Juan Carlos Hernández-Boluda, MD, PhD
Allo-HCT (Allo-HCT) remains an important part of the therapeutic paradigm of Mier fibrosis, but the emergence of new treatment options and risk assessment considerations often determine that patients are ideal candidates for the implant complex. In response to these barriers, a team of investigators from the European Association for Blood and Bone Marrow Transplants (EBMT) has developed a machine learning model designed to identify and stratify transplant risks in patients with myelofibrosis.1
On March 27, 2025, EBMT announced the release of its machine learning model. Overall survival (OS) in myelofibrosis patients following Allo-HCT can be predicted based on patient characteristics such as age, performance status, and comorbidity index. Open Access Tools are now available for free online.
“I have been working on myelofibrosis transplants for almost 20 years, and despite that experience, it remains difficult to select the right time and patient to advance to the transplant,” MBBCH, MRCP, PhD and FRCPath said in an interview with Oncologylive. “This is even more [true] This year, there are many clinical trials that patients are interested in, and the whole host of new drugs that don't fully understand how to integrate them into transplant algorithms. This tool, along with many other factors we need to consider, will help determine the right patient to advance into implantation. ”
At the 51st Annual Meeting of EBMT, OncologyLive spoke with McLornan. Juan Carlos Hernández-Bluda, MD, PhD; Adrián Mosquera Orgueira, MD, PhD, who were also part of the EBMT team that developed machine learning tools, learned more about how models were developed and its usefulness in predicting the risk of myelofibrosis transplantation.
Machine learning model design and training
To create a machine learning tool, investigators collected data from adult patients with primary or secondary myelofibrosis who had undergone their first AroHCT between 2005 and 2020 at the EBMT centre. prevention.
In total, data from 5,183 patients from 288 centers were included, informing machine learning models. The training cohort included 3887 patients, and the validation cohort included 1296 patients. Once the models were created, investigators conducted a retrospective study to compare the performance of the new machine learning approach with four levels of COX regression-based scores, other machine learning-based models derived from the same data set, and International Center for Blood and Bone Marrow Transplantation Research (CIBMTR) scores.
“This study was conducted prior to Allo-HCT to investigate the visibility and impact of machine learning tools in risk stratification in patients with myelofibrosis. “We utilized that baseline data. [were] To better understand whether these patients are these patients, they are included in the registry to train models that can predict outcomes for these patients, both in terms of OS and recurrence-free survival. [experiencing] Not only disease-related factors, but toxicity outcomes also worsen. ”
In the overall cohort, the median age at Allo-HCT was 58.3 years (range, 52.0-63.5). Most patients were male (62.6%), had primary myelofibrosis (72.2%), constitutional symptoms at Allo-HCT (59.6%), had low risk disease at Allo-HCT per dynamic international prognostic scoring system (2.4%), and had low CIBMTR risk scores at Allo-HCT (40.1%). The type of Allo-HCT donor consisted of identical siblings (29.6%), non-sibling related donors (0.9%), inconsistent related donors (6.5%), related unrelated donors (42.0%), unrelated donors (13.0%), or unrelated number of unrelated numbers (8.0%).
EBMT models identify high-risk patients for implantation
Median OS was 79.4 months (95% CI, 69.2-89.6) and 73.7 months (95% CI, 54.7-92.7) at follow-up in the training set and 60.0 months (95% CI, 55.7-63.2) at the test set. No significant differences have been reported between the two cohorts with high platelet counts at Allo-HCT time points and the use of antimammary cell globulin in the test cohort.
Regarding the overall transplant outcome, 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%). The estimated progressive survival rates for each of these were 62% (95% CI, 60%-63%), 44% (95% CI, 43%-46%), and 35% (95% CI, 33%-37%). Estimated non-relapsed 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 match indexes for OS and NRM in both the training and test sets compared to the other three machine learning approaches. The EBMT model showed significant reallocation rates for other risk groups in the intermediate 2 risk group from the COX score perspective.
In particular, the EBMT machine learning model assigned 25% of patients as high risk, compared with 10.1% using COX-derived scores and 10.1% using CIBMTR model. For the high-risk groups identified by COX-derived scores (n = 180), the 12-month OS rates (n = 471) for the machine learning high-risk group training set (n = 471) were 58.9% and 51.5%, respectively. In the test set, the rates for each of these were 61.0% and 48.1% using the COX model (n = 55) using the machine learning approach (n = 164) and 61.8% and 50.1% respectively.
“CIBMTR scores identify 8% to 10% of patients as high risk, but the reality tells us. [this number] Orgeira said. “When we identified the optimum threshold, our tool identified A. [group of] twenty five% [of patients] A person who performs very poorly [after allo-HCT]. This more than doubles the amount of high-risk patients that other tools can predict and better match what we see at the clinic. Furthermore, the machine learning model identified a larger risk group with regard to NRM compared to COX-derived scores.
The EBMT model also features an interactive web-based calculator that visualizes risk scores for patients who are candidates for Allo-HCT. The calculator features inputs of 10 variables used in machine learning tools and generates text and percentile displays for OS and NRM curves, NRM and mortality rates for 1 and 2 years, and patient risk groups.
“We're good now [transplant risk projection] Scoring system, but you [also] You need to consider toxicity and transplant outcomes, and there are no good tools at this time [for that],” Hernandez Boruda said in an interview with OncologyLive. [machine learning] There is information about tools, disease and transplant risk, and this information can be shared with patients. There is also a visual web application that allows you to share this image with your patients if you are interested. [Using this tool]we can make more informed decisions. This is important because transplantation is a therapeutic treatment, but it poses many risks [in terms of] The life of the patient. I think combining what I had before with this tool will help you make that decision. ”
reference
1. Machine learning programs enhance transplant risk assessments in patients with myelofibrosis over current models. News release. American Society of Hematology. March 27th, 2025. Accessed April 8, 2025. bit.ly/3rkiopa
2. Hernandez-Boluda JC, Mosquera Orgueira A, Gras L, et al. The use of machine learning techniques to predict poor survival after hematopoietic cell transplantation for myelofibrosis. blood. Released online on March 27th, 2025. doi:10.1182/blood.2024027287
