AI-ECG for LVSD detection: a systematic review and first-in-kind multinational head-to-head comparison

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Background and aims

Several artificial intelligence-enhanced electrocardiogram (AI-ECG) models have shown promise in detecting left ventricular systolic dysfunction (LVSD), but their head-to-head agreement and performance have not been independently evaluated within the same cohort. This study aimed to systematically characterize available AI-ECG models for LVSD detection, and for the first time evaluate their head-to head performance.

Methods

We systematically reviewed published AI-ECG models for LVSD detection, extracting data on architecture, cohort characteristics, and performance. Authors of these studies were invited to share models for external validation in a deeply phenotyped registry of all patients undergoing cardiac magnetic resonance imaging (CMR) in the Amsterdam UMC with paired ECGs. Performance was assessed in all consecutive patients and in a subset reflecting a lower-complexity sample (15% prevalence of LVEF <40%).

Results

We identified 35 studies describing 51 AI-ECG models, all showing high performance (area under the receiver operating characteristic curve [AUROC]>80) or excellent (AUROC >90) performance. Four research groups (Korea, the United States, Taiwan, The Netherlands) shared models for external validation. AUROCs ranged from 0.83–0.93 in all consecutive patients (n = 1,306; mean age 59 ± 15 years; 450 [35%] female) and 0.87–0.96 in the lower complexity subset. Performance remained consistent across subgroups, with minor degradations in ECGs displaying a wide QRS-complex or atrial fibrillation.

Conclusions

In this first-in-kind independent validation and head-to-head comparison study, AI-ECG for LVSD detection exhibited robust performance despite disparate training populations. However, limited availability of models impedes independent validation and clinical adoption.

Article activity feed