advanced deep learning enables prediction of allogenic stem cell mobilization success

Efficient mobilization of donor hematopoietic stem and progenitor cells (HSPCs) to peripheral blood (PB) is essential to the success of HSPC transplantation [1]. Although granulocyte colony-stimulating factor (G-CSF) is widely used to induce mobilization of HSPCs, healthy donors exhibit marked inter-individual variability in CD34⁺ cell yield [2,3,4]. Across institutions, variability may be in part due to differences in mobilization regimens, dosing, and apheresis procedures. Inadequate mobilization can compromise cell collection, and has consequences, including delayed engraftment, graft failure, and relapse. Donors may also face complications of prolonged G-CSF treatment and multiple apheresis sessions [5]. However, there remains no reliable method to identify these individuals prior to apheresis collection [6]. Prior efforts to predict mobilization outcome have largely been confined to single-time-point analyses with limited accuracy [2,3,4, 7]. Pre-G-CSF models, while useful for early risk stratification, may lack predictive resolution in borderline donors. Conversely, post-G-CSF models capture real-time biological response but offer only reactive interventions. A combined modeling strategy that leverages both pre- and post-G-CSF data offers the dual advantage of early prediction and refined risk assessment once G-CSF has begun. Here, we utilize baseline and post-mobilization laboratory values to improve accuracy of single-timepoint prediction models. We demonstrate the performance of advanced deep learning models that can flexibly weigh routine laboratory data from both baseline and post-GCSF timepoints to accurately predict poor mobilizers and support personalized decision-making throughout the mobilization timeline.