Development and validation of an interpretable machine learning model using routine laboratory biomarkers to stratify severe pneumonia risk in young children

• Interpretable ML framework: Stratifying pediatric severe pneumonia risk effectively. • Clinical utility: Achieving high performance for admission diagnosis and progression prediction. • Translational implication: Deploying clinical decision support with case-level interpretability. • Novelty biomarkers: Identifying chloride and glucose for high-risk pediatric pneumonia. • Validation scale: Leveraging large-scale pediatric cases to ensure robustness. Severe pneumonia is a leading infectious cause of mortality in children under 5 years globally. Early identification of high-risk cases remains challenging due to the lack of reliable stratification tools. This study aimed to develop an interpretable machine learning (IML) model using routine laboratory biomarkers for simultaneous diagnosis of severe pneumonia at admission and prediction of progression risk during hospitalization. This retrospective cohort study analyzed 85,886 children with pneumonia from a Chinese tertiary hospital (2013–2023). Two matched cohorts were established: Cohort I (n = 7,132) for admission diagnosis, and Cohort II (n = 1,064) for progression prediction. Fifty-seven laboratory parameters collected within 24 h of admission were extracted from electronic health records. Nine machine learning (ML) algorithms underwent systematic evaluation. Model performance was assessed using the area under the receiver-operating-characteristic curve (AUC), among other metrics. Model interpretability was achieved via SHapley Additive exPlanations (SHAP) analysis. We evaluated the performance of nine machine learning algorithms and the CatBoost model incorporating 11 laboratory features demonstrated superior performance (AUC: 0.879 for admission diagnosis; 0.839 for progression prediction). Key feature thresholds were optimized using Youden’s index (e.g., chloride ≤ 99 mmol/L). A real-time web application with case-level interpretability was deployed. This interpretable CatBoost model accurately stratifies pediatric severe pneumonia risk using routine laboratory data. Clinical implementation via the web tool may facilitate early intervention in resource-limited settings, though extensive external validation is warranted.