A Hematopoietic Stem Cell Transplantation AI Foundation Model for Individualized Prediction of Relapse and Non-Relapse Mortality

Accurate individualized prediction of relapse and non-relapse mortality (NRM) after hematopoietic cell transplantation (HCT) remains challenging because it requires integration of clinical and transplant-related variables. The Center for International Blood and Marrow Transplant Research (CIBMTR) provides extensive publicly available datasets. However, their reuse at scale has been limited by heterogeneity in study populations and the restricted variable set within each dataset. Tabular foundation models are well suited to this setting. Their transformer-based architectures are pretrained across diverse tabular structures, enabling knowledge transfer across heterogeneous datasets. Such models have also outperformed conventional machine-learning and statistical methods. We developed HCTab-FM, an HCT-specific tabular foundation model trained on CIBMTR datasets, and integrated it into a pipeline generating individualized predictions from natural-language clinical descriptions.

Clinical and transplant-related variables were harmonized across 88 public CIBMTR datasets, yielding 243,738 observations for NRM and 226,740 for relapse. TabPFN-v3 was fine-tuned separately for each endpoint, and a censor-aware discrete-time survival layer generated individualized cumulative incidence curves. Performance was evaluated in 709 institutional HCT recipients against unfine-tuned TabPFN-v3, AutoGluon, CatBoost, XGBoost, and cause-specific Cox regression. For natural-language evaluation, 50 outcome-blinded clinical scenarios were processed by a large language model that extracted relevant variables, selected compatible CIBMTR datasets, mapped variables to study data dictionaries, and supplied mapped patient variables and matched CIBMTR records to HCTab-FM.

Internal CIBMTR validation yielded AUROCs of 0.86 for NRM and 0.74 for relapse. In the institutional cohort, HCTab-FM achieved AUROCs of 0.74 and 0.69, compared with 0.72 and 0.66 for unfine-tuned TabPFN-v3. AutoGluon, CatBoost, and XGBoost achieved AUROCs of 0.69/0.53, 0.61/0.55, and 0.57/0.50 for NRM/relapse, respectively. Survival C-indices were 0.72 for NRM and 0.73 for relapse, versus 0.62 and 0.72 with cause-specific Cox regression. Across 50 scenarios, observed and predicted 24-month cumulative incidences of NRM were 14.7% and 13.4%, respectively, yielding a predicted-to-observed cumulative subdistribution hazard ratio (HR) of 0.91 (95% CI, 0.67–2.19), where 1.0 denotes population-level calibration. For relapse, observed and predicted 24-month cumulative incidences were 22.0% and 16.6%, respectively, corresponding to a HR of 0.83 (95% CI, 0.62–1.91).

HCTab-FM demonstrates that heterogeneous public CIBMTR datasets can be integrated into an HCT-specific tabular foundation model for individualized prediction of relapse and NRM. This framework offers a scalable approach for developing individualized risk models across diverse HCT populations and transplant settings.