Challenging a Clinical Dogma with Multimodal Machine Learning: A Retrospective Analysis of Transplant Mismatched Donor Selection

The “younger is better” maxim for hematopoietic cell transplantation (HCT) donor selection often drives risky treatment delays. We challenged this dogma in the modern era of post-transplantation cyclophosphamide (PTCy) prophylaxis. Using advanced machine learning models on a CIBMTR registry cohort of 5,143 patients undergoing haploidentical or mismatched unrelated donor (MMUD) HCT with PTCy, we quantified the precise impact of donor age on outcomes. We found the survival penalty from an older donor is modest and varies by donor type; for haploidentical HCT, a donor's age had to reach 42 years to confer just a 10-day loss in 2-year survival compared to an 18-year-old donor. MMUDs were less sensitive to donor age, with no equivalent survival decrement observed for donors up to 50 years old. While increasing donor age was independently associated with a higher risk of graft-versus-host disease, it had no effect on relapse. Our findings demonstrate that the small survival benefit of a younger donor must be weighed against the significant morbidity of graft-versus-host disease and the well-documented harms of treatment delay. This quantitative framework supports the practice of proceeding with a suitable, available donor rather than waiting for a marginally younger one.