Optimal Donor Selection for Hematopoietic Cell Transplantation Using Bayesian Machine Learning

Abstract

Purpose: Donor selection practices for matched unrelated donor (MUD) hematopoietic cell transplantation (HCT) vary, and the impact of optimizing donor selection in a patient-specific way using modern machine learning (ML) models has not been studied.

Methods: We trained a Bayesian ML model in 10,318 patients who underwent MUD HCT from 1999 to 2014 to provide patient- and donor-specific predictions of clinically severe (grade 3 or 4) acute graft-versus-host disease or death by day 180. The model was validated in 3,501 patients from 2015 to 2016 with archived records of potential donors at search. Donor selection optimizing predicted outcomes was implemented over either an unlimited donor pool or the donors in the search archives. Posterior mean differences in outcomes from optimal donor selection versus actual practice were summarized per patient and across the population with 95% intervals.

Results: Event rates were 33% (training) and 37% (validation). Among donor features, only age affected outcomes, with the effect consistent regardless of patient features. The median (interquartile range) difference in age between the youngest donor at search and the selected donor was 6 (1-10) years, whereas the number of donors per patient younger than the selected donor was 6 (1-36). Fourteen percent of the validation data set had an approximate 5% absolute reduction in event rates from selecting the youngest donor at search versus the actual donor used, leading to an absolute population reduction of 1% (95% interval, 0 to 3).

Conclusion: We confirmed the singular importance of selecting the youngest available MUD, irrespective of patient features, identified potential for improved HCT outcomes by selecting a younger MUD, and demonstrated use of novel ML models transferable to optimize other complex treatment decisions in a patient-specific way.

FIG 1.

Donor age effect on outcome. (A) Risk of acute GVHD grade 3 or 4 or death by 180 days versus donor age, for 40 different risk profiles of patient characteristics known at the time of search initiation, specifically recipient age (18-29, 30-39, 40-49, 50-59, and 60+), sex, and disease (AML, ALL, CML, and MDS). Predictions come from the BART model on the training data and are averaged over the distribution of other patient, disease, and transplant characteristics in the training cohort. (B) Donor age (median and quartiles) versus year of transplant in combined training and validation data sets. ALL, acute lymphoblastic leukemia; AML, acute myeloid leukemia; BART, Bayesian Additive Regression Trees; CML, chronic myeloid leukemia; GVHD, graft-versus-host disease; MDS, myelodysplasia; Tx, transplant.

FIG 2.

Impact of optimal donor selection in training cohort. For training cohort, we evaluated optimal donor selection under unlimited donor availability, so that donor of age 18 years was optimal for each patient. (A) Posterior mean (blue), interquartile range (light blue), and 95% intervals (very light blue) for absolute risk difference for optimal age 18-year-old donor versus actual donor. (B) Posterior probability that donor of age 18 years has a better outcome than actual donor. (C) Posterior distribution of the population-level risk for donor of age 18 years and actual donor. (D) Posterior distribution of the difference in population-level risk for donor of age 18 years versus actual donor. (A and B) Plotted in increasing order according to the magnitude of the posterior mean improvement (more negative means improved outcomes).

FIG 3.

Impact of optimal donor selection in validation cohort. For validation cohort, we evaluated optimal donor selection under either unlimited donor availability (donor of age 18 years for each patient), or by selecting the optimal (youngest) donor from among the available donors in the search archive for each patient. (A) Posterior mean (blue), interquartile range (IQR; light blue), and 95% intervals (very light blue) for absolute risk difference for donor of age 18 years versus actual donor. (B) Posterior probability that donor of age 18 years has a better outcome than actual donor. (C) Posterior mean (blue), IQR (light blue), and 95% intervals (very light blue) for absolute risk difference for optimal donor from search archive versus actual donor. (D) Posterior probability that optimal donor from search archive has a better outcome than actual donor. (E) Posterior distribution of the population level risk for optimal donor from search archive, donor of age 18 years, and actual donor. (F) Posterior distribution of the difference in population level risk for optimal donor from search archive or for donor of age 18 years versus actual donor. (A-D) Plotted in increasing order according to the magnitude of the posterior mean improvement (more negative means improved outcomes).

FIG A1.

Margins of error for risk predictions by donor age. Margins of error for predictions of risk of acute GVHD grade 3 or 4 or death by 180 days versus donor age, for 40 different risk profiles of patient characteristics known at the time of search initiation, specifically recipient age (18-29, 30-39, 40-49, 50-59, and 60+), sex, disease (AML, ALL, CML, and MDS). Predictions come from the BART model on the training data and are averaged over the distribution of other patient, disease, and transplant characteristics in the training cohort. ALL, acute lymphoblastic leukemia; AML, acute myeloid leukemia; BART, Bayesian Additive Regression Trees; CML, chronic myeloid leukemia; GVHD, graft-versus-host disease; MDS, myelodysplasia.

FIG A2.

Calibration plots of the observed outcome probability versus the predicted outcome probability from the BART model, with LOESS smooth and 95% pointwise CIs. Predicted outcome from BART uses a shifted intercept to account for the increased overall event rate in the validation cohort compared with the training data set. BART, Bayesian Additive Regression Trees; LOESS, Locally Weighted Smoothing.