EQUITY IN HCT SURVIVAL PREDICTIONS

Hematopoietic Cell Transplantation (HCT) remains a critical therapeutic option for various hematological conditions, yet predicting post-transplant survival outcomes remains a complex challenge—particularly across diverse racial and socioeconomic groups. This project proposes a machine learning–driven system that enhances survival prediction for allogeneic HCT patients by addressing existing disparities related to race, geography, and socioeconomic status. Leveraging ensemble models such as XGBoost, CatBoost, and LightGBM, the system integrates clinical and demographic data to estimate personalized survival risk scores. A user-friendly interface is developed using Next.js (frontend) and Flask (backend) to enable clinicians and researchers to interact with the model and obtain interpretable predictions. For fair and accurate model evaluation, we employ the Stratified Concordance Index (C-index), a metric specifically adapted to assess predictive performance across racial groups independently. This metric not only evaluates the model’s ability to rank survival times reliably but also penalizes variability across racial subgroups, promoting equitable healthcare outcomes. By incorporating both algorithmic sophistication and a focus on fairness, this project aims to assist clinical decision-making, improve patient stratification, and foster greater trust in predictive healthcare systems.