Use of machine learning techniques to predict poor survival after hematopoietic cell transplantation for myelofibrosis.

Abstract

With the incorporation of effective therapies for myelofibrosis (MF), accurately predicting outcomes after allogeneic hematopoietic cell transplantation (allo-HCT) is crucial for determining the optimal timing for this procedure. Using data from 5,183 MF patients who underwent first allo-HCT between 2005 and 2020 at EBMT centers, we examined different machine learning (ML) models to predict overall survival (OS) after transplant. The cohort was divided into a training set (75%) and a test set (25%) for model validation. A Random Survival Forests (RSF) model was developed based on 10 variables: patient age, comorbidity index, performance status, blood blasts, hemoglobin, leukocytes, platelets, donor type, conditioning intensity, and graft-versus-host disease prophylaxis. Its performance was compared with a four-level Cox regression-based score and other ML-based models derived from the same dataset, and with the CIBMTR score. The RSF outperformed all comparators, achieving better concordance indices across both primary and post-essential thrombocythemia/polycythemia vera MF subgroups. The robustness and generalizability of the RSF model was confirmed by Akaike's Information Criterion and time-dependent Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) metrics in both sets. While all models were prognostic for non-relapse mortality, the RSF provided better curve separation, effectively identifying a high-risk group comprising 25% of patients. In conclusion, ML enhances risk stratification in MF patients undergoing allo-HCT, paving the way for personalized medicine. A web application (https://gemfin.click/ebmt) based on the RSF model offers a practical tool to identify patients at high risk for poor transplantation outcomes, supporting informed treatment decisions and advancing individualized care.