Artificial intelligence assisted risk prediction in organ transplantation: a UK Live-Donor Kidney Transplant Outcome Prediction tool

Abstract

Introduction: Predicting the outcome of a kidney transplant involving a living donor advances donor decision-making donors for clinicians and patients. However, the discriminative or calibration capacity of the currently employed models are limited. We set out to apply artificial intelligence (AI) algorithms to create a highly predictive risk stratification indicator, applicable to the UK's transplant selection process.

Methodology: Pre-transplant characteristics from 12,661 live-donor kidney transplants (performed between 2007 and 2022) from the United Kingdom Transplant Registry database were analyzed. The transplants were randomly divided into training (70%) and validation (30%) sets. Death-censored graft survival was the primary performance indicator. We experimented with four machine learning (ML) models assessed for calibration and discrimination [integrated Brier score (IBS) and Harrell's concordance index]. We assessed the potential clinical utility using decision curve analysis.

Results: XGBoost demonstrated the best discriminative performance for survival (area under the curve = 0.73, 0.74, and 0.75 at 3, 7, and 10 years post-transplant, respectively). The concordance index was 0.72. The calibration process was adequate, as evidenced by the IBS score of 0.09.

Conclusion: By evaluating possible donor-recipient pairs based on graft survival, the AI-based UK Live-Donor Kidney Transplant Outcome Prediction has the potential