The ability to predict future outcomes of deceased-donor kidney grafts improves allocation decision-making for transplant clinicians. To improve the UK transplant selection process, we set out to utilize novel artificial intelligence (AI) algorithms to develop improved risk stratification.
The United Kingdom Transplant Registry (UKTR) database was used to analyse pre-transplant variables from 29,714 deceased-donor kidney transplants carried out between 2008 and 2022. Overall graft survival served as the primary performance metric. We tested four machine learning models that were evaluated for calibration and discrimination using the integrated Brier score (IBS) and Harrell's concordance index. We assessed the potential clinical utility using decision curve analysis. External validation and Health inequalities were evaluated by assessing the model metrics on patients with higher and lower deprivation scores.
The IBS score of the XGBoost model was 0.14, demonstrating accurate calibration. At 3, 5, 7, and 9 years after transplant, among all the involved AI algorithms, XGBoost gave the best discriminative performance for survival (AUC=0.74, 0.75, 0.76, and 0.75, respectively) along with a concordance index of 0.74. When applied to the same cohort, the UK kidney donor risk index (KDRI) had a concordance of only 0.62, with AUC scores 0.61 at 3 years post-transplant, 0.60 at 5 years, 0.62 at 7 and 9 years.
This novel model, termed the United Kingdom Deceased-Donor Kidney Transplant Outcome Prediction (D-TOP), can potentially optimize deceased donor selection with a better prediction of overall-graft survival and improving the effectiveness of kidney allocation schemes.