This study evaluated a composite artificial intelligence–enabled electrocardiogram (AI-ECG) designed to identify prevalent and predict incident structural heart disease (SHD). The composite combines two independently developed models targeting left ventricular systolic dysfunction (LVSD) and left ventricular diastolic dysfunction (LVDD), and a positive composite was defined as a positive result on either component model.
The analysis reported results across multinational cohorts and assessed both cross-sectional detection and longitudinal prediction among individuals without baseline SHD or heart failure.
The composite AI-ECG aggregates outputs from separate models for LVSD and LVDD. The authors defined SHD to include one or more of the following echocardiographic phenotypes: reduced left ventricular ejection fraction, moderate or severe valvular disease, left ventricular hypertrophy, or pulmonary hypertension. The composite was considered positive when either the LVSD model or the LVDD model signaled positivity.
Detection performance was assessed in two detection cohorts: a Korean clinical cohort from Incheon Sejong Hospital and a US dataset from Columbia University Irving Medical Center. The detection analyses included adults with paired electrocardiography and echocardiography studies. Incident risk was evaluated among individuals without baseline SHD or heart failure in the Korean cohort and in the UK Biobank. Exact eligibility criteria and additional baseline characteristics beyond cohort sizes were reported in the source but are not detailed further in this summary.
The composite AI-ECG detected SHD across both detection cohorts. In the two detection cohorts of 46,082 and 36,286 participants, the composite achieved the following performance measures:
The composite showed positivity across all included SHD phenotypes, indicating it captured signals associated with multiple structural abnormalities rather than a single target feature.
Among individuals without SHD or heart failure at baseline, composite positivity was associated with higher risk of developing SHD during follow-up. The association with incident SHD was reported as hazard ratios of 3.75 and 2.75 in the two at-risk cohorts evaluated. Model discrimination for incident SHD, reported by C statistic, ranged from 0.69 to 0.78 across cohorts.
These results indicate that composite positivity on the AI-ECG identified individuals at greater risk for subsequent development of SHD in cohort-level analyses.
The authors concluded that the composite AI-ECG identified prevalent and predicted incident structural heart disease across multinational cohorts and that it captured signals beyond the original training targets of the component models. They suggest potential utility as a scalable cardiovascular screening tool.
However, the report explicitly notes that prospective evaluation is required to determine whether ECG-based risk stratification using this composite approach improves clinical outcomes. As the work is presented as a preprint, it has not been certified by peer review and should not be used to guide clinical practice without further validation.
The study is a multinational cohort analysis with cross-sectional detection and longitudinal prediction components. The source reports Institutional Review Board approval at ISH; because the analysis used retrospective, previously collected and publicly available data, the requirement for informed consent was waived.
Details such as inclusion/exclusion criteria, follow-up duration, and other methodological specifics are presented in the source manuscript; if not fully described above, those details were not reported in the summary available here.
Multiple authors are employees of and hold stock in Medical AI Co., Ltd.; two authors hold equity in the same company. One author reported extensive research support and advisory relationships with multiple industry entities. Some authors declared no competing interests.
The preprint is made available under a CC-BY-NC-ND 4.0 International license. The report states that the findings are new and have not undergone peer review.