Machine learning models for predicting plastic bronchitis in hospitalized children with Mycoplasma pneumoniae pneumonia undergoing clinically indicated bronchoscopy: a single-center retrospective study

Objective To develop and compare multiple machine learning models for predicting the risk of plastic bronchitis (PB) in children with Mycoplasma pneumoniae pneumonia (MPP) and to establish an interpretable early risk identification tool. Methods A retrospective study was conducted involving 702 hospitalized children with MPP (477 non-PB and 225 PB) admitted to the Affiliated Hospital of Yan'an University from January 2023 to January 2026. Clinical characteristics, laboratory findings, and imaging data were collected. Key predictors were identified using a two-stage Lasso–Boruta approach. The dataset was randomly divided into training and test sets at a ratio of 7:3. Eight machine learning models, including CatBoost, XGBoost, LightGBM, Random Forest, Logistic Regression, Support Vector Machine, Gaussian Naive Bayes, and K-Nearest Neighbors, were developed. Hyperparameters were optimized using five-fold stratified cross-validation with grid search or random search based on the F 1 score. Model performance was evaluated using accuracy, precision, recall, F 1 score, ROC-AUC, PR-AUC, calibration, and decision curve analysis. SHAP analysis was applied for model interpretation. Temporal validation and bootstrap optimism correction were further performed to assess temporal generalizability and optimism in model performance. Results The Lasso–Boruta approach identified nine PB-related predictors, including fever duration, peak temperature, WBC, CRP, ALB, LDH, FDP, D-dimer, and pulmonary consolidation. All eight models demonstrated predictive capability in the test set, with ROC-AUC values ranging from 0.854 to 0.881. RF achieved an accuracy of 0.853, precision of 0.804, recall of 0.721, F 1 score of 0.759, and ROC-AUC of 0.881, whereas CatBoost achieved an accuracy of 0.853, precision of 0.793, recall of 0.736, F 1 score of 0.762, and ROC-AUC of 0.871. In temporal validation, RF achieved the highest accuracy (0.925) and F 1 score (0.900), whereas CatBoost achieved the highest ROC-AUC (0.970) and PR-AUC (0.956). Considering the complementary performance profiles across different evaluation dimensions and the results of bootstrap optimism correction, RF and CatBoost were retained as candidate models for further interpretation and risk prediction tool development. SHAP analysis provided individual-level explanations of the contributions of the selected predictors to PB risk. Conclusions Machine learning models based on clinical manifestations, laboratory indicators, and imaging features showed potential for PB risk identification in children with MPP. RF and CatBoost exhibited complementary predictive strengths across different evaluation dimensions, while SHAP analysis facilitated individual-level interpretation. Further multicenter external validation is required.