New EBMT Machine Learning Model Outperforms Available Tools in Predicting Transplant Risk in Myelofibrosis

Juan Carlos Hernández-Boluda, MD, PhD

Although allogeneic hematopoietic cell transplantation (allo-HCT) remains an important part of the therapeutic paradigm of myelofibrosis as the only curative option for the disease, the emergence of new treatment options and risk assessment considerations often make deciding which patients are ideal candidates for transplant complex. In response to these barriers, a team of investigators from the European Society for Blood and Marrow Transplantation (EBMT) have developed a machine learning model designed to identify and stratify transplant risk for patients with myelofibrosis.1

On March 27, 2025, the EBMT announced the launch of the machine learning model. It can predict overall survival (OS) for patients with myelofibrosis following allo-HCT based on patient characteristics including age, performance status, and comorbidity index. The open-access tool is now available for use free of charge online.

“I’ve worked in transplantation for myelofibrosis for almost 2 decades, and, despite that experience, choosing the right time and patient to bring forward for transplant is still challenging,” Donal McLornan, MBBCh, MRCP, PhD, FRCPath, said in an interview with OncologyLive. “This is even more [true] in the current year because we have many clinical trials that patients are interested in and a whole host of new drugs that we don’t fully understand how to integrate into the transplantation algorithm. This tool will help us decide on the right patient to bring forward to transplant, alongside many other factors that we need to consider.”

During the 51st Annual Meeting of the EBMT, OncologyLive spoke with McLornan; Juan Carlos Hernández-Boluda, MD, PhD; and Adrián Mosquera Orgueira, MD, PhD, who were part of the EBMT team that developed the machine learning tool, to learn more about how the model was developed and its utility in predicting transplant risk in myelofibrosis.

Designing and Training the Machine Learning Model

To create the machine learning tool, investigators gathered data from adult patients with primary or secondary myelofibrosis who underwent their first allo-HCT between 2005 and 2020 in EBMT centers.2 A random survival forests model was then created using 10 variables: age, comorbidity index, performance status, blood blasts, hemoglobin, leukocytes, platelets, donor type, conditioning intensity, and graft-vs-host disease prophylaxis.

In total, data from 5183 patients from 288 centers were included to inform the machine learning model. The training cohort included 3887 patients, and the validation cohort included 1296 patients. Once the model was created, investigators performed a retrospective study to compare the performance of the new machine learning approach with a 4-level Cox regression-based score, other machine learning–based models derived from the same data set, and Center for International Blood and Marrow Transplant Research (CIBMTR) score.

“The study was conducted to explore the visibility and impact of machine learning tools in risk stratification for patients with myelofibrosis before allo-HCT because this is a very complex decision-making process and we wanted to develop a tool that could enhance the capacities of clinicians through datadriven approaches,” Orgueira said in an interview with OncologyLive. “We took advantage of the baseline data that [were] included in the registry to train a model that could predict the outcomes of these patients, both in terms of OS and relapse-free survival, to better understand whether these patients are [experiencing] worse outcomes due not only to disease-related factors but also due to toxicity.”

In the overall cohort, the median age at the time of allo-HCT was 58.3 years (range, 52.0-63.5). Most patients were male (62.6%), had primary myelofibrosis (72.2%), had constitutional symptoms at the time of allo-HCT (59.6%), had low-risk disease at the time of allo-HCT per the Dynamic International Prognostic Scoring System (2.4%), and had a low CIBMTR risk score at the time of allo-HCT (40.1%). Allo-HCT donor types were composed of identical sibling (29.6%), matched related donor other than a sibling (0.9%), mismatched related donor (6.5%), matched unrelated donor (42.0%), mismatched unrelated donor (13.0%), or unrelated with an unknown number of mismatches (8.0%).

