Introduction: Severe adenovirus pneumonia (SAP) in children is associated with high morbidity ,a propensity for multi-system complications, and a significant risk of long-term sequelae. Objectives: To facilitate early intervention, this study aimed to develop and validate an interpretable machine learning model for predicting SAP risk in pediatric patients. Methods: A total of 930 pediatric patients diagnosed with adenovirus pneumonia were included in this retrospective analysis, with 43 variables encompassing clinical and demographic characteristics as well as laboratory indicators. The dataset was randomly partitioned into a training set (70%, n=652) and an independent test set (30%, n=278). Feature selection utilized Least Absolute Shrinkage and Selection Operator (LASSO) regression and the Boruta algorithm, identifying 20 key variables, followed by the construction of predictive models using eight mainstream machine learning algorithms. Model performance was evaluated based on the area under the receiver operating characteristic curve (AUC), decision curve analysis, and calibration curves, and the interpretability of the optimal model was assessed using SHapley Additive exPlanations (SHAP). Results: A predictive model incorporating 20 variables was developed. Among the algorithms tested, the LightGBM model demonstrated the best performance, achieving an AUC of 0.918. Ten-fold cross-validation and sensitivity analysis confirmed its robustness and reliability . SHAP analysis identified the top five predictive features: extensive lobar consolidation, fever duration, lung collapse, decreased breath sounds, and peak fever. Conclusion: This effective machine learning-based model enhances understanding of pediatric SAP and supports early identification and timely intervention.