Risk classification of severe pneumonia at admission in children with human metapneumovirus infection: development of a machine learning model based on retrospective data from 854 cases

To develop and validate a machine learning–based model to accurately classify the risk of severe pneumonia in children infected with hMPV. This retrospective cohort study consecutively enrolled clinical data from 854 hospitalized children with hMPV-associated pneumonia at Kunming Children’s Hospital between December 2019 and March 2025. According to guideline-based criteria, patients were classified into a severe pneumonia group ( n = 203) and a non-severe pneumonia group ( n = 651). The dataset was randomly divided into a training set and a testing set at a 7:3 ratio. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression. Six machine learning models, including logistic regression, k-nearest neighbors, naïve Bayes, multilayer perceptron, random forest, and extreme gradient boosting (XGBoost), were constructed and compared in the training set. Model performance in the independent testing set was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Clinical net benefit was assessed using decision curve analysis, and SHapley Additive exPlanations (SHAP) were applied to visually interpret the optimal model. LASSO regression identified 9 core predictors: age, wheezing, pulmonary rales, pleural effusion, ferritin, lactate dehydrogenase, neutrophil percentage, neutrophil count, and albumin. Among the six models, XGBoost and logistic regression demonstrated comparable overall performance. XGBoost achieved a training set AUC of 0.87 (95% CI: 0.84–0.90) and a test set AUC of 0.78 (95% CI: 0.72–0.85). The accuracy on the training set was 0.81, the sensitivity was 0.99, the specificity was 0.23, and the F1 score was 0.89; the accuracy on the test set was 0.79, the sensitivity was 0.98, the specificity was 0.17, and the F1 score was 0.88.Decision curve analysis indicated that the model provided a favorable clinical net benefit across a broad range of threshold probabilities. SHAP analysis revealed that wheezing, age, and pulmonary rales were the three most influential predictors. We successfully developed and validated machine learning models, particularly XGBoost and logistic regression, for risk classification of severe pneumonia in children with hMPV infection. The model demonstrated strong discriminative performance, clinical utility, and interpretability, and may serve as a valuable decision-support tool for risk stratification of high-risk pediatric patients. Not applicable.