EBMT Model Identifies High-Risk Patients for Transplant

At a median follow-up of 58.2 months (95% CI, 55.6-59.8) in the training set and 60.0 months (95% CI, 55.7-63.2) in the test set, the median OS was 79.4 months (95% CI, 69.2-89.6) and 73.7 months (95% CI, 54.7-92.7), respectively. No significant differences were reported between the 2 cohorts beyond a higher platelet count at the time of allo-HCT and a lower rate of antithymocyte globulin use in the test cohort.

In terms of overall transplantation outcomes, the estimated 1-, 5-, and 10-year OS rates were 70% (95% CI, 69%-71%), 53% (95% CI, 51%-54%), and 43% (95% CI, 41%-45%), respectively. These respective estimated progression- free survival rates were 62% (95% CI, 60%-63%), 44% (95% CI, 43%-46%), and 35% (95% CI, 33%-37%). The estimated nonrelapse mortality (NRM) rates were 23% (95% CI, 22%-24%), 32% (95% CI, 31%-33%), and 36% (95% CI, 35%-38%), respectively.

The EBMT model displayed higher concordance indices for OS and NRM in both the training and test sets compared with 3 other machine learning approaches. The EBMT model showed a significant rate of reassignment to other risk groups from the intermediate-2 risk group in terms of Cox score.

Notably, the EBMT machine learning model assigned 25% of patients as high risk with a poor outcome following transplantation (approximately 35% NRM and 40% overall mortality 1-year rates) compared with 10.1% using the Cox-derived score and 8.2% using the CIBMTR model. The 12- and 24-month OS rates in the training set (n = 471) of the machine learning high-risk group were 58.9% and 51.5%, respectively, compared with 58.3% and 52.7%, respectively, in the high-risk group identified by Cox-derived scores (n = 180). In the test set, these respective rates were 61.0% and 48.1% using the machine learning approach (n = 164) and 61.8% and 50.1%, respectively, using the Cox model (n = 55).

“CIBMTR scores identify 8% to 10% of patients as being high risk, but reality tells us [this number] is closer to 30%,” Orgueira noted. “When we identified the optimal threshold, our tool identified a [group of] 25% [of patients] who are performing very poorly [after allo-HCT]. That more than doubles the amount of high-risk patients that you can predict with other tools, and it is better aligned with what we see in the clinic.” Additionally, the machine learning model identified a larger high-risk group in terms of NRM compared with the Cox-derived score. In the test set, the 12- and 24-month NRM rates in the machine learning high-risk group were 36.4% and 42.6%, respectively. In the test set of the high-risk group via Cox-derived score, these respective rates were 36.0% and 42.8%.

The EBMT model also features an interactive web-based calculator that visualizes the risk score for a patient who is a candidate for allo-HCT. The calculator features inputs for the 10 variables used by the machine learning tool and produces OS and NRM curves, 1- and 2-year NRM and mortality rates, and a text and percentile display of the patient’s risk group.

“We currently have good [transplant risk projection] scoring systems, but you [also] have to consider the toxicity and the results of transplant, and we don’t have very good tools at this time [for that],” Hernández-Boluda said in an interview with OncologyLive. “With this [machine learning] tool, we will have information on the disease and transplantation risk, and we can share this information with the patient. We also have a web application that is visual, and we can share this image with the patient if they are interested. [Using this tool], we can make a more informed decision, which is critical because transplantation is a curative treatment, but also it poses a lot of danger [in terms of] the life of the patient. I believe combining what we had before with this tool will help us to make that decision.”

References

1. Machine learning program enhances transplant risk assessment in myelofibrosis patients better than current models. News release. American Society of Hematology. March 27, 2025. Accessed April 8, 2025. bit.ly/3RKIOPa

2. Hernandez-Boluda JC, Mosquera Orgueira A, Gras L, et al. Use of machine learning techniques to predict poor survival after hematopoietic cell transplantation for myelofibrosis. Blood. Published online March 27, 2025. doi:10.1182/blood.2024027287