Can a Single ECG Trace Reveal How Well the Heart Pumps? AI Study Explores Post–Myocardial Infarction Follow-up

Follow-up after myocardial infarction involves understanding not only whether the heartbeat is regular, but also whether the heart can pump blood effectively. If a single ECG trace recorded at home could offer clues about left ventricular function, patients and medical teams might find it easier to track recovery. A study published on October 2 in *npj Cardiovascular Health* provides new evidence in this direction, but has not yet established a role in routine clinical care.

A team at Leiden University Medical Center in the Netherlands used an explainable deep learning model to analyze ECGs from a population following myocardial infarction and compared the predictive performance of single-lead and 12-lead ECGs. The journal abstract reports an area under the curve (AUC) of 0.883 for the single-lead model and 0.897 for the 12-lead model. AUC measures a model’s ability to distinguish between different states; it cannot be directly interpreted as “88.3% of patients were correctly diagnosed.” Similar values also do not mean that the two tests are clinically interchangeable.

This work did not simply use hospital ECGs to simulate wearable devices. According to a preprint of the same study released in February, the single-lead data came from 30-second ECGs recorded by patients using smartwatches in a home remote monitoring program; the raw data covered 708 people and approximately 64,000 recordings. However, not all of the raw recordings could be used to validate ventricular function, and the test set listed in the preprint came from 144 patients. These methodological details belong to an early version and should still be distinguished from the final journal version.

Home measurements also introduce problems encountered less often in hospital testing. The preprint notes that wearing the watch on the wrong wrist could invert the signal, while shared devices and recordings made while moving could introduce noise; the researchers therefore screened recordings for quality. This highlights a practical limitation of wearable assessments: model performance depends on the reliability of the input signals, and whether unsuitable recordings can be identified automatically in actual use also requires validation.

Another focus of the study was to understand the basis for the model’s judgments. The journal abstract states that explainability analyses showed that the model used physiologically plausible ECG features. This adds credibility to the results, but the relationship between electrical signals and pumping function still cannot be regarded as a direct measurement of the heart’s ability to contract, nor is it sufficient to determine treatment solely from the model’s output.

Conference materials provide additional context for the presentation of this work: a medical imaging symposium held in Leiden on January 27, 2026, had already listed research with the same title as poster No. 14, presented by Viktor van der Valk; the study was subsequently released as a preprint in February and appeared in a peer-reviewed journal in October. The symposium booklet can corroborate that the research was shared, but does not constitute another independent validation.

The key next step is to determine whether this ability to distinguish between states can be maintained across different hospitals, devices, and patient populations, and how alert thresholds can balance missed cases and false alarms. To become a clinical tool, it would also require prospective studies that closely reflect routine care to establish how results lead to further testing, whether they improve care, and to clarify the corresponding medical device regulatory requirements. The current findings support continued exploration of home assessments of cardiac function, but are not yet sufficient to justify omitting existing tests.