Left ventricular hypertrophy (LVH), left atrial enlargement (LAE), and left ventricular enlargement (LVE) are key echocardiographic markers of adverse cardiac remodelling. As universal echocardiographic screening is impractical, AI-enabled electrocardiogram (AI-ECG) models have been developed to identify patients who may benefit from targeted imaging. However, whether such models can simultaneously detect LVH, LAE, and LVE across diverse clinical settings — including lower-prevalence community populations — remains unclear.
To develop and validate an echocardiography-supervised AI-ECG model for simultaneously detecting LVH, LAE, and LVE across diverse clinical settings, and compare it with an alternative AI model trained on conventional ECG criteria.
A multicentre retrospective study used paired 12-lead ECGs and echocardiograms obtained within 30 days. A 1D squeeze-and-excitation residual network was developed using a tertiary-hospital-based public cohort of 205,916 ECG–echocardiography pairs from 44,583 patients, split 9:1 patient-disjoint into training and internal validation sets. Three external validation cohorts represented secondary care, primary care, and health screening settings. Echocardiographic labels were derived from quantitative guideline-based measurements. The comparator model used the same architecture but was trained on conventional rule-based ECG criteria. Discrimination was assessed by AUROC and compared using DeLong's test.
Across all cohorts, prevalence ranged from 5.5%–30.4% (LVH), 2.4%–31.5% (LAE), and 0.5%–15.8% (LVE). AUROC across the four cohorts:
| Target | Internal | Secondary Hospital | Primary Care | Health Screening |
|---|---|---|---|---|
| LVH | 0.783 (0.769–0.798) | 0.786 (0.778–0.794) | 0.892 (0.870–0.914) | 0.854 (0.827–0.882) |
| LAE | 0.782 (0.768–0.797) | 0.826 (0.817–0.834) | 0.912 (0.874–0.949) | 0.840 (0.794–0.886) |
| LVE | 0.832 (0.810–0.852) | 0.782 (0.769–0.796) | 0.944 (0.852–1.000) | 0.774 (0.725–0.824) |
The echocardiography-supervised model outperformed the rule-based comparator for LVH and LAE across all cohorts (all p<0.001 by DeLong's test). In pooled validation, discrimination was maintained across age, sex, comorbidity, and ECG acquisition settings.
An echocardiography-supervised AI-ECG model enabled simultaneous detection of LVH, LAE, and LVE across hospital, primary care, and health screening populations. Its superiority over a rule-based comparator suggests that echocardiographic supervision adds diagnostic value beyond conventional ECG criteria and may support scalable prioritisation for echocardiography across variable-prevalence settings.