Hematopoietic stem and progenitor cell (HSPC) transplantation offers a potentially curative therapy for aggressive hematologic malignancies and bone marrow failure syndromes. Successful transplantation depends on effective mobilization of donor CD34 + cells, yet some healthy donors fail to achieve adequate CD34 + yields despite standard granulocyte colony–stimulating factor (G–CSF)–based regimens. Early identification of such donors enables timely intervention, improving transplantation outcomes and reducing healthcare costs. We analyzed demographic and pre– and post–G–CSF laboratory data from 1,160 healthy donors from across multiple institutions and developed two complementary machine–learning frameworks to predict mobilization outcome. A transformer–based probabilistic model (TabPFN) trained on baseline complete blood counts (CBCs) rigorously discriminates poor from good mobilizers. Applying the same architecture to donor data after mobilization attains near–perfect discrimination. To unify the predictions across time points, we introduce an attention–aware neural network that ingests either baseline or post–mobilization data via a ″lab–type″ context flag, enabling accurate prediction of poor mobilizers both before and after G–CSF mobilization. We further validated the framework on data from over 19,000 healthy donors compiled by the Center for International Blood and Marrow Transplant Research. These interpretable models enable early triage and ″just–in–time″ rescue interventions, providing a data driven foundation for personalized donor mobilization strategies.
Adil et al. (Sat,) studied this question.