Development and validation of an interpretable machine learning model for early prediction of refractory mycoplasma pneumoniae pneumonia: a multi-center study

Early prediction of refractory Mycoplasma pneumoniae pneumonia (RMPP) in children remains challenging with conventional approaches. Machine learning (ML) offers a promising approach to improve risk stratification by identifying patterns in multidimensional clinical data. This study aims to develop and externally validate an interpretable ML model for early RMPP identification. This multicenter retrospective study collected clinical data from pediatric patients with MPP across four medical centers between 2019 and 2023. Using data from two centers as the derivation cohort, we developed and compared six ML models to predict refractory outcomes. Model performance was assessed by discrimination, calibration, and decision curve analysis. The optimal model was externally validated in two independent cohorts and systematically interpreted using SHapley Additive exPlanations (SHAP) to reveal individual-level and cohort-level risk drivers. The analysis included 6,587 children across four centers: 5,546 in the derivation cohort and 696 and 345 in two external validation cohorts. The random forest model showed the best overall performance, with an AUC of 0.809 in the untouched internal hold-out test set. In the two external cohorts, the model achieved AUROCs of 0.774 and 0.807, respectively. Calibration drift was observed in external populations with different baseline RMPP prevalence, and local intercept recalibration improved calibration. SHAP analysis identified fever duration before hospitalization, neutrophil percentage, and age as the leading predictors. This externally validated machine learning model demonstrated robust generalizability across multiple centers and provided individualized risk assessment for children with MPP at admission. Local recalibration may be needed when applied to populations with markedly different baseline risks.