# Risk prediction for gastrointestinal bleeding in pediatric Henoch-Schönlein purpura using an interpretable transformer model

## TL;DR

This study developed a predictive model using a Transformer-based algorithm to assess gastrointestinal bleeding risk in children with IgA vasculitis, improving clinical decision-making.

## Contribution

An interpretable Transformer-based model for predicting GI bleeding in pediatric IgAV patients using clinically accessible biomarkers.

## Key findings

- The Transformer-based TabPFN-V2 model achieved high accuracy (0.88) and AUC-ROC (0.98) in predicting GI bleeding risk.
- SHAP analysis identified D-dimer, total cholesterol, platelet count, apolipoprotein, and C-reactive protein as key biomarkers.
- The model provides practical guidance for clinical decision-making in pediatric IgAV care.

## Abstract

Henoch-Schönlein purpura (HSP), clinically recognized as IgA vasculitis (IgAV), a prevalent systemic vasculitis in pediatric populations, frequently involves gastrointestinal (GI) tract manifestations that may lead to serious complications including hemorrhage and tissue necrosis. Timely identification of GI bleeding risk enables prompt clinical intervention and improves therapeutic outcomes. This study aims to develop and clinically validate an interpretable Transformer-based predictive model for assessing GI bleeding risk in pediatric patients with IgAV.

This retrospective cohort study analyzed 758 pediatric IgAV cases (ages 0–14 years) admitted to the Department of Pediatrics at the Affiliated Hospital of North Sichuan Medical College between 1 May 2020, and 31 January 2024. Comprehensive clinical data including symptoms and laboratory parameters were systematically collected. GI complications were stratified into three severity tiers: 1) no complications, 2) abdominal pain without bleeding), and 3) documented rectal bleeding or hemorrhage, based on standardized diagnostic criteria. Five machine learning algorithms (Random Forest, XGBoost, LightGBM, CatBoost, and TabPFN-V2) were optimized through nested cross-validation. Model performance was evaluated using multiple metrics: accuracy, precision, recall, F1-score, the Kappa coefficient, and ROC-AUC. The optimal model was subsequently interpreted using Shapley Additive Explanations (SHAP) values to elucidate feature importance.

Among the evaluated models, the Transformer-based TabPFN-V2 demonstrated superior predictive performance, achieving a validation accuracy of 0.88, precision of 0.88, recall of 0.87, F1-score of 0.88, Kappa coefficient of 0.82, and AUC-ROC of 0.98. SHAP analysis revealed the five most influential biomarkers for global interpretability: D-dimer, total cholesterol, platelet count, apolipoprotein, and C-reactive protein.

The interpretable Transformer-based TabPFN-V2 model demonstrated robust predictive performance for GI bleeding risk in pediatric IgAV patients. Clinically accessible laboratory parameters identified by this model not only offer practical guidance for clinical decision-making but also establish a foundation for advancing medical artificial intelligence integration in pediatric care.