A multicenter retrospective study included 419 children with Mycoplasma pneumoniae pneumonia, divided into training cohort (n=248), testing (n=62) and external validation (n=109), where patients were classified into groups without and with refractory form (RMPP) according to clinical recommendations. Radiomic features were extracted from chest CT scans using PyRadiomics, with feature selection using the SelectKBest and least absolute shrinkage methods. Three models based on random forests were developed: clinical-imaging, radiomic and integrated. The integrated model achieved the highest predictive performance in the validation cohort with an AUC of 0.811 (95% CI: 0.704–0.917), versus the radiomic model (AUC 0.788, 95% CI: 0.788–0.85) and the clinical-imaging model (AUC 0.675, 95% CI: 0.603–0.833). McNemar tests confirmed significant differences between models (p=0.001 between radiomic and clinical-imaging, p=0.013 between radiomic and integrated, p<0.001 between clinical-imaging and integrated). Net reclassification improvement (NRI) was higher for the integrated model compared to the others (both p<0.001). Key predictors included D-dimer, fever type, systemic immune inflammation index, gray-level co-occurrence matrix, and wavelet kurtosis. The integrated model improves RMPP risk stratification.