Addressing the translation gap in SCT donor selection

Machine learning (ML) offers genuine promise for improving donor selection in hematopoietic stem cell transplantation (SCT), yet the gap between model development and clinical adoption remains wide. Drawing on experience from a German federally-funded translational incubation project at the Berlin Institute of Health and Charité Hospital (2021-2023), we offer a practitioner's perspective on what it actually takes to move ML-driven SCT tools from research into clinical practice. We make two distinct contributions. First, we offer concrete prescriptions on data strategy and translational methodology: registry data was not designed for ML and must be rethought from the ground up; multi-center data infrastructure is a prerequisite, not an aspiration; approaches which increase the breadth of clinical parameters are theoretically sound and should be embraced; and the field must move beyond retrospective validation toward prospective studies designed around standard-of-care endpoints. Second, we provide deliberate sensitization to three forces that will surprise clinically-focused teams: the product and regulatory burden of sustaining a clinical tool; the true resource requirements for the full translational path; and the market dynamics and misaligned institutional incentives that impede adoption. Despite these barriers, the technical feasibility of ML-supported SCT donor selection is real, and the community's demonstrated willingness to share data provides the foundation on which to build. What is needed now is not more retrospective modelling, but projects that enter the space with a clear-eyed understanding of the full landscape.