Prediction of acute heart failure in pediatric myocarditis using a calibrated machine learning approach: a 13-year retrospective cohort study

Acute heart failure (HF) is a serious endpoint in pediatric myocarditis, yet early risk stratification remains challenging. We aimed to develop a machine-learning model to support early identification of children at elevated risk of in-hospital HF. In this single-center retrospective study, 1,208 children with myocarditis admitted from January 2013 to December 2025 were enrolled and split 7:3 into development and validation sets. Predictors were selected by combining the least absolute shrinkage and selection operator regression with Random Forest importance. Logistic regression, Random Forest (RF), and eXtreme Gradient Boosting were calibrated by isotonic regression. Discrimination, calibration, decision curve analysis, net reclassification improvement (NRI), and integrated discrimination improvement were assessed. Prespecified sensitivity analyses covered two composite endpoints, 100 repeated random splits, and temporal validation (development 2013–2022, validation 2023–2025). SHapley Additive exPlanations provided interpretability, and the best model was deployed as a web calculator. In-hospital HF occurred in 119 children (9.85%); ten consensus predictors were retained. On validation, calibrated RF reached an area under the curve (AUC) of 0.843 (95% CI 0.754–0.921; Brier 0.064); at a 10% threshold, sensitivity was 77.8% and negative predictive value 97.1%. The gain over N-terminal pro-brain natriuretic peptide (NT-proBNP) alone was nonsignificant (ΔAUC 0.030, P = 0.320), whereas reclassification improved over cardiac troponin T alone (NRI 0.476, P = 0.006). NT-proBNP had not been ordered in 44.2% of admissions; without it the model still reached 0.897. The same features predicted a major adverse cardiac event (MACE) composite (0.900) and acute HF or MACE (0.865); AUC across 100 random splits averaged 0.847 ± 0.034. On temporal validation, AUC was 0.813 (95% CI 0.736–0.882) and the advantage over NT-proBNP became significant (ΔAUC 0.067, P = 0.038), but risk was under-predicted (5.6% vs. 9.9% observed; calibration slope 0.77). This model may support early HF risk stratification in this single-center cohort; external validation is required before clinical use. Discrimination was preserved on temporal validation, but calibration drifted between eras, so the tool is a research prototype and an early screening aid rather than a standalone diagnostic instrument. Not applicable.