An AI model of transplantation risk for myelofibrosis Free

Allogeneic hematopoietic cell transplantation (allo-HCT) remains the only curative treatment for myelofibrosis (MF) but is associated with significant toxicity. It is therefore crucial to identify which high-risk patients benefit the most from allo-HCT and which would do better with alternative therapies. In this issue of Blood, Hernández-Boluda et al¹ develop an improved machine-learning-based model and web-based application to predict high risk of allo-HCT for the treatment of MF.¹ 

MF is a reactive bone marrow (BM) scarring condition due to chronic inflammation, usually found in myeloproliferative neoplasms (MPNs). Patients with MPNs can present with MF at diagnosis (primary MF [pMF]) or may acquire secondary MF (sMF) following disease evolution or transformation from polycythemia vera (PV) and, less frequently, essential thrombocythemia (ET).² Different allele burden, loss of heterozygosity, acquisition of disease-modifying mutations, and interactions with specialized BM niches may all contribute to the disparate risk of sMF in different MPNs.^2,3 

Although the management of MF in MPN has improved over the years, the main benefits of current treatments are ameliorations in constitutional symptoms and overall quality of life; however, the allele burden or the risk of acquiring secondary acute myeloid leukemia has not significantly changed.² The only curative treatment for MF remains allo-HCT, which is only recommended for a minority of patients due to its toxicity.⁴ Given the development of alternative symptomatic treatments and experimental agents currently being tested in clinical studies, it is imperative to improve allo-HCT outcome prediction to reserve that treatment for those most likely to benefit from allo-HCT vs those who would be more likely to benefit from alternative therapies.⁴ 

Existing prognostic models for overall survival (OS) of MF after allo-HSCT have focused on nonrelapse mortality (NRM) for informing clinical treatment decisions.^4,5 Currently, the identification of patients with MF who are eligible for HCT relies on several criteria. First is a high-molecular-risk profile, using systems such as the Mutation-Enhanced International Prognostic Score Systems MIPSS70⁶ (mutations in ASXL1, EZH2, SRSF2, IDH1/2) and, more recently, MIPSS70+ and MIPSS70+ v.2.0, which added U2AF1Q157 variant to the previously identified mutations.⁷ In sMF, the Myelofibrosis Secondary to PV and ET-Prognostic Model (MYSEC-PM) identified 4 different risk categories.⁸ Finally, the Myelofibrosis Transplant Scoring System (MTSS) was developed to more precisely predict outcomes of HCT in pMF or sMF.⁵ MTSS considers HCT-specific risk factors, including donor source and Karnofsky score or the Dynamic International Prognostic Scoring System (DIPSS/DIPSS+).⁹ 

Recently, a panel of MF experts representing the European Society for Blood and Marrow Transplantation (EBMT) and the European LeukemiaNet (ELN) revised the previous recommendations for HCT in MF. These experts advised that higher-risk patients, determined by specific risk scores (DIPSS and MIPSS70/MIPSS70+), should undergo transplant assessment if they exhibit a low or moderate risk of NRM according to the MTSS tool. Additionally, intermediate-risk patients can also be considered for HCT if they have a low HCT-related risk⁴ (see table, left).

Hernández-Boluda et al examined different machine learning models to better predict the risk of HCT in patients with MF. The authors analyzed data from 5,183 patients with MF who underwent first allo-HCT between 2005 and 2020 at EBMT centers. The investigators extended previous models by incorporating additional variables, such as patient comorbidities and new therapeutic strategies.

The machine learning approaches included a random survival forests (RSF), an oblique random survival forests, an XGBoost-based survival modeling (via the XGBSE library), and a deep learning method (DeepSurv). The cohort was divided into a training set (75%) and a test set (25%). The machine learning models were developed based on 10 key variables identified, consisting of patient age, comorbidity index (HCT-CI), performance status, peripheral blood blasts, hemoglobin, leukocytes, platelets, donor type, conditioning intensity, and graft-versus-host disease (GVHD) prophylaxis (see table, right). The investigators found that RSF predicts OS and NRM similarly or better than the alternative techniques. Although the same imputation methods were not consistently applied across all machine learning models (which might limit the analysis and the assumptions of the RSF), overall, the RSF model appears to outcompete the other models in identifying patients at high risk for poor outcome from HCT.

Compared with a 4-level Cox regression-based score and other machine learning–based models, the RSF model provided better separation and identification of the poor outcomes group. Although traditional performance metrics, such as the c-index, showed only modest differences overall, the performance of RSF was particularly enhanced compared with the one currently used in clinical practice based on the Center for International Blood and Marrow Transplant Research (CIBMTR) score.¹⁰ The CIBMTR model does not account for haploidentical transplants or posttransplant cyclophosphamide use, both of which are increasingly common in HCT centers. The inclusion of these factors underscores the clinical relevance of the RSF model.

The main difference between the CIBMTR and RFS models is the relative risk assigned to patient populations. Although the CIBMTR model classifies 40% of patients as low risk, 52% as intermediate risk, and only 8% as high risk for poor outcome, in the RSF model 25% of patients are classified as low risk, 50% as intermediate risk, and as many as 25% as high risk. Therefore, although the RSF does not separate intermediate-risk groups over other approaches, it identifies a much larger proportion of high-risk patients, comprising ≈35% NRM rate and ≈40% overall mortality at 1 year. The authors observed that drop-off in survival outcomes for the highest-risk group is primarily driven by well-established prognostic factors, including older age, worse performance status, and hematologic markers of advanced disease. These variables are strongly linked to post-HCT mortality, leading to enhanced stratification at the higher end of the risk spectrum. Although the RSF model does not try to provide a mechanistic explanation of disease progression, it effectively captures the combined influence of these clinical factors on HCT outcomes.

Finally, a web application (https://gemfin.click/ebmt) emerging from the RSF model offers a practical tool to identify patients at high risk of poor outcome after allo-HCT, to better guide treatment decisions. The web-based calculator allows the identification of a subset of patients with a high risk for poor outcome with a predicted 1-year OS of <50%, thus prompting consideration of other tailored therapeutic interventions for patients identified by this model.

Conflict-of-interest disclosure: S.M.-F. declares no competing financial interests.