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.
Scientifica
|August 11, 2026
Summary
Machine learning survival models offer improved prediction for pediatric HSCT patients compared to Cox regression. Explainable AI provides transparent, clinically relevant risk assessments for better patient care.
Area of Science:
- Hematopoietic Stem Cell Transplantation (HSCT)
- Machine Learning in Medicine
- Survival Analysis
- Pediatric Oncology
Background:
- Clinical adoption of machine learning (ML) survival models in pediatric HSCT is hindered by concerns about information leakage, time-dependent structures, and interpretability.
- This study addresses these limitations by evaluating explainable survival modeling aligned with clinical decision points.
Purpose of the Study:
- To develop and validate an information-leakage-free, explainable survival modeling framework for pediatric HSCT.
- To compare the performance of ML models against Cox regression using a landmark approach.
- To assess the clinical interpretability and utility of ML-based survival predictions.
Main Methods:
- A single-center cohort of 839 transplant-naïve children (<14 years) undergoing HSCT was analyzed.
- A day +100 landmark cohort was constructed to prevent information leakage from post-transplant events.
- Cox proportional hazards, random survival forests (ranger), and randomForestSRC (rfSRC) were benchmarked using C-index, iAUC, and IBS.
- Model-agnostic explainability (survex) methods were employed for transparent risk assessment.
Main Results:
- ML models, particularly random survival forests, outperformed Cox regression in both baseline and landmark analyses.
- The landmark approach successfully prevented information leakage and improved model performance.
- Explainability analyses yielded clinically coherent global and individualized risk narratives.
Conclusions:
- An information-leakage-free, landmark-based explainable AI framework enables clinically credible survival prediction in pediatric HSCT.
- Random survival forests demonstrate superior performance over Cox regression, offering transparent explanations for clinical communication and risk updating.
