Key Points:
- Despite an increasing prevalence, patients with valvular heart disease (VHD) often remain undiagnosed or treated too late to meaningfully influence outcomes. AI-based ECG tools show promise in detecting structural heart disease.
- The PREVUE-VALVE study enrolled 3,000 community-dwelling adults aged 65–85, who underwent home-based echocardiograms and ECGs. The EchoNext AI-ECG model was then applied to this dataset to evaluate its performance for SHD detection.
- AI-ECG performance was moderate in the community setting compared to the original hospital-based derivation cohort likely reflecting differences in disease spectrum. Performance improved meaningfully in subgroups with abnormal ECGs or impaired health status, underscoring the importance of validating AI tools in their intended use population.
With an aging population, the incidence and prevalence of valvular heart disease (VHD) and broader structural heart disease (SHD) are rising. Despite the availability of effective therapies, many patients remain untreated or are identified too late to benefit. AI-based ECG tools have demonstrated strong performance for SHD detection in hospital and clinic-based populations, but their utility for screening in the community setting with less clinical phenotyping remains unknown.
The Age and Sex-Specific PREValence of AcqUirEd VALVular Heart DiseasE (PREVUE-VALVE) study was a decentralized clinical study in which community-dwelling adults aged 65–85 underwent home-based 12-lead ECGs and echocardiograms performed by trained technicians. The main results presented at TCT 2025 shed light on the prevalence of valvular heart disease (VHD) and its subtypes among older Americans, as well as the influence of age, sex, race and ethnicity on VHD prevalence. The final analytic cohort comprised 2,402 participants. The EchoNext deep learning model — previously developed and validated for SHD detection across multiple hospital-based cohorts — was applied to the PREVUE-VALVE dataset and compared to its performance in the original derivation cohort (New York Presbyterian Hospital). Prespecified subgroup analyses were performed to identify community-dwelling populations in whom AI-ECG performance was enhanced.
Among the 2,402 participants analyzed, the average age was 71 years, 58% were female, and 74% were non-Hispanic White. The overall prevalence of SHD was 7.5% — considerably lower than the approximately one-third prevalence observed in the hospital-based derivation cohort. The EchoNext model achieved an AUROC of 71.0 (sensitivity 70%, specificity 70%) in PREVUE-VALVE compared to 82.5 in the hospital-based cohort, with a corresponding PPV of only 14% (compared to 70% in the hospital derivation cohort) reflecting the lower disease prevalence in the community. This performance gap was largely due to differences in disease spectrum: PREVUE-VALVE participants had predominantly mild to moderate disease with high rates of tricuspid regurgitation, whereas the hospital cohort was enriched for more severe phenotypes, which generate more discriminatory ECG signals. Propensity score matching of the hospital cohort to the PREVUE-VALVE population substantially narrowed the performance gap, confirming that disease spectrum — rather than the model itself — was the primary driver. Importantly, subgroups with abnormal ECGs or impaired health status as measured by KCCQ overall summary score had meaningfully improved AI-ECG performance, with PPV rising to 20–26%, approaching the performance seen in hospital-based populations.
Presenting at the American College of Cardiology Scientific Sessions on March 30, 2026, lead investigator Dr. Timothy Poterucha concluded that “AI performance is critically dependent on population characteristics and disease spectrum,” and that “AI models developed in hospital-based populations may have markedly different characteristics in community settings.” He emphasized that these findings re-emphasize the importance of validating AI tools in their planned use cases before broad clinical deployment.