Identifying patterns in GvHD patients

Abstract

Acute Graft-versus-Host Disease (aGvHD) is a common and serious complication following allogeneic hematopoietic cell transplantation (allo-HCT) in pediatric patients. Although corticosteroids are the standard treatment, nearly half of the affected children do not respond adequately. This thesis aims to identify diagnostic features that distinguish steroid non-responsive patients using high-dimensional clinical data collected post-transplant. A dataset of 607 pediatric alloHCT patients, including 266 with aGvHD, of whom 41 were non-responders, was analyzed. After preprocessing, feature selection was performed using mutual information, followed by modeling with Random Forest and XGBoost classifiers. Results showed that Random Forest achieved high accuracy in identifying treatment responsiveness (AUC = 0.85), outperforming XGBoost. Key predictive features included reticulocyte fractions, albumin, and bilirubin levels. These findings suggest that machine learning models can effectively support early risk stratification and personalized treatment strategies in pediatric aGvHD care. Limitations include class imbalance and the need for external validation. Future work should focus on prospective studies and time-to-event modeling to enable clinical applicability.