Machine learning models trained on synthetic patient data achieved a concordance index of 0.709 on baseline synthetic data and 0.617 when tested against real patient records, reflecting a performance drop of 0.092 under domain shift, according to an evaluation from an international data science competition.
The initiative sought to validate whether synthetic cohorts could serve as viable stand-ins for sensitive clinical data while enabling rigorous, equity-aware evaluation across diverse patient groups.
To establish the challenge benchmarks, the investigation deployed generative modeling techniques to fabricate realistic post-transplant patient records. The baseline competition dataset was generated using SurvivalGAN, while an enhanced ensemble pipeline combined AdsGAN, SurvivalGAN, Survival-CTGAN, and normalizing flows alongside adaptive demographic balancing.
The research team subjected the generated cohorts to validation protocols assessing distributional realism, statistical utility, and cross-domain transferability. The analytical framework incorporated Kaplan–Meier survival curves alongside Kolmogorov–Smirnov and log-rank statistical tests to evaluate distribution alignment between synthetic and observed records. Realism and utility were directly examined through train-on-synthetic and test-on-real comparisons benchmarked against real-only baselines, specifically testing how well survival algorithms transferred across the domain shift from baseline synthetic data to real-world patient records.
A key focus of the competition involved evaluating subgroup equity across distinct patient populations. Participants were tasked with engineering equity-aware survival models capable of delivering fair outcomes across demographic groups. The models were evaluated using specialized fairness metrics, specifically the race-stratified concordance index, denoted as MMSD, alongside a post-hoc robustness metric, denoted as SSD. These evaluation criteria enabled the researchers to measure subgroup predictive performance stratified directly by race group and to monitor model equity under domain transfer.
The collaborative benchmarking event achieved substantial global participation, attracting 11,311 registrants representing 108 countries. While top-performing survival algorithms demonstrated solid internal accuracy on the baseline synthetic benchmark with a concordance index of 0.709, their discriminative ability declined to 0.617 upon real-world testing. Furthermore, the analysis revealed that the equity advantages observed when training and testing within synthetic cohorts attenuated under the domain shift to real clinical data, highlighting the persistent challenges of transferring fairness properties between synthetic environments and clinical reality.
A preliminary privacy evaluation recorded a membership inference area under the receiver operating characteristic curve of 0.512 and a nearest-neighbor distance ratio of 1.44, indicating no empirical evidence of training-record memorization.
Subject of Research: Evaluation of synthetic datasets and equity-aware survival modeling in hematopoietic cell transplantation
Article Title: Survival modeling on synthetic post-HCT data: validation and equity-aware evaluation from a CIBMTR–Kaggle data science competition
Article References: Deshpande, T., Akdemir, D., Reade, W., & Bolon, Y.-T. (2026). Survival modeling on synthetic post-HCT data: validation and equity-aware evaluation from a CIBMTR–Kaggle data science competition. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03925-z
Image Credits: AI Generated
DOI: 10.1186/s12911-026-03925-z
Keywords: hematopoietic cell transplantation, survival modeling, synthetic data, CIBMTR, Kaggle, SurvivalGAN, Survival-CTGAN, AdsGAN, normalizing flows, domain shift, algorithmic fairness, concordance index, membership inference, data privacy, open science, demographic balancing, donor selection
News Source: Denise Maddox. (November 14, 2026). CIBMTR Kaggle Competition Validates Synthetic Data for Equity-Aware Post-Transplant Survival Modeling. Scienmag.



