Development and validation of a predictive model for myocardial injury in children with pneumonia based on interpretable machine learning.

OBJECTIVE: This study aimed to identify independent factors for pneumonia-related myocardial injury in children, and construct and validate a risk prediction model and stratification system to enable early screening of high-risk patients and guide individualized clinical interventions. METHODS: A total of 222 children hospitalized with pneumonia from January 2023 to May 2026 were enrolled in this retrospective study and equally divided into myocardial injury and non-injury groups (111 cases per group). Admission baseline clinical and laboratory data were collected. Core predictive variables were screened via univariate analysis and LASSO regression. Restricted cubic spline regression was performed to explore the nonlinear relationships between variables and myocardial injury. Multivariate logistic regression was used to identify independent risk and protective factors, while the SHAP algorithm validated variable importance and effect directions. We further analyzed variable interactions and established a nomogram model. Model performance was evaluated using ROC curves, bootstrap calibration, and DCA. Subgroup analysis and a three-tier risk stratification system were conducted for systematic verification. RESULTS: Univariate analysis screened 11 differential variables, which were compressed to seven core predictors by LASSO regression. Multivariate analysis identified AST (OR = 1.11) and TP (OR = 1.12) as independent risk factors, while serum sodium (OR = 0.81), carbon dioxide combining power (OR = 0.86), and urea (OR = 0.67) were protective factors (all P < 0.05), consistent with SHAP validation results. Significant synergistic interaction was found between AST and TP (S = 1.32), whereas AST exerted antagonistic interactions with sodium (S = 0.74) and carbon dioxide combining power (S = 0.78). The nomogram yielded AUCs of 0.902 (training set) and 0.822 (validation set), with calibration errors of 0.021 and 0.035. DCA demonstrated favorable clinical net benefit, and all subgroup AUCs exceeded 0.85. The three-tier stratification system presented significantly different myocardial injury rates (6.12%, 25.36%, and 62.79% for low-, medium-, and high-risk groups, P < 0.001). CONCLUSION: AST, TP, serum sodium, carbon dioxide combining power, and urea independently affect the risk of pediatric pneumonia-associated myocardial injury with distinct synergistic and antagonistic interactions. The validated nomogram and risk stratification system serve as accurate and stable bedside tools for early risk assessment, facilitating timely clinical intervention and optimized medical resource allocation.