Dr Hernández-Boluda on the Rationale for Using Machine Learning to Predict Post-HCT Outcomes in Myelofibrosis

“[There are] some limitations of the prognostic scores that we had so far [for determining patient transplant eligibility], and it will be very good to have a tool—a prognostic tool—in order to decide which patients should be taken to transplant, to conventional treatment, or even to clinical trials.”

Juan Carlos Hernández-Boluda, MD, PhD, an associated physician at the Hematology and Medical Oncology Service of the Hospital Clínico of Valencia, discussed the rationale for investigating the use of an EBMT machine learning model to predict survival outcomes following allogeneic hematopoietic cell transplant (allo-HCT) in patients with myelofibrosis.

Hernández-Boluda began by stating that he has over 15 years of experience using nontransplant therapies to treat patients with myelofibrosis. In this setting, he has used prognostic scoring systems to help guide treatment decisions, he stated. Recently, Hernández-Boluda and colleagues applied machine learning methodologies to develop more reliable prognostic tools for patients with myelofibrosis receiving nontransplant treatment, he explained. Based on this work, Hernández-Boluda identified limitations in existing transplant-related prognostic models as well as the potential value of a more robust prognostic tool to better inform decisions regarding patient eligibility for transplant vs other standard therapies vs enrollment in clinical trials.

Accordingly, Hernández-Boluda and colleagues developed an interactive, web-based machine-learning model to predict overall survival (OS) outcomes following allo-HCT among patients with myelofibrosis. A study evaluating the efficacy of different machine-learning models used data from 5183 patients with myelofibrosis who first underwent allo-HCT between 2005 and 2020 at EBMT centers. Investigators developed a Random Survival Forests (RSF) model that used patient baseline characteristics to predict OS outcomes. This RSF model outperformed all comparator machine-learning models used in the study, as it had better concordance indices across subgroups of patients with primary and secondary myelofibrosis. Furthermore, the RSF model identified 25% of patients as being at high risk for non-relapse mortality.