Survival modeling on synthetic post-HCT data: validation and equity-aware evaluation from a CIBMTR–Kaggle data science competition
Year: 2026
Type: article
Abstract: Survival modeling in Hematopoietic Cell Transplantation (HCT) informs donor selection, conditioning regimens, and post-transplant care. The Center for International Blood and Marrow Transplant Research® (CIBMTR®) partnered with Kaggle to run a data science competition using synthetic HCT datasets. We evaluated (i) synthetic fidelity and transferability using train-on-synthetic/test-on-real comparisons alongside real-only baselines, and (ii) subgroup equity using the race-stratified concordance index (MMSD) and a post-hoc robustness metric (SSD) under domain shift. Synthetic cohorts were generated using SurvivalGAN for the baseline competition dataset and an enhanced ensemble pipeline combining AdsGAN, SurvivalGAN, Survival-CTGAN, and normalizing flows with adaptive demographic balancing. Validation employed Kaplan–Meier curves, statistical tests (Kolmogorov–Smirnov and log-rank), realism/utility (TSTR/TRTR‑style transfer) and domain shift assessment (Baseline \(\rightarrow\) Real). Participants developed equity-aware survival models evaluated via fairness metrics (MMSD; SSD post hoc) with subgroup performance stratified by race group. The competition attracted 11,311 registrants from 108 countries. Top models achieved C-index 0.709 on Baseline \(\rightarrow\) Baseline and 0.617 on Baseline \(\rightarrow\) Real ( \(\Delta = -0.092\) ). Equity advantages observed on synthetic data attenuated under domain shift to real data. Synthetic data are feasible for privacy-preserving, equity-focused research in HCT. The competition enabled open science and international participation, consistent with Sustainable Development Goals [1] (primarily SDG 9). Preliminary privacy analysis (membership inference AUC = 0.512; nearest-neighbor distance ratio = 1.44) found no evidence of training-record memorization. Not applicable.
Institutions: National Marrow Donor Program, Google (United States), Kaggle
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,990
OpenAlex ID: W7229506558
