ECG-based artificial intelligence for risk stratification and outcome prediction after shunt closure in congenital heart disease

Introduction The population of adults with congenital heart disease (ACHD) is growing rapidly. Despite successful closure of atrial (ASD) and ventricular septal defects (VSD), patients remain at lifelong risk for hemodynamic sequelae, particularly right ventricular (RV) dysfunction and pulmonary hypertension (PH). Current surveillance relies on routine transthoracic echocardiography (TTE), which is resource-intensive and often inefficient due to variation in disease progression. There is an unmet need for efficient triage to identify high-risk patients. While AI-enabled ECG (AI-ECG) is a validated screening tool in the general and hospital population, its utility for risk stratification and triage in post-shunt closure ACHD patients remains undefined. Purpose To develop and validate an AI-ECG model for the detection of RV abnormalities and long-term risk stratification in ACHD patients following shunt closure. Methods We developed a deep convolutional neural network, fine-tuned from the ECGFounder foundational model, using ECG/TTE pairs from an academic hospital population. The model was trained to predict a composite endpoint of RV abnormalities (moderate-to-severe RV systolic dysfunction, dilatation, tricuspid regurgitation, or PH). Validation was performed in an independent cohort of ACHD patients after surgical or percutaneous shunt closure. Discriminative performance was assessed using Area Under the Receiver Operating Characteristic curves (AUC). Predictive value for new-onset abnormalities was evaluated using Kaplan-Meier analysis and multivariable Cox proportional hazards models adjusting for baseline age, sex, and procedure-type. Results The model was trained on 71.070 ECG/TTE pairs and validated in 332 post-closure ACHD patients (median age 25 [IQR 20 - 26], 57% female), with 1164 ECG/TTE pairs in total. Patients were diagnosed with isolated ASD (171, 52%), VSD (94, 28%), and combined AVSD (67, 20%) repaired via surgical (275, 83%) or percutaneous (57, 17%) procedures. For the composite endpoint, the algorithm achieved an AUC of 0.78 (95% CI 0.75 – 0.81). At a 90% sensitivity threshold, the negative predictive value was 0.94 (95% CI 0.92 – 0.96), facilitating effective rule-out of complications. Individual AUCs ranged from 0.77 for tricuspid regurgitation to 0.83 for RV systolic dysfunction. In multivariable Cox analysis, AI-ECG was a significant predictor for all new-onset abnormalities with hazard ratios ranging from 1.37 to 1.48 per unit increase in logit-probability (all p ≤ 0.01). Model concordance indices for predicting future events ranged from 0.77 to 0.87. Conclusion Our AI-ECG model accurately identifies RV abnormalities and predicts new-onset complications in ACHD patients after shunt closure. By providing scalable, automated risk stratification, this tool may enhance follow-up efficiency and facilitate timely intervention in this high-burden population.