Predicting relapse in pediatric IgA vasculitis in the era of machine learning: an explainable hybrid ensemble model

This study aimed to develop an explainable hybrid ensemble machine learning (ML) model to predict relapse in pediatric IgA vasculitis (IgAV) using routine clinical and laboratory data.

This retrospective cohort study included 653 children diagnosed with IgAV according to EULAR/PRINTO/PRES criteria. Demographic, clinical, laboratory data, and the Pediatric Vasculitis Activity Score (PVAS) at diagnosis were evaluated. Several tree-based ML algorithms were evaluated and combined using a soft voting strategy. The hybrid ensemble model combining Random Forest, AdaBoost, and Extra Trees was selected as the final model. SHapley Additive exPlanations (SHAP)-based explainability analysis was performed, and a nomogram was developed for individualized risk prediction. Conventional logistic regression analyses were also performed to identify independent predictors of relapse.

Relapses were observed in 117 (17.9%) patients during follow-up. The hybrid ensemble model achieved the highest area under the receiver operating characteristic curve (AUC-ROC) (0.7504 ± 0.0856) with competitive area under the precision-recall curve (AUC-PR) (0.4540 ± 0.1308) and Brier score (0.1315 ± 0.0130). SHAP analysis identified older age at diagnosis, higher PVAS, female sex and abnormal urinalysis findings as the most influential predictors of relapse. On conventional regression analysis, PVAS was strongly associated with relapse. In addition, female sex, older age at diagnosis, rash involving the face, and abnormal urinalysis findings were independent predictors.

Disease relapse in pediatric IgAV can be predicted using clinical and laboratory data with an explainable hybrid ensemble ML approach. The predictive model and nomogram may support risk stratification and targeted follow-up in children with IgAV.