AUTHORS
Jiasheng Wang, MD, Duoduo Zhao, MD2, Dhriti Sood, MBBS3, David Swoboda, MD4, Aziz Nazha, MD5
AFFILIATIONS
1 The Ohio State University Comprehensive Cancer Center, Columbus, United States
2 The Ohio State University Comprehensive Cancer Center, Columbus, United States
3 Jefferson Einstein Philadelphia Hospital, Philadelphia, United States
4 Tampa General Cancer Institute, Tampa, United States
5 Thomas Jefferson University, Philadelphia, United States
ABSTRACT
Background: Accurate individualized prediction of relapse and non-relapse mortality (NRM) after hematopoietic cell transplantation (HCT) remains challenging, as it requires the integration of detailed 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 an HCT-specific tabular foundation model (HCTab-FM), trained on nearly half a million endpoint-specific records from CIBMTR datasets. We further integrated it into an end-to-end pipeline that generates individualized post-HCT outcome predictions directly from natural-language clinical descriptions.
Methods: We first harmonized clinical and transplant-related variables across 88 publicly available CIBMTR datasets. These datasets yielded 243,738 patient-level observations evaluable for the NRM endpoint and 226,740 evaluable for the relapse endpoint. A tabular foundation model, TabPFN-v3, was fine-tuned separately for each endpoint to generate HCTab-FM. To extend the model from fixed-time risk prediction to time-to-event analysis, we incorporated a censor-aware discrete-time survival layer, yielding individualized cumulative incidence curves for NRM and relapse. Model performance was evaluated in an institutional cohort of 709 HCT recipients, with HCTab-FM benchmarked against TabPFN-v3, CatBoost, XGBoost, AutoGluon, and cause-specific Cox regression.
To assess prediction from natural-language input, we converted 50 institutional cases into outcome-blinded clinical scenarios (e.g., "a 68-year-old woman with NPM1-mutated AML in CR3, KPS 60, atrial fibrillation, and EF 45% undergoing matched-sibling peripheral-blood HCT with fludarabine/melphalan 100 and PTCy"). A large language model extracted the relevant variables, selected specific CIBMTR datasets with compatible variables, and mapped each case to the corresponding data dictionaries. The mapped patient variables and matched CIBMTR records were then supplied to HCTab-FM to generate individualized relapse and NRM predictions. Model code and weights will be released publicly.
Results: On internal validation using CIBMTR data, HCTab-FM achieved AUROCs of 0.86 for NRM and 0.74 for relapse. In the held-out institutional cohort, HCTab-FM achieved AUROCs of 0.74 for NRM and 0.69 for relapse, versus 0.72 and 0.66, respectively, for unfine-tuned TabPFN-v3. HCTab-FM also outperformed AutoGluon, CatBoost, and XGBoost, which achieved AUROCs of 0.69, 0.61, and 0.57 for NRM and 0.53, 0.55, and 0.50 for relapse, respectively. In time-to-event analyses, the censor-aware HCTab-FM achieved C-indices of 0.72 for NRM and 0.73 for relapse, versus 0.62 and 0.72, respectively, for cause-specific Cox regression.
In the scenario-based evaluation across 50 clinical 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).
Conclusion: 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.
DISCLOSURES
Jiasheng Wang, MD: No relevant conflicts of interest to declare.