Machine-learning prediction of cardiovascular diagnoses from pediatric ECG reports in 11 643 children

Cardiovascular disease (CVD) in children is heterogeneous, often paucisymptomatic, and easily missed at first contact, yet the electrocardiogram (ECG) remains the most widely available cardiac test. In routine care the ECG is stored and exchanged as a structured report rather than as a raw waveform. Using the ZZU-pECG database of 14,190 twelve- and nine-lead recordings from 11,643 hospitalized children aged 0 to 14 years with International Classification of Diseases, 10th revision (ICD-10) diagnoses, we developed and internally validated seven machine-learning classifiers (elastic-net logistic regression, random forest, histogram gradient boosting, LightGBM, support vector machine, k-nearest neighbours, and a multilayer perceptron) to predict an ICD-10-coded diagnosis of any of 19 pediatric CVDs from 85 structured report features, comprising patient age and sex, the number and identity of expert ECG diagnostic statements, and per-lead signal-quality indices. Recording duration and lead configuration were deliberately excluded as potential workflow shortcut variables. Models were tuned by three-fold patient-grouped cross-validation and evaluated on a patient-disjoint hold-out test set of 2,814 recordings. The primary model (histogram gradient boosting) achieved an area under the receiver-operating-characteristic curve of 0.821 (95% CI 0.805 to 0.837), was well calibrated (slope 1.02, Brier 0.141), and showed positive net benefit across the range of decision thresholds examined. At a high-sensitivity operating point it identified 89.7% of recordings with a coded CVD diagnosis, with a negative predictive value of 93.5%. Discrimination fell to 0.748 using demographics and signal quality alone and to 0.776 using ECG statements alone (both DeLong P < 0.001 versus the primary model), and was largely preserved after removing the most disease-proximal statements (0.813, P = 0.014). Performance was materially lower in infants (0.635) and nine-lead recordings (0.746). Structured pediatric ECG reports contained a reproducible report-level signal for predicting coded CVD diagnoses in internal validation, but performance heterogeneity in these subgroups and the absence of external validation preclude clinical deployment. Because the predictors are themselves expert interpretations of the tracing, these results describe the information content of the ECG report rather than of the ECG signal.