Machine Learning Finds Heart Disease Clues Hidden in Routine Pediatric ECG Reports

Pediatric cardiovascular disease often lurks asymptomatically, evading detection until acute events force emergency intervention. Now, researchers from Iran and Michigan State University have demonstrated that machine learning can extract diagnostic signals from the structured text and metadata of routine ECG reports—without ever analyzing the underlying heart rhythm waveforms—to flag children at risk of hidden heart disease.

The study, published in BMC Pediatrics, analyzed 14,190 ECG recordings from 11,643 hospitalized children aged 0–14 years, achieving an area under the receiver operating characteristic curve (AUROC) of 0.821 and a 93.5% negative predictive value for coded cardiovascular diagnoses. The approach suggests a practical pathway for AI-assisted triage that integrates directly into existing hospital information systems.

Mining Diagnostic Intelligence from Structured Reports

Led by Pooya Eini of Tehran's Rajaie Cardiovascular Institute, the research team deliberately avoided raw ECG waveforms, instead feeding machine learning models 85 structured features extracted from routine clinical reports: patient demographics, expert ECG diagnostic statements, and per-lead signal-quality indices. The team excluded recording duration and lead configuration data to prevent models from learning workflow shortcuts rather than genuine diagnostic patterns.

Seven classifier algorithms were tested, with histogram gradient boosting emerging as the top performer. The model achieved a calibration slope of 1.02 and Brier score of 0.141, demonstrating both discrimination and reliability. At a high-sensitivity operating point, the system flagged 89.7% of children with cardiovascular disease while maintaining strong negative predictive value—suggesting utility as a rule-out screening tool to identify the minority of reports requiring urgent specialist review.

Feature importance analysis revealed that demographics and signal quality alone achieved AUROC 0.748, while ECG diagnostic statements alone reached 0.776. The full model significantly outperformed both individual feature sets (DeLong test P<0.001), indicating that integration of multiple data streams enhances diagnostic power.

Performance Gaps Highlight Deployment Challenges

Despite promising overall results, the model showed significant performance degradation in critical subgroups. AUROC dropped to 0.635 in infants and 0.746 for nine-lead recordings, compared to twelve-lead studies. These limitations constrain immediate applicability in neonatal intensive care and emergency settings where reduced-lead configurations are common.

The researchers employed rigorous validation methodology, using three-fold cross-validation with patient-grouped folds and testing on a patient-disjoint hold-out set of 2,814 recordings. This approach prevents data leakage from multiple ECGs per patient—a common pitfall in clinical AI research that can artificially inflate performance estimates.

The study targeted 19 pediatric cardiovascular diseases coded under ICD-10, drawing from the ZZU-pECG dataset of hospitalized children. No external funding supported the work, and authors declared no competing interests.

Implications for Clinical AI Integration

The research validates that structured clinical reports contain extractable diagnostic information suitable for machine learning, potentially offering a more practical integration pathway than waveform-based AI systems. Because structured ECG report data already flows through hospital information systems, such tools could deploy without requiring specialized data pipelines or waveform storage infrastructure.

However, the authors' methodological rigor in excluding workflow variables and preventing shortcut learning underscores a critical gap: external validation across multiple hospital systems remains essential before clinical deployment. The performance limitations in infants and with reduced-lead configurations further emphasize that promising research results do not automatically translate to deployment-ready clinical tools.

For pediatric cardiology, the potential payoff is substantial—systematic flagging of high-risk ECG reports could reduce missed diagnoses of serious but asymptomatic heart conditions, potentially catching diseases before they cause acute events in vulnerable children.

Sources

1. AI Reads Children's ECG Reports to Flag Hidden Heart Disease — Scienmag