Validating ECG-Based Artificial Intelligence Models for Cardiomyopathy Detection in Obstetric Patients: A Multisite International Study

Abstract

Background
Cardiomyopathy significantly impacts maternal health worldwide. Artificial intelligence–enabled electrocardiograms (AI-ECG) have shown effectiveness in detecting left ventricular systolic dysfunction (LVSD) in adults and limited obstetric cohorts; however, broad validation in diverse obstetric populations remains unexplored.

Objectives
Our primary objective was to validate the performance of 2 versions of an AI-ECG model for detecting cardiomyopathy in an international obstetric cohort.

Methods
We conducted a retrospective, multisite study of pregnant or postpartum women (≤12 months) across 11 institutions in 4 countries (Israel, Nigeria, Pakistan, and the United States). Two AI-ECG models (Food and Drug Administration–cleared v2.5.0 and refined v3.0.0) were evaluated against echocardiography. LVSD was defined as left ventricular ejection fraction ≤40% in keeping with the model’s training target. Model performance was assessed by area under the receiver-operating characteristic curve (AUROC).

Results
We included 1,820 women with paired ECG-echocardiogram data, of whom 7.3% had LVSD. Model 1 (v2.5.0) demonstrated strong discriminative performance with an AUROC of 0.865, sensitivity of 46.1%, and negative predictive value of 95.8%. Model 2 (v3.0.0) showed significantly improved performance, with an AUROC of 0.925, sensitivity of 64.1%, and negative predictive value of 97.2% (P < 0.001) for LVSD detection.

Conclusions
AI-ECG models demonstrated strong external validity in a global obstetric cohort, with significantly improved performance in the refined model trained on expanded data. These findings support the potential role of AI-ECG for a scalable LVSD screening tool in obstetric populations, while underscoring the importance of population-specific validation prior to clinical implementation.