Expert-Level Automated Diagnosis of the Pediatric ECG Using a Deep Neural Network.

Summary

Using 583,134 pediatric ECGs from 201,620 patients, a CNN achieved expert-level performance, outperforming commercial software for any abnormality (AUROC 0.94), WPW (0.99), and prolonged QTc (0.96). Blinded re-adjudication favored AI classifications over original reads when discordant.

Key Findings

• CNN outperformed commercial software across tasks: any abnormality (AUROC 0.94; AUPRC 0.96), WPW (AUROC 0.99; AUPRC 0.88), prolonged QTc (AUROC 0.96; AUPRC 0.63).
• Blinded expert readjudication agreed with AI more often than with the original cardiologist reads for discordant cases (P=0.001 for any abnormality).
• Massive dataset included diverse ages (median 11.7 years) and conditions (11% congenital heart disease), supporting generalizability.

Clinical Implications

AI-ECG can be deployed to screen and triage pediatric ECGs for abnormalities, WPW, and QT prolongation, potentially reducing diagnostic delays and variability. Integration with clinician oversight and prospective external validation will enable safe adoption.

Limitations

• Single-center, retrospective design without external prospective validation
• Potential label noise and spectrum bias inherent to clinical datasets; unclear performance in low-resource acquisition settings

Future Directions

Prospective, multi-center external validation; workflow integration studies (assisted reading, triage); regulatory evaluation and fairness assessment across demographics and device vendors.