A wrist-worn single-lead ECG now flags 13 structural heart conditions, supervised by routine echo reports

A single-lead ECG waveform can now yield per-subtype probabilities for 13 structural heart diseases spanning myocardial, chamber, valvular and great-vessel pathology. Most prior AI-ECG screening models were built for 12-lead configurations, incompatible with the single-lead sensors in consumer wearables, and the existing single-lead contrastive work was validated only inside one health system and covered few coarse subtypes. Validation ran in three tiers: an internal test set of 4,078 patients at Peking University People's Hospital, a temporal set of 4,523 patients split at November 2019, and a geographically independent external set of 16,621 patients at the Second Hospital of Tianjin Medical University about 140 km away, totaling 25,222 patients.

The model treats physician-authored echocardiography reports as its supervisory signal, removing the need for per-condition manual annotation. The prevailing supervised route relies on large, condition-specific, manually labeled datasets that are expensive to assemble and coarse in diagnostic granularity. A single-lead ECG encoder is first pre-trained by aligning ECG with report representations in a shared space, jointly with an ECG-only self-supervised contrastive objective, after which the encoder is frozen and only a linear classification head is trained to output 13 subtype probabilities.

Model outputs track quantitative echocardiographic measurements monotonically, letting the waveform act as a continuous digital biomarker. Binary diagnostic labels only say whether disease is present and cannot express the severity of structural compromise. Predicted probabilities correlate negatively with left ventricular ejection fraction and positively with interventricular septal thickness and left ventricular posterior wall thickness, and echocardiographic measurements change monotonically across quartiles of predicted probability.

The demonstrated scope is a retrospective, multicenter diagnostic accuracy study: development and validation cohorts were of East Asian descent and drawn from patients clinically referred for echocardiography, so the reported AUROC values describe performance in these referral populations. The intended setting is opportunistic single-lead ECG screening outside the hospital, in communities and primary care, used to rank patients for echocardiographic referral. Community number-needed-to-screen estimates were calculated from literature-derived prevalence and have not been validated against observed community screening outcomes.

The authors note that the supervisory signal depends on the diagnostic consistency of physician-authored reports; contrastive learning can tolerate some noise, but sensitivity to interval cardiac changes within the plus or minus 10-day matching window has not been assessed through dedicated ablation studies. Missing sex data, 41.5% in specific cohort subsets, may limit the robustness of sex-stratified bias analyses. Discrimination was limited for atrial pathologies, potentially reflecting the difficulty of capturing subtle P-wave morphology from a single lead compared with the more prominent QRS complex, and hospital-recorded Lead I is only a proxy for wearable recordings that does not capture motion artifacts or signal-quality variation. Performance declined monotonically across most categories as age increased from under 50 to 75 and above, which the authors attribute to age-associated electrophysiological confounders such as conduction delays and atrial fibrillation.