Comprehensive Summary
The retrospective study aimed to analyze predictors of acute graft versus host disease (aGVHD) in pediatric patients using Random Forest (RF), Logistic Regression (LR), and Boruta machine learning methods. The “Bone Marrow Transplant: Children” dataset from the UCI Machine Learning Repository was used after Synthetic Minority Oversampling Technique (SMOTE) balanced the class distribution of the variables. Ultimately 124 pediatric patient data with ALL or AML that included demographic and clinical variables were analyzed. There were statistically significant differences noted for recipient age, recipient weight, donor age, allele mismatch, antigen mismatch and recipient CMV status with p-values under 0.05 identified by the machine learning methods. Allogenic hematopoietic cell transplantation (allo-HCT) is a crucial treatment for leukemia but it has a significant risk of aGVHD and significant morbidity and mortality.These machine learning models demonstrate potential for predicting post-transplant outcomes and may be further applied in the future to guide donor selection and treatment regimens.
Outcomes and Implications
The study concludes that machine learning models are capable of identifying risk factors for aGVHD in pediatric patients that undergo allo-HCT. These models can improve outcomes in allo-HCT by identifying patient risk factors, allowing clinicians to better prevent aGVHD in pediatric patients. With improvements to the model and training on datasets that integrate immunogenic profiling, comprehensive viral serostatus evaluation, and precise cell dose quantification alongside a multicenter design, future research can better create a model that can be applied in clinical settings.