Objective To develop and validate a radiomics-based model for the early prediction of delayed radiological recovery from Mycoplasma pneumoniae pneumonia (MPP). Methods The study cohort included MPP patients admitted to three medical centers from July 2023 to October 2023. Patients were divided into two groups based on clinical and imaging outcomes at two weeks post-treatment: recovery and delayed recovery. Image segmentation, feature extraction, and selection were performed using the uAI Research Portal V1.1. A logistic regression model was constructed. A total of 1904 radiomics features were automatically extracted from the raw images using the Pyradiomics V3.0 tool built into the uAI Research Portal V1.1. Feature selection was performed with Z-score normalization, Spearman correlation analysis, and least absolute shrinkage and selection operator methods. Clinical-imaging, radiomics, and integrated (clinical + imaging + radiomics features) models were constructed. Model performance was assessed by calculating the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Decision curve analysis was used to evaluate the net benefit. Results A total of 576 eligible patients were included in the study. Multivariate logistic regression analysis revealed that D-dimer levels (odds ratio [OR]: 4.06, 95% confidence interval [CI]: 2.52-6.55; p < 0.001), the systemic immune-inflammation index (SII) (OR: 6.61, 95% CI: 2.08-21.00; p = 0.001), and the consolidation pattern (OR: 2.82, 95% CI = 1.52-5.24; p = 0.001) were significantly associated with delayed recovery. With the training set, the integrated model outperformed the radiomics model in terms of predictive efficacy (AUC: 0.982 vs. 0.872, p = 0.01). However, there was no statistically significant difference between the integrated and clinical-imaging models (AUC: 0.982 vs. 0.894, p = 0.07). There was no significant difference in performance between the radiomics and clinical-imaging models with both the training and external validation sets (AUC: 0.894 vs. 0.872, p = 0.75; AUC: 0.807 vs. 0.837, p = 0.638). With the external validation set, the integrated model achieved an AUC of 0.865 (95% CI: 0.770-0.960), sensitivity of 0.933, and specificity of 0.720. Conclusion The MPP delayed recovery prediction model, based on 16 radiomics features, D-dimer, SII, and a consolidation pattern, showed excellent predictive performance. It can be used by clinicians as an effective tool for early identification and intervention.