In hematology, treatment decisions, particularly those related to transplantation, necessitate nuanced and multifaceted risk assessments. Now, a web application (gemfin.click/ebmt) based on the Random Survival Forests (RSF) model can help identify patients at high risk for poor transplantation outcomes, thereby supporting informed treatment decisions and advancing individualized care. Juan Carlos Hernández-Boluda, MD, PhD, a hematologist at Hospital Clínico Valencia in Spain, and colleagues published the description and validation of their machine learning model in Blood.
The prediction model included 10 variables: age, comorbidity index, performance status, blood blasts, hemoglobin, leukocytes, platelets, donor type, conditioning intensity, and graft-versus-host disease (GVHD) prophylaxis.
“The clinical implications of our study are centered around improving how we identify and manage risk in patients with myelofibrosis (MF) undergoing allogeneic hematopoietic cell transplantation (alloHCT),” said Adrián Mosquera Orgueira, MD, PhD, computational and genomic hematology group leader at the Health Research Institute of Santiago de Compostela in Spain and coauthor of the study. “Historically, we’ve relied on traditional baseline prognostic scores that, while useful, have notable limitations — they often overlook important variables like comorbidities or the evolving landscape of transplant strategies, including haploidentical donors and modern GVHD prophylaxis.”
The investigators performed retrospective modeling to develop an RSF using a heterogeneous cohort of 5,183 patients with MF who underwent their first alloHCT between 2005 and 2020 at European Society for Blood & Marrow Transplantation (EBMT) centers. The authors acknowledged that the study is limited by the fact that alloHCT is a multifaceted procedure influenced by numerous interrelated and independent variables that can significantly constrain the power of prognostic models. Nevertheless, they found that their model demonstrated superior performance in predicting both overall survival and non-relapse mortality when compared with established prognostic tools such as the Center for International Blood and Marrow Transplant Research (CIBMTR) score.
Dr. Mosquera Orgueira emphasized that the RSF was able to identify a high-risk subgroup comprising 25% of patients. “We didn’t anticipate such a large shift, and it really emphasized how much risk may be concealed within the so-called intermediate categories of older models,” Dr. Mosquera Orgueira said. “That has clear clinical implications because it suggests we might be under-recognizing vulnerability in a sizable portion of patients who could potentially benefit from alternative strategies or intensified follow-up.”
The researchers have made the model available as a web-based tool that enables clinicians to inform pretransplant discussions in real time. A routine pretransplant evaluation yields all the variables necessary for the tool, which integrates the information in a manner that reflects the complexity of individual patients. While highly effective in risk stratification, the investigators found that, even after adjustment, the web application struggled to predict the effects of modifiable transplant variables, such as donor type and conditioning regimen, or GVHD prophylaxis. This means, Dr. Mosquera Orgueira said, that it is not a tool for selecting specific transplant strategies but rather a tool to help clinicians identify which patients are likely to benefit from transplant and which may require alternative approaches or enrollment in a clinical trial.
Dr. Mosquera Orgueira said the model represents a step toward more personalized transplant medicine in MF because it can provide clinicians with a new tool to weigh the risks and benefits of treatment in a complex and evolving therapeutic landscape; however, it cannot directly guide treatment choices.
“Hematologists should be aware that machine learning is becoming an indispensable tool in modern medicine — not as a replacement for clinical expertise, but as a way to deepen and refine it,” he said. Such models, he explained, excel at integrating large amounts of clinical, molecular, and procedural data and identifying complex, often nonlinear relationships that traditional statistics can miss. The model also represents, Dr. Mosquera Orgueira said, a step into a new era shaped by generative artificial intelligence (AI). Ultimately, he said, the realization of the promise of AI in hematology lies in the collaboration between clinicians, data scientists, and algorithms that are designed to support, not supplant, human care.
Any conflicts of interest declared by the authors can be found in the original article.
Reference
Hernández-Boluda JC, Mosquera Orgueira A, Gras L, et al. Use of machine learning techniques to predict poor survival after hematopoietic cell transplantation for myelofibrosis [published online ahead of print, 2025 March 27]. Blood. doi: 10.1182/blood.2024027287.
