BREAKING THE BLACK BOX IN HEMATOPOIETIC STEM CELL TRANSPLANTATION: AN EXPLAINABLE MACHINE LEARNING WEB APPLICATION FOR PERSONALIZED SURVIVAL PREDICTION

@article{stelet2026breaking,
  title = {{BREAKING THE BLACK BOX IN HEMATOPOIETIC STEM CELL TRANSPLANTATION: AN EXPLAINABLE MACHINE LEARNING WEB APPLICATION FOR PERSONALIZED SURVIVAL PREDICTION}},
  author = {Vinicius Navega STELET and Eliana ABDELHAY and Renata BINATO and Anna Beatriz Fusco de Souza FERREIRA and Diogo Antonio TSCHOEKE and Leticia Martins RAPOSO},
  journal = {Hematology, Transfusion and Cell Therapy},
  year = {2026},
  month = oct,
  doi = {10.1016/j.htct.2026.106832},
  url = {https://doi.org/10.1016/j.htct.2026.106832},
  note = {CC BY},
}

TY  - JOUR
TI  - BREAKING THE BLACK BOX IN HEMATOPOIETIC STEM CELL TRANSPLANTATION: AN EXPLAINABLE MACHINE LEARNING WEB APPLICATION FOR PERSONALIZED SURVIVAL PREDICTION
AU  - Vinicius Navega STELET
AU  - Eliana ABDELHAY
AU  - Renata BINATO
AU  - Anna Beatriz Fusco de Souza FERREIRA
AU  - Diogo Antonio TSCHOEKE
AU  - Leticia Martins RAPOSO
PY  - 2026
DA  - 2026/10/23
JO  - Hematology, Transfusion and Cell Therapy
DO  - 10.1016/j.htct.2026.106832
UR  - https://doi.org/10.1016/j.htct.2026.106832
AB  - Background: Allogeneic hematopoietic stem cell transplantation (allo-HSCT) is a major curative therapy for acute leukemias and myelodysplastic syndromes, yet post-transplant mortality remains a clinical challenge. Traditional prognostic tools are easily interpretable but limited by strict statistical assumptions, such as linearity and proportional hazards. Machine learning (ML) models overcome these limitations by capturing complex, non-linear interactions, but often at the cost of interpretability. Explainable AI (XAI) bridges this gap, enabling clinicians to assess the magnitude and direction of individual variables for personalized predictions, thereby increasing clinical applicability. Objective: To develop an accessible web-based tool driven by an ML model to predict and classify survival risk after allo-HSCT based on pre-transplant characteristics, providing local explainability for individualized clinical decision support. Methods: We analyzed a retrospective multicenter cohort of 16,458 pediatric and adult patients who underwent allo-HSCT (up to one HLA mismatch) for acute leukemias or MDS. Pre-HSCT variables were selected using the Boruta algorithm. An Extreme Gradient Boosting (XGBoost) model, adapted for survival analysis, was trained and tested using a 67/33 data split. For risk stratification, we utilized risk percentiles, quartiles, and Kaplan-Meier survival curves derived from the training set. SHapley Additive exPlanations (SHAP), including waterfall and force plots, were implemented to demonstrate local feature contributions. The model and the web application were developed in R using the Shiny framework. Results: Eleven pre-HSCT predictors were selected: recipient and donor age, conditioning regimen, primary disease, disease stage, graft type, GVHD prophylaxis, HCT-CI, time from diagnosis to HSCT, KPS, and HLA match level. The XGBoost model achieved a concordance index of 0.622 (95% CI: 0.612 - 0.632) in the test set. The model was successfully deployed via the ''HCT Survival Explorer'' web app (https://steletvinicius.shinyapps.io/HSCTSurvxAI/). Through the application, clinicians can input the 11 predictive variables to simulate specific HSCT scenarios, immediately obtaining the patient’s risk context (survival function and risk percentiles) and the individualized magnitude of each variable's contribution via local SHAP analysis. Conclusion: Risk stratification is crucial for guiding therapeutic choices in allo-HSCT. Our XAI-based web application, trained on robust multicenter data, serves as a powerful translational tool. By integrating multiple clinical factors and breaking the black box of predictive modeling, it provides personalized, transparent prognostic insights that directly support bedside clinical decision-making.
SN  - 2531-1379
SN  - 2531-1387
ER  -