Summary
A medRxiv preprint reports that a fine-tuned ECG foundation model detected rheumatic mitral stenosis across 12-, 6- and 1-lead ECG configurations. Its external-validation AUROC ranged from 0.77 to 0.82, supporting further study of ECG screening and echocardiography triage.
A medRxiv preprint reports that a fine-tuned electrocardiogram (ECG) foundation model identified patterns associated with rheumatic mitral stenosis (RMS) across standard and reduced-lead ECG configurations. The model’s area under the receiver operating characteristic curve (AUROC) was 0.77–0.82 on external validation, depending on the number of leads used. The authors describe the results as support for further evaluation of ECG screening and echocardiography triage in settings where access to echocardiography is limited.
RMS is narrowing of the mitral valve associated with rheumatic heart disease. An ECG records the heart’s electrical activity; it does not directly image the valve. The approach therefore uses patterns in the ECG as indirect signals that may help identify people who need further assessment.
What the model was tested on
The researchers fine-tuned an ECG foundation model using 804 RMS-positive ECGs per fold. The labels were derived indirectly from diagnostic codes, rather than being described in the abstract as direct measurements of valve severity. The study used openly available human ECG data from the Harvard-Emory ECG Database and evaluated performance with 12-lead, 6-lead and 1-lead configurations.
For internal evaluation, the reported AUROCs were 0.89 with 12 leads, 0.87 with 6 leads and 0.87 with 1 lead. In external validation, they were 0.82, 0.80 and 0.77, respectively. The authors also report comparable performance across demographic subgroups.
AUROC measures how well a model ranks positive cases above negative cases across possible decision thresholds. It is not the same as the percentage of patients correctly diagnosed. The abstract reports AUROC values but not the results for a specific screening threshold.
A possible route to triage
The results suggest that reduced-lead ECGs may warrant further study as a way to flag people for echocardiography, rather than as a replacement for imaging. That could be relevant where echocardiography access is constrained, the motivation identified by the researchers.
This is a preprint feasibility study using indirectly labelled ECGs. Its findings support further evaluation; they are not evidence that the model is ready for clinical screening. Prospective assessment would be needed to establish how it performs in the intended care settings and how well a chosen threshold identifies people who should receive echocardiography.