Overnight polysomnography ECG model predicts cardiovascular outcomes in external validation cohorts

Researchers evaluated whether single-lead electrocardiograms routinely recorded during polysomnography could support cardiovascular risk prediction. A neural network initially trained for arrhythmia detection was fine-tuned using ECGs and sleep-stage information from 15,809 Massachusetts General Hospital patients to predict 10-year atrial fibrillation, stroke, myocardial infarction, heart failure and mortality risk. External validation included 9,810 Emory University Hospital patients and 12,576 Beth Israel Deaconess Medical Center patients; outcomes were identified through electronic health record diagnostic codes.

Associations persisted after adjustment for demographic, cardiovascular and sleep-related factors. Per standard-deviation increase in model output, hazard ratios at Emory and Beth Israel, respectively, were 2.03 and 2.72 for atrial fibrillation, 1.19 and 1.40 for stroke, 1.69 and 2.33 for heart failure, and 1.82 and 1.65 for all-cause mortality. Myocardial infarction hazard ratios were 1.38 (95% CI 1.22–1.55) and 1.10 (0.86–1.40), with the latter confidence interval including no association.

These findings suggest that routinely collected sleep-study ECGs contain prognostic information beyond measured risk factors and sleep characteristics. However, the available abstract does not report discrimination, calibration or clinical utility, and associations alone do not establish benefit from model-guided care. Diagnostic-code-based outcomes and restriction to patients undergoing polysomnography also warrant consideration when assessing applicability.