Abstract
A large number of association studies have related donor characteristics to survival after bone marrow transplantation, for leukemia in general and specifically for acute myeloid leukemia (AML) patients. However, population-based differences often do not hold at the single transplant level. We test whether transplantation outcomes can be predicted at the single-patient level and whether such predictions can be used to better choose donors. The analysis was performed on a mixture of different diseases or with AML only, and with either patient and donor information or donor information only. We analyzed 3671 8-of-8 HLA-matched AML donor-recipient pairs and tested whether the outcome, including 1-year total and event-free survival, can be predicted from patient and donor-related factors. We used multiple machine learning and survival analysis methods. The best method is a fully connected neural network. Multiple outcomes can be predicted, with area under the specificity-sensitivity curve (AUC) values between 0.54 and 0.67 for the different outcomes. The patient age has a strong impact on prediction. However, for a given patient, when only donor or transplant information is used, limited prediction accuracy of 0.54 to 0.56 AUC for event-free survival and survival is obtained. Graft-versus-host disease and rejection after 1 year have slightly higher AUC values of around 0.59, whereas the relapse prediction accuracy was random. All donors' characteristics have a limited influence on the quality of hematopoietic stem cell transplantation for fully matched donors. Many factors with a population effect on survival have a very limited effect when combined with all other factors in a single-donor predictive model.