Related Experiment Video
Updated: Feb 17, 2026
Real-Time Cardiac Mapping with a Noninvasive Imageless Electrocardiographic Imaging System
Published on: April 11, 2025
Development and multicentre validation of an artificial intelligence electrocardiogram model for ventricular
Son Q Duong1,2,3, Akhil Vaid2,4, Pengfei Jiang5,6
1Department of Pediatrics (Cardiology), Icahn School of Medicine at Mount Sinai, 1468 Madison Ave, Annenberg 3rd Fl, New York, NY 10029, USA.
Aims:
Periodic cardiac MRI (CMR) is recommended to identify adverse ventricular remodelling in repaired tetralogy of Fallot (TOF), but access to CMR is uneven, and compliance is poor. We developed a 12-lead electrocardiogram (ECG) artificial intelligence (AI) biomarker to identify CMR-quantified adverse biventricular remodelling in repaired TOF.
Methods And Results:
Six (1 train/5 external test) North American retrospective cohorts with paired ECG and CMR were included. The main outcome was a composite of ≥2 TOF-specific CMR abnormalities: right ventricular (RV) end-diastolic volume ≥ 160 mL/m2, RV end-systolic volume ≥ 80 mL/m2, RV ejection fraction (EF) <47%, and left ventricular EF <55%. Model discrimination, calibration, and net benefit as a screening test to rule out ventricular remodelling were assessed. Nine hundred and eight patients (2552 ECGs) were included in training, and 782 patients (1795 ECGs) in external validation (outcome prevalence 57%). The area under the receiver-operating curve (AUROC) was 0.85 (95% confidence interval 0.83-0.87), and average precision was 0.88. At a screening risk-threshold of 0.25, there was 92% sensitivity, 41% specificity, 87% negative predictive value, and 55% positive predictive value for ventricular remodelling, which yielded a 13% net reduction in CMR use on net benefit analysis. There was no difference by sex or race/ethnicity, but there were differences by age and site, with two of five sites with lower AUROC than the others, and three of five sites met criteria for miscalibration, which improved after centre-specific calibration.
Conclusion:
An artificial intelligence analysis of electrocardiogram (AI-ECG) biomarker in repaired TOF effectively identifies ventricular remodelling to inform timing of advanced imaging. Extensive external validation revealed variation in discrimination and calibration that are important considerations for clinical implementation and regulatory approval pathways of AI-ECG in congenital heart disease.