Perspective
The U.S. Food and Drug Administration approval of four JAK inhibitors (JAKis) for the treatment of MF has led to increased recognition of this disease and earlier referral for hematopoietic cell transplantation (HCT). Referral of newly diagnosed patients and JAKi-induced symptom control and decreased splenomegaly have resulted in improved post-HCT survival in the last decade, especially among JAKi responders.1,2
Indication for HCT is generally determined based on disease stage using validated staging systems, including the Dynamic International Prognostic Scoring System or the Molecular International Prognostic Scoring System.3-6 However, predicting outcomes after HCT for MF has been less well studied. Published in 2019, the Myelofibrosis Transplant Scoring System (MTSS) identified clinical and molecular predictors of post-HCT survival in 361 patients.7 Although attempts to validate the MTSS have proven somewhat unsuccessful, there is evidence that the MTSS can discriminate a subset of patients that do particularly poorly.8
The current study used machine learning techniques to process data from more than 5,000 patients with MF who had received HCT to determine which patient-, disease-, and HCT-related variables were associated with post-HCT survival. The model outperformed traditional statistical approaches at identifying patients who were at highest risk of post-HCT mortality. Machine learning confirmed many of the high-risk variables included in the MTSS (age, Karnofsky performance status, platelet count, leukocyte count, and donor type) but also identified blood blasts, comorbidity index, hemoglobin level, conditioning intensity, and GVHD prophylaxis as significant. Unlike the MTSS, the machine learning model did not include molecular predictive factors, which confer risk for post-HCT relapse, and, of note, only 25% of patients in the cohort had received pre-HCT JAKis.
Identifying variables that are likely to increase post-HCT mortality based on systematic evidence is tantamount to providing patients with MF with the best possible advice regarding this life-altering procedure. Early referral, incorporating pre-HCT JAKi therapy, and careful selection of GVHD and conditioning regimens are strategies that can effectively modify many of the high-risk variables identified. While designed as a prognostic strategy for outcomes following HCT for MF, machine learning will prove the most meaningful if we can use the data to prevent poor outcomes in future patients.
Rachel B. Salit, MD
Associate Professor, Clinical Research Division
Fred Hutchinson Cancer Research Center
References
- Maze D, Arcasoy MO, Henrie R, et al. Upfront allogeneic transplantation versus JAK inhibitor therapy for patients with myelofibrosis: a North American collaborative study. Bone Marrow Transplant. 2024;59(2):196-202.
- Salit RB. The role of JAK inhibitors in hematopoietic cell transplantation. Bone Marrow Transplant. 2022;57(6):857-865.
- Gangat N, Caramazza D, Vaidya R, et al. DIPSS plus: a refined Dynamic International Prognostic Scoring System for primary myelofibrosis that incorporates prognostic information from karyotype, platelet count, and transfusion status. J Clin Oncol. 2011;29(4):392-397.
- Tefferi A, Guglielmelli P, Lasho TL, et al. MIPSS70+ Version 2.0: Mutation and karyotype-enhanced International Prognostic Scoring System for primary myelofibrosis. J Clin Oncol. 2018;36(17):1769-1770.
- Kröger N, Giorgino T, Scott BL, et al. Impact of allogeneic stem cell transplantation on survival of patients less than 65 years of age with primary myelofibrosis. Blood. 2015;125(21):3347-50.
- Gowin K, Ballen K, Ahn KW, et al. Survival following allogeneic transplant in patients with myelofibrosis. Blood Adv. 2020;4(9):1965-1973.
- Gagelmann N, Ditschkowski M, Bogdanov R, et al. Comprehensive clinical-molecular transplant scoring system for myelofibrosis undergoing stem cell transplantation. Blood. 2019;133(20):2233-2242.
- Hernández-Boluda JC, Pereira A, Alvarez-Larran A, et al. Predicting survival after allogeneic hematopoietic cell transplantation in myelofibrosis: performance of the Myelofibrosis Transplant Scoring System (MTSS) and development of a new prognostic model. Biol Blood Marrow Transplant. 2020;26(12):2237-2244.
