Explainable Survival Modeling in Pediatric Hematopoietic Stem Cell Transplantation Using Landmark-Based Machine
Khawar Siddiqui1, Abdulrahman Al-Musa1, Mona Al-Saleh2
1Department of Pediatric Hematology/Oncology, King Faisal Specialist Hospital and Research Center, Riyadh, Saudi Arabia, kfshrc.edu.sa.
Background:
Machine learning survival models can outperform conventional Cox regression in heterogeneous pediatric HSCT cohorts, but clinical adoption is limited by concerns regarding information leakage, time-dependent structure, and interpretability. We evaluated explainable survival modeling aligned to real-world decision points before transplantation and after early post-transplant complications.
Materials And Methods:
In a single-center cohort of 839 transplant-naïve children (< 14 years) undergoing HSCT (2010-2016), overall survival was modeled from infusion to death or last follow-up. A baseline model used only pretransplant variables (age at HSCT, primary diagnostic group, stem cell source, and donor HLA type). To prevent leakage from postinfusion events, a day + 100 landmark cohort (n = 721; 121 deaths beyond day + 100) was constructed and extended with early complications known by day + 100 (SOS/VOD and acute GvHD). Cox proportional hazards, random survival forests (ranger), and randomForestSRC (rfSRC) were benchmarked using C-index, integrated cumulative/dynamic AUC (iAUC), and integrated Brier score (IBS). Cox proportional hazard assumptions were tested using Schoenfeld residuals. Model-agnostic explainability (survex) included time-dependent permutation importance, partial dependence survival profiles, SurvSHAP(t), ceteris paribus profiles, and additive attributions (breakdown and SHAP).
Results:
Overall mortality was 28.1% (236/839). Baseline Cox violated proportional hazards (global p = 0.00181), whereas the day + 100 landmark Cox did not (global p = 0.52). Baseline performance favored ML models (C-index: Cox 0.673, ranger 0.677, rfSRC 0.668; iAUC: 0.698, 0.866, 0.826; IBS: 0.159, 0.126, 0.135). In the landmark setting, discrimination and accuracy improved with ML (C-index: Cox 0.721, ranger 0.870, rfSRC 0.840; iAUC: 0.609, 0.668, 0.650; IBS: 0.125, 0.0969, 0.1016). Explainability analyses produced clinically coherent global and individualized risk narratives beyond day + 100.
Conclusion:
An information-leakage-free, landmark-based XAI framework allows clinically credible survival prediction in pediatric HSCT, with random survival forests outperforming Cox regression and providing transparent cohort-level and patient-level explanations suitable for clinical communication and risk updating.
