A stacking machine learning model for identifying comorbid anxiety disorder in children and adolescents with functional abdominal pain disorders

This is the authors' abstract. We don't add key points for this paper.

Objective Comorbid anxiety disorder is common in children and adolescents with functional abdominal pain disorders, but clinically usable risk-stratification tools integrating abdominal pain, psychosocial, school-related, and family-level information remain limited. Methods In this observational predictive modeling study, children and adolescents aged 7–18 years with Rome IV-defined functional abdominal pain disorders were evaluated between December 2021 and December 2025. Participants were randomly allocated at a 7:3 ratio into training and internal validation cohorts using stratified sampling by anxiety status. Candidate predictors included demographic, abdominal pain-related, somatic, psychosocial, school-related, and family-level variables. After LASSO-based feature selection, four prespecified base learners—explainable boosting machine, calibrated support vector machine, random forest, and CatBoost—were combined using an L2-penalized logistic regression stacking meta-learner. Performance was assessed using discrimination, threshold-dependent metrics, calibration, decision curve analysis, learner diversity, and SHAP-based interpretability. Results Among 629 eligible patients, 210 (33.4%) had comorbid anxiety disorder. The training cohort included 440 patients, and the validation cohort included 189 patients, with similar anxiety prevalence in both cohorts. Clinically relevant correlates of anxiety included older age, longer symptom duration, higher pain frequency, night-time awakening, pain temporally associated with stressful events, higher SDSC and CSI-24 scores, higher parental GAD-7 score, and greater school absence. In out-of-fold training estimates, the stacking model achieved an AUC of 0.831 and a Brier score of 0.151. In the validation cohort, it showed the highest numerical performance among candidate models, with an AUC of 0.846 (95% CI, 0.789–0.894), accuracy of 0.788, sensitivity of 0.762, specificity of 0.802, F1 score of 0.706, and Brier score of 0.146. Decision curve analysis suggested potential clinical utility across selected threshold ranges. SHAP analysis identified school absence, parental GAD-7 score, CSI-24 score, pain-stress association, symptom duration, pain frequency, SDSC score, age, and night-time awakening as leading contributors to the final model output. Conclusion A stacking machine learning model integrating clinical, psychosocial, school-related, and family-level variables showed acceptable internal-validation performance for identifying comorbid anxiety disorder in children and adolescents with functional abdominal pain disorders. External validation is needed before clinical implementation.