Development of a Risk Prediction Model for Refractory Mycoplasma Pneumoniae Pneumonia in Children Using a machine learning predictive model based on the SHAP methodology

Objective To analyze the clinical characteristics of children Refractory Mycoplasma pneumoniae Pneumonia (RMPP) and construct a clinical prediction model of RMPP.Methods Children with mycoplasma pneumoniae pneumonia (MPP) hospitalized in Zhengzhou Children's Hospital of Henan Children's Hospital (Affiliated Children's Hospital of Zhengzhou University) from January 2023 to December 2024 were collected. According to the clinical diagnosis results, 300 cases of RMPP and 300 cases of MPP were randomly selected as research samples. Six ML algorithms were used to construct the RMPP prediction model. Visual tools such as area under receiver operating characteristic curves (AUC), accuracy, sensitivity, accuracy, and F1 values, as well as ROC curves, calibration curves, and clinical decision curves (DCA) were used to evaluate model performance. In addition, the shapley Additive Interpretation (SHAP) method was used to select important features and interpret the final model.Results In the patients with RMPP, 64.33% (193/300) combined with lung consolidation; The proportion of adenovirus mixed infection was the highest (7.33%, 22/300). Among the six ML algorithms, LightGBM showed the best diagnostic performance, with AUC of 0.913 (95%CI: 0.859 -- 0.959), accuracy of 0.843, sensitivity of 0.852, accuracy of 0.839, and F1 value of 0.845.Conclusions RMPP is easily associated with lung consolidation, adenovirus is the most common type of mixed infection of RMPP. The RMPP prediction model based on the LightGBM algorithm and SHAP interpretability framework demonstrates excellent predictive performance and can be utilized for early warning of RMPP.