Detection of right ventricular heart disease in patients with adult congenital heart disease using artificial intelligence assisted electrocardiogram interpretation
D Ahmetagic, L Bosch, H B Van Der Zwaan, M I F J Oerlemans, R Van EsAbstract
Background
Advancements in care have led to a growing population of adults with congenital heart disease (ACHD). These patients require frequent follow-up, especially when atrial or septal defects (ASD/VSD) impose chronic volume or pressure overload. Current echocardiographic (TTE) follow-up strategies are often inefficient due to the variable rates of disease progression among individuals. While artificial intelligence-enabled interpretation of electrocardiograms (AI-ECG) offers a scalable solution for detecting structural heart disease in general cardiac care, its performance in the ACHD population has not been established.
Purpose
This study sought to develop and validate an AI-ECG algorithm to identify right ventricular (RV) abnormalities and pulmonary hypertension (PHT) in ACHD patients with a history of ASD or VSD.
Methods
For model development, 12-lead ECGs from a general hospital population were retrospectively paired with TTEs performed within 90 days at a Dutch academic centre. A deep convolutional network, fine-tuned from the pre-trained ECGFounder foundational model, was developed to predict a composite endpoint of RV abnormalities, its individual components, and PHT. The composite was defined as the presence of moderate-to-severe RV systolic dysfunction, RV dilatation or tricuspid regurgitation, or pulmonary hypertension (defined as a tricuspid regurgitant jet velocity ≥3.4 m/s and/or a right ventricular systolic pressure ≥40 mmHg). The model’s performance was validated in a separate, independent test set consisting exclusively of ACHD patients with a documented history of ASD or VSD, using only the first ECG per patient.
Results
The model was trained on 67,821 ECGs from 34,224 patients and validated on 535 ACHD patients (median age 29 [IQR 20–33] years; 58% female). In the test set, diagnoses included isolated VSD (44%), isolated ASD (34%), AVSD (20%), and Tetralogy of Fallot (2%). The composite endpoint was present in 102 patients (19%). The algorithm achieved an area under the receiver operating characteristic curve (AUC) of 0.76 (95% CI: 0.70–0.81) for the composite endpoint. Using the G-mean optimized threshold, the model demonstrated a sensitivity of 0.77 (95% CI: 0.69–0.86), specificity of 0.63 (95% CI: 0.59–0.68), positive predictive value of 0.34 (95% CI: 0.28–0.40), and a negative predictive value of 0.92 (95% CI: 0.89–0.95). Individual AUCs ranged from 0.87 (95% CI: 0.77–0.94) for RV systolic dysfunction to 0.67 (95% CI: 0.59–0.76) for tricuspid regurgitation. A complete overview of model performance can be seen in Table 1.
Conclusion
Our convolutional neural network accurately predicts RV abnormalities in patients with ACHD, a high-risk population in which AI-ECG algorithms have not previously been validated. This approach may enhance diagnostic screening efficiency and guide clinical decision-making for the longitudinal follow-up of ACHD patients.For image description, please refer to the figure legend and surrounding text.