Novel machine learning technique further clarifies unrelated donor selection to optimize transplantation outcomes

Abstract

Abstract: We investigated the impact of donor characteristics on outcomes in allogeneic hematopoietic cell transplantation (HCT) recipients using a novel machine learning approach, the Nonparametric Failure Time Bayesian Additive Regression Trees (NFT BART). NFT BART models were trained on data from 10 016 patients who underwent a first HLA-A, B, C, and DRB1 matched unrelated donor (MUD) HCT between 2016 and 2019, reported to the Center for International Blood and Marrow Transplant Research, then validated on an independent cohort of 1802 patients. The NFT BART models were adjusted based on recipient, disease, and transplant variables. We defined a clinically meaningful impact on overall survival (OS) or event-free survival (EFS; survival without relapse, graft failure, or moderate to severe chronic graft-versus-host disease) as >1% difference in predicted outcome at 3 years. Characteristics with <1% impact (within a zone of indifference) were not considered to be clinically relevant. Donor cytomegalovirus, parity, HLA-DQB1, and HLA-DPB1 T-cell epitope matching fell within the zone of indifference. The only significant donor factor that associated with OS was age, in which, compared with 18-year-old donors, donors aged ≥31 years old were associated with lower OS. Both donor age (≤32 years) and use of a male donor, regardless of recipient sex, improved EFS. We, therefore, recommend selecting the earliest available donor within the 18 to 30 years age range for HCT to optimize OS. If several donors in the 18 to 30 years age range are available, a male donor may be chosen to optimize EFS.