Uncoupling toxicity from efficacy in donor selection for allogeneic hematopoietic cell transplantation: a retrospective cohort study using double machine learning

Haploidentical hematopoietic cell transplantation (HCT) offers potent graft-versus-tumor effects but is historically limited by higher non-relapse mortality (NRM) compared to matched unrelated (MUD) or related donors (MRD). To algorithmically decouple the competing risks of NRM and relapse, we analyzed 1,713 adult patients (MD Anderson discovery cohort) and 6,829 patients (registry replication cohorts) using a double machine-learning framework of Causal Survival Forests to estimate individualized treatment effects, while Random Survival Forests defined absolute prognostic risk. We found biological uncoupling of relapse and NRM: relapse was driven by disease biology, while NRM was governed by physiological reserve. Specifically, cardiovascular or cerebrovascular comorbidities acted as selective amplifiers of haploidentical toxicity-a "vascular penalty" that doubled mortality risk in the bootstrapped model (sub-distribution hazard ratio 2.09, 95% confidence interval 1.26-3.35). We developed a 3-factor nomogram (recipient age, vascular comorbidity, B-leader matching status) to identify an "Optimized Haploidentical" phenotype. Exploratory multivariable modeling suggested MUD consistently maximized predicted disease-free survival across all subgroups. Among haploidentical versus MRD comparisons, B-leader mismatch appeared to favor MRD across ages, while B-leader matched haploidentical donors in older patients (≥ 60 years) modeled near-equivalent DFS to MRD, representing a hypothesis-generating "zone of equivalence" requiring prospective validation.