AI-augmented ECG for pre-echocardiography triage: a tool to optimize cardiac imaging utilization

Echocardiography is a key diagnostic modality for cardiac dysfunction but is often over utilized due to variability in pre-test clinical assessment. There is a need for a scalable, cost-effective screening tool that can reduce unnecessary referrals without compromising diagnostic accuracy. To develop and validate an AI tool that uses standard 12-lead ECG images to predict the presence of major echocardiographic abnormalities, including reduced ejection fraction (EF ≤35%), valvular heart disease, and elevated pulmonary artery pressure, as a triage tool prior to echocardiography.

51,055 patients aged ≥15 years from a tertiary cardiac care centre, which underwent ECG and echocardiography on the same day between January 2021, and February 2024 were identified. ECGs were stored as images and pre-processed for model input. Echocardiographic findings were extracted using structured reports and regular expression-based keyword searches. The final dataset (n = 52,817) was split into training (40,796), epoch-monitoring (2,148), and testing (9,873) sets. An ensemble of 3 deep learning models was trained. Model performance was assessed using AUROC, PRAUC, sensitivity, specificity, positive predictive value, and negative predictive value. The internal test set demonstrated an AUROC of 0.87 (95% CI: 0.86–0.88) and PRAUC of 0.66 (95% CI: 0.65–0.69). At the Youden threshold (0.27), sensitivity, specificity, PPV, and NPV were 0.80, 0.80, 0.46, and 0.95, respectively. External validation was performed on 20,053 patients. It yielded an AUROC of 0.84 and PRAUC of 0.50.

The proposed AI model accurately identifies major echocardiographic abnormalities from ECG images, achieving high NPV and demonstrating strong generalizability.