Prediction of Poor Survival after Hematopoietic Cell Transplantation in Myelofibrosis Using Machine Learning Techniques

Introduction: Allogeneic hematopoietic cell transplantation (allo-HCT) can cure myelofibrosis (MF) but requires careful risk-benefit analysis due to its toxicity. This is crucial as effective agents for treating MF have been incorporated into clinical practice. Current prognostic models for risk stratification after allo-HCT do not account for new strategies like haploidentical donor transplants or cyclophosphamide use. There is also a need to better identify patients at high risk of transplant complications who may benefit more from a medical approach. We aim to determine if new statistical approaches can enhance prognostication in allo-HCT for MF.

Methods: This registry-based study, approved by the Chronic Malignancies Working Party (CMWP) of EBMT, included adult primary and secondary MF patients undergoing first allo-HCT between 2005-2020. Patients transplanted from cord blood or with leukemic transformation were excluded. A total of 5,183 patients met the inclusion criteria. Study goal was to develop a prognostic model for overall survival (OS) using machine learning (ML) techniques (a random survival forest [RSF] model) and compare its performance with a Cox regression-based model developed in the same dataset, and with the CIBMTR risk score (Tamari, Blood Adv 2023). Two independent groups created the models, using the same random distribution of patients into a training set (75% of the cohort) and a test set (25% of the cohort). The models’ accuracy in predicting OS and non-relapse mortality (NRM) was compared using Harrel’s c‐indexes in both sets.

Results: After a median follow-up of 58.7 months, the estimated OS rate at 1 and 5 years was 70% (95% Confidence Interval [CI]: 69-71) and 53% (95% CI: 51-54), respectively. The estimated NRM rate at 1 and 5 years was 23% (95% CI: 22-24) and 32% (95% CI: 31-33), respectively.

In multivariable Cox-regression analysis, five factors predicted OS: patient age (<50/50-59/>59 years), donor type (HLA mismatched/other), Karnofsky performance status (KPS, <90/other), Hematopoietic Cell Transplantation Comorbidity Index (HCT-CI, high/other) and genotype (JAK2-triple negative/other). A score based on hazard ratios defined four risk categories with 5-year OS rates of each one in the training and test set being 82% (95% CI: 74-90) and 65% (95% CI: 41-89) for low risk (6% of the cohort); 62% (95% CI: 58-66) and 65% (IC 95%: 57-73) for intermediate-1 (36% of the cohort); 52% (95% CI: 48-56) and 47% (95% CI: 40-54) for intermediate-2 (48% of the cohort); and 39% (95% CI: 30-47) and 30% (95% CI: 13-47) for high risk (10% of the cohort), respectively. A RSF model was created to predict OS using 52 initial variables. After dimensionality reduction, the model was reduced to a minimum set of 10 prognostic factors: patient age, HCT-CI, KPS, blood blasts %, Hb value, WBC and platelet counts at transplant, donor type, conditioning intensity, and graft-versus-host disease prophylaxis.

In mortality prediction, the ML model showed modest superior performance to the Cox-derived score in the training set (c-index: 0.603 vs. 0.594) and test set (c-index: 0.611 vs. 0.587). The ML model outperformed the CIBMTR model in the training set (N=1,925; c-index: 0.611 vs. 0.557) and test set (N=618; c-index: 0.626 vs. 0.581). In predicting NRM, the ML model had a similar c-index compared to that of the Cox-derived score in the training set (0.599 vs. 0.597), but outperformed it in the test set (0.612 vs. 0.595), indicating higher reproducibility. Notably, the ML model identified 25% of high-risk patients, compared to 10% by the Cox-derived score and 8% by the CIBMTR model. In the test set, the ML model's high-risk group had 1 year and 2-years OS rates of 61% and 48% respectively, similar to the Cox-derived score’s rates of 62% and 50%. The high-risk group by the ML model had 1-year and 2-years NRM rates of 36% and 43% in the test set, matching those observed for this risk category by the Cox-derived score (36% at 1 year and 43% at 2 years).

Conclusion: The ML approach effectively identified high-risk patients with poor outcomes after allo-HCT, more than doubling the detection rate compared to Cox regression-based models. The model demonstrated substantial predictive accuracy across the training and test sets, indicating enhanced generalizability. This Ml model enables personalized patient management, potentially leading to better treatment decisions and improved survival.