SHAP based feature selection for machine learning prediction of mycoplasma pneumoniae pneumonia with atelectasis in children.

backgroundMycoplasma pneumoniae pneumonia (MPP) with atelectasis in children is the most reported complication of MPP in children. It is associated with serious respiratory issues, pulmonary tissue damage and fibrosis, and systemic complications. Present diagnostic methods, which are mainly based on radiographic imaging, have drawbacks related to delayed case detection and radiation exposure. There is a need for a simple, non-invasive approach in the diagnosis. We aimed to develop machine learning (ML) models to identify MPP with atelectasis in children and support early risk stratification.

methodsWe conducted a retrospective study of 508 hospitalized pediatric patients with MPP from July 2022 to June 2024. Baseline clinical data were extracted from the electronic medical record system. Six tree-based ML models—decision tree (DT), random forest (RF), adaptive boosting (ADB), extreme gradient boosting (XGB), bayesian decision tree (BDT), and neural decision tree (NDT)—were developed. The SHapley Additive exPlanations (SHAP) was used to assess model feature importance, and final predictor subsets were determined through SHAP-guided forward feature selection. Models were tuned by Bayesian optimization within a nested five-fold cross-validation framework, and their performance was evaluated in an independent stratified validation cohort using receiver operating characteristic curve (AUROC), accuracy and decision curve analysis (DCA).

resultsSHAP-guided feature selection generated a specific feature set for each model. These features reflect systemic inflammatory burden, febrile status, adaptive immune response, and target-organ injury. In the training cohort, the RF model demonstrated the best performance, achieving an AUROC of 0.91 (95% CI 0.88–0.93). The performance of RF remained stable in the validation cohort, with an AUROC of 0.89 (95% CI 0.83–0.95). The model exhibited strong classification performance across other key metrics, including sensitivity, specificity, and negative predictive value. The results suggest that the model is generalizable with minimal overfitting.

conclusionThe interpretable RF model helps clinicians identify early cases of MPP that is complicated with atelectasis in children. This stable and interpretable method offers a practical tool for early risk stratification and timely decision-making.