful variation between endpoints.[1]

What this could change for patients and clinicians

If prospective validation succeeds, a sleep service might use the score to identify patients who merit a closer cardiovascular history, a conventional risk assessment or targeted rhythm monitoring. That could be especially relevant for atrial fibrillation, which can be intermittent and silent. But an alert would need a defined follow-up route; otherwise it may create anxiety and additional appointments without improving care.

A safe pathway would separate screening from diagnosis. Clinicians would review symptoms, existing conditions and ordinary risk factors, then decide whether a standard ECG, ambulatory rhythm monitor or another established test is appropriate. Services would need to track false positives, missed cases, waiting time, workload and whether alerts improve outcomes across age, sex and racial or ethnic groups. The publication does not provide evidence for those operational choices.

Patients should also know when a recording collected for sleep assessment is reused to estimate other risks. Consent, data retention and the possibility of incidental findings need clear governance. A low score must not override symptoms or block ordinary assessment, and a high score should not become a diagnosis in the medical record without confirmation.[1]

Evidence limits, funding and interests

The retrospective design cannot show that the model causes better care. Diagnosis codes were not clinically adjudicated, cause of death was unavailable, and the records lacked biomarkers that might explain or improve predictions. The authors could not pair the original sleep-study ECG with an ECG at later diagnosis. Undiagnosed atrial fibrillation at baseline remains possible; only five 30-second segments were manually reviewed in recordings linked to future atrial fibrillation.

The National Heart, Lung, and Blood Institute supported the work through grant R01 HL161253. The authors reported no financial conflicts. They disclosed that one author, B.M.W., was a co-founder, adviser and consultant to Beacon Biosignals and held equity in the company. Code was made public, while patient-level data remained restricted. Independent testing would help determine whether the reported performance depends on local equipment, coding or referral patterns.

The assessment would strengthen with a prospectively registered, multicentre study that freezes the model before testing, reports calibration and decision-curve results, and compares it with simpler ECG and clinical baselines. Most importantly, a trial should test whether score-guided follow-up finds treatable disease earlier or improves outcomes without disproportionate investigations. It would weaken if calibration drifts, simpler models perform as well, or extra testing produces harm without measurable benefit.[1]