Inflammatory phenotyping by latent class analysis and machine learning-based prediction of postoperative complications in pediatric appendicitis: a retrospective cohort study

Background Acute appendicitis exhibits heterogeneous inflammatory responses that conventional single-marker assessments fail to capture. This study identified distinct inflammatory phenotypes using latent class analysis (LCA) and developed machine learning models to predict postoperative complications in pediatric appendicitis. Methods This retrospective cohort study included 402 pediatric patients who underwent laparoscopic appendectomy. LCA was performed using nine inflammatory and clinical indicators to identify inflammatory phenotypes, and the optimal number of classes was determined by entropy, information criteria, and clinical interpretability. Three machine learning models—logistic regression, random forest, and XGBoost—were developed using a 70/30 train-test split. Five-fold cross-validation was performed within the training set for hyperparameter tuning and internal stability assessment, while the held-out 30% test set was reserved for final performance reporting. Model performance was evaluated using AUC, sensitivity, specificity, and calibration analysis. Results Three inflammatory phenotypes were identified: Low (29.4%), Moderate (40.8%), and High (29.9%) inflammation, with significant differences in inflammatory markers, perforation rates, hospital stay, and complication rates across classes (all P < 0.001). The overall complication rate was 17.4%. Multivariate analysis identified perforation as the sole independent predictor (OR 3.03, 95% CI 1.03–8.89; P = 0.044). On the test set, XGBoost achieved the highest AUC (0.911), followed by random forest (0.905) and logistic regression (0.860). All models demonstrated high sensitivity (0.833–0.917), NPV exceeding 0.98, and good calibration. Conclusion LCA revealed clinically meaningful inflammatory phenotypes in pediatric appendicitis with distinct outcome profiles. Machine learning models, particularly XGBoost, demonstrated excellent predictive performance for postoperative complications, supporting their potential future application in preoperative risk stratification pending external validation and prospective evaluation.