New in JACC: AI-ECG performance drops

🚨 New in JACC: AI-ECG performance drops when moving from hospital cohorts to real-world community settings

Many AI diagnostic models, including those designed to detect structural heart disease (SHD) from ECGs, are developed and validated in hospital populations enriched for higher disease prevalence and more severe phenotypes.

But what happens when these models are deployed for community screening?

The PREVUE-VALVE study evaluated the EchoNext AI-ECG model in community-dwelling adults aged 65–85 undergoing in-home ECG and echocardiography. The findings highlight an important challenge in AI translation:

• SHD prevalence was only 8% in the community cohort (vs ~43% in hospital-based cohorts)
• Patients had milder disease phenotypes and a different clinical case mix
• Model performance declined: AUC 71% (95% CI 66–76%) vs 83% in the hospital derivation setting

Adjusting for prevalence and case mix narrowed the performance gap — but did not fully eliminate it.

Interestingly, performance improved in higher-risk community subgroups, such as individuals with abnormal ECGs (AUC 79%).

Key takeaway: AI model performance is not automatically transportable across settings. Disease spectrum, patient characteristics, and clinical context matter. Models must be evaluated in the populations where they are intended to be used.

This is an important lesson for everyone building, validating, regulating, or deploying AI in cardiology and population health.

#AIinCardiology #AIECG #StructuralHeartDisease #DigitalHealth #JACC #PREVUEVALVE

Full open-access paper: https://www.jacc.org/doi/10.1016/j.jacc.2026.06.013