Deep Learning Models Enable Healthy Donor Management through Prediction of Mobilization Success

Hematopoietic stem cell (HSC) 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 mobilize adequately, despite standard granulocyte colony-stimulating factor (G-CSF)-based regimens. Early identification of such donors enables timely intervention, improving outcomes and reducing healthcare costs. Previous efforts to identify suboptimal donors have limited accuracy or are restricted to a single timepoint of data collection. Here, we utilize baseline and post-mobilization laboratory values to demonstrate the performance of advanced deep learning models that can flexibly weigh routine laboratory data from both timepoints. We compiled clinical and laboratory data from 1,160 healthy donors across multiple institutions, and a large CIBMTR dataset of healthy donors (n=19,207) to train and test predictive models. In the pre-G-CSF cohort (n=799), a transformer-based probabilistic model (TabPFN) trained on all 42 baseline variables achieved an accuracy of 89% and area under the receiving operator curve (AUC) of 0.96. Restriction of variables to just demographic and CBC features (14 variables) retained similar discrimination (accuracy=91%, AUC=0.96). Using the same 14 variables, TabPFN prediction in the post-GCSF cohort (n=361) yielded an accuracy of 94% (AUC=0.99). Utilizing the full cohort of 1,160 donors, we trained an attention-aware deep learning model incorporating data from both timepoints via lab-type indicator (0=pre-G-CSF, 1=post-G-CSF) achieving an accuracy of 81% (AUC of 0.89). To establish generalizability, we applied this model to a large CIBMTR dataset of healthy donors (n=19,207), achieving an accuracy of 79% (AUC=0.88). SHapley Additive exPlanations (SHAP) analysis highlighted white blood cell (WBC) count and differentials as the major predictors of mobilization. These interpretable models accurately predict poor mobilizers to enable both early donor triage and “just-in-time” rescue interventions, providing a data driven foundation for personalized donor mobilization strategies.