: The selection of a donor for allogeneic hematopoietic stem cell transplantation (HSCT) is a complex, multifactorial process requiring the balance of traditional human leukocyte antigen (HLA) compatibility with non-HLA factors including modern pharmacological approaches such as post-transplant high-dose cyclophosphamide (PTCy) for graft-versus-host disease (GvHD) prophylaxis. This work addresses the resulting need for interpretable, data-driven decision support by developing a comprehensive framework based on a Bayesian hierarchical Accelerated Failure Time (AFT) survival model. Trained on a multicenter national registry (Gruppo Italiano per il Trapianto di Midollo Osseo, cellule staminali emopoietiche e terapia cellulare, GITMO, n=7,609 transplants from 85 centers; 2011-2024), the model directly models survival times using a log-normal distribution, explicitly accounts for heteroscedasticity, and incorporates center-level random effects. The core output is an interactive Shiny web application enabling transplant physicians and search coordinators to compare multiple donor candidates through personalized survival curves with Bayesian credible intervals. Model validation demonstrates robust performance, with a concordance index of 0.64 and strong discriminative ability at the clinically relevant one-year mark (AUC: 0.71 at 12 months; Brier score: 0.187). A key finding is the quantification of how PTCy substantially mitigates the survival penalty otherwise associated with mismatched unrelated donors (MMUD), with an acceleration factor of 1.57 (95% credible interval: 1.12-2.16) favoring PTCy over ATG in this setting. Both patient age (AF 0.64; 36% reduction in median survival per standard deviation) and donor age (AF 0.90; 10% reduction per SD) are independently significant. By bridging rigorous Bayesian survival modeling with clinical usability and explainability via SHAP values, this work provides a practical tool to navigate the complexities of contemporary donor selection and support more personalized, evidence-based decisions.
A Bayesian Survival Model to Support the Selection of the Best Hematopoietic Stem Cell Donor
Malagola, MicheleWriting – Review & Editing
;Polverelli, NicolaWriting – Review & Editing
;
2026-01-01
Abstract
: The selection of a donor for allogeneic hematopoietic stem cell transplantation (HSCT) is a complex, multifactorial process requiring the balance of traditional human leukocyte antigen (HLA) compatibility with non-HLA factors including modern pharmacological approaches such as post-transplant high-dose cyclophosphamide (PTCy) for graft-versus-host disease (GvHD) prophylaxis. This work addresses the resulting need for interpretable, data-driven decision support by developing a comprehensive framework based on a Bayesian hierarchical Accelerated Failure Time (AFT) survival model. Trained on a multicenter national registry (Gruppo Italiano per il Trapianto di Midollo Osseo, cellule staminali emopoietiche e terapia cellulare, GITMO, n=7,609 transplants from 85 centers; 2011-2024), the model directly models survival times using a log-normal distribution, explicitly accounts for heteroscedasticity, and incorporates center-level random effects. The core output is an interactive Shiny web application enabling transplant physicians and search coordinators to compare multiple donor candidates through personalized survival curves with Bayesian credible intervals. Model validation demonstrates robust performance, with a concordance index of 0.64 and strong discriminative ability at the clinically relevant one-year mark (AUC: 0.71 at 12 months; Brier score: 0.187). A key finding is the quantification of how PTCy substantially mitigates the survival penalty otherwise associated with mismatched unrelated donors (MMUD), with an acceleration factor of 1.57 (95% credible interval: 1.12-2.16) favoring PTCy over ATG in this setting. Both patient age (AF 0.64; 36% reduction in median survival per standard deviation) and donor age (AF 0.90; 10% reduction per SD) are independently significant. By bridging rigorous Bayesian survival modeling with clinical usability and explainability via SHAP values, this work provides a practical tool to navigate the complexities of contemporary donor selection and support more personalized, evidence-based decisions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
