Pediatric atrial septal defect detection using a federated deep learning model from electrocardiograms

Introduction Atrial septal defect (ASD) is one of the most common congenital heart diseases.1 ASD is often asymptomatic and missed until the development of irreversible complications such as atrial arrhythmias, stroke and congestive heart failure.2 Early detection is crucial to prevent long-term complications.3 We previously developed an artificial intelligence (AI) model capable of detecting ASD in adults.4 However, it remains unclear whether this model can be generalized to pediatric patients to enable earlier diagnosis. Purpose To develop an AI model capable of detecting ASD in pediatric patients. Methods ECG data were independently collected from INST1, a tertiary hospital, and INST2, a tertiary pediatric hospital. The presence or absence of ASD was confirmed by transthoracic/transesophageal echocardiography. Only the first available ECG obtained before ASD closure per patient was used. Case and control ECGs were matched at a 1:5 ratio based on age and sex. Model training was performed using a federated learning approach.5 Area under the receiver operating characteristic curve (AUROC) values from each institution were centrally averaged, and the model with the highest mean AUROC was selected. Test AUROCs of the final model were reported with the 95% confidence interval (CI), estimated using 2,000 bootstrap samples. Sensitivity and specificity for ASD detection were calculated and compared between the AI model and conventional ECG criteria (any ECG abnormality, right atrial enlargement and RBBB). Model performance was further evaluated across age-stratified subgroups. Results A total of 32,834 ECGs from INST1 and 14,457 ECGs from INST2 were used for model training. The previously developed adult ASD detection model trained at INST1 (AUROC 0.88 in the adult test set) displayed reduced performance when applied to pediatric test set, with AUROCs of 0.84 at INST1 and 0.72 at INST2. The model trained with both adult and pediatric data achieved AUROCs of 0.91 (95% CI, 0.89–0.94) in the INST1 test set and 0.86 (95% CI, 0.84–0.89) in the INST2 test set (Fig. 1). At a cutoff point achieving specificity >0.80, the sensitivities were 0.86 at INST1 and 0.70 at INST2 (specificities 0.81 and 0.83, respectively). Operating points derived from conventional ECG criteria (any abnormality, right atrial enlargement, and RBBB) were all located below the ROC curve of the ASD AI model. Age-stratified subgroup analyses demonstrated consistently high performance across age groups. In the adult cohort of INST1, the AUROC was 0.92 for patients aged <20 years and 0.91 for those aged ≥60 years. In pediatric patients at INST1, AUROCs were 0.88, 0.92, 0.88, and 0.93 for age groups <5, 5–10, 11–15, and 16–20 years, respectively. Corresponding AUROCs in INST2 were similarly high at 0.87, 0.87, 0.86, and 0.86 (Fig. 2). Conclusion The federated AI model demonstrated robust performance for detecting ASD in pediatric patients across institutions.Fig.1Fig.2