Construction of an Early Risk Stratification Model for Severe Mycoplasma pneumoniae Pneumonia in Children: A LASSO-Stability Selection and Multi-Model Comparative Study

Authors

• Liu Tianyang, Jiamusi University, Jiamusi 154000, Heilongjiang, China
• Huang Xinyuan, Clinic of Jiamusi Children's Welfare Institute, Jiamusi 154000, Heilongjiang, China
• Wang Ying, Department of Pediatrics, The First Affiliated Hospital of Jiamusi University, Jiamusi 154000, Heilongjiang, China
• Zou Yanhong, Department of Pediatrics, The First Affiliated Hospital of Jiamusi University, Jiamusi 154000, Heilongjiang, ChinaFollow

Abstract

Objective: To develop an early risk stratification model for severe Mycoplasma pneumoniae pneumonia (SMPP) in children and compare Logistic regression, Random Forest (RF), and XGBoost. Methods: This retrospective study included 1,000 children with MPP at a single center (May 2023–May 2025). The primary outcome was severe clinical progression requiring intensive intervention (oxygen therapy, PICU admission, corticosteroid pulse therapy, or septic shock), defined independently of candidate predictors. Missing data were handled using multiple imputation (20 datasets); variables with >30% missingness were excluded. LASSO stability selection (≥19/20 datasets) identified predictors. Three models were developed in a 70% training set and validated in a 30% test set. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, Brier score, and decision curve analysis (DCA). Temporal validation and bootstrap optimism correction were performed. A nomogram and an online risk calculator were developed based on the best-performing model. Results: LASSO stability selection identified 8 core predictors: age, fever duration, C-reactive protein (CRP), lactate dehydrogenase (LDH), pleural effusion, creatine kinase (CK), adenosine deaminase (ADA), and interleukin-6 (IL-6). Logistic regression achieved the highest test AUC (0.795; 95% CI: 0.734–0.856), with a statistically significant difference compared with RF (P=0.019) but not with XGBoost (P=0.050). Logistic regression also demonstrated the best calibration (Brier score=0.140) and the highest clinical net benefit across clinically relevant threshold probabilities. RF and XGBoost overfitted (training AUCs 1.000 and 0.844). Temporal validation yielded an AUC of 0.780 (95% CI 0.725–0.835), and the optimism-corrected AUC was 0.726. Conclusion: The Logistic model based on 8 readily available predictors showed promising performance for early risk stratification of SMPP. Prospective multicenter external validation is warranted before clinical implementation.

Recommended Citation

Tianyang, Liu; Xinyuan, Huang; Ying, Wang; and Yanhong, Zou (2026) "Construction of an Early Risk Stratification Model for Severe Mycoplasma pneumoniae Pneumonia in Children: A LASSO-Stability Selection and Multi-Model Comparative Study," Journal of Pediatric Infectious Diseases: Vol. 21: Iss. 5, Article 5.
DOI: https://doi.org/10.53391/1305-7707.1085
Available at: https://jpid.researchcommons.org/journal/vol21/iss5/5