This study evaluated whether EchoNext, an AI-ECG model trained in hospital-based cohorts, could reliably detect structural heart disease (SHD) when deployed in a community-dwelling population (PREVUE-VALVE; n=2,402 adults aged 65–85) undergoing in-home ECG and echocardiography.
Model discrimination was meaningfully lower in the community cohort (AUC 71%, 95% CI 66–76%) versus hospital-based cohorts (AUC 83%, 95% CI 82–83%), driven by lower SHD prevalence (8% vs 43%) and milder disease phenotypes; performance improved modestly in higher-risk subgroups (abnormal ECG: AUC 79%; impaired health status: AUC 76%).
- Single AI-ECG model (EchoNext) tested; findings may not generalize to other AI-ECG tools. - Propensity matching attenuated but did not fully explain the performance gap, suggesting unmeasured confounders remain. - Community cohort was limited to adults aged 65–85, restricting generalizability to younger populations.
AI-ECG models validated in hospital populations should not be assumed to perform equivalently in community or primary care settings, where disease is less prevalent and milder — clinicians should seek community-level validation data before adopting such tools for screening. Prioritizing AI-ECG use in higher-risk subgroups (e.g., those with an abnormal ECG) may preserve meaningful diagnostic yield.
Explore related topics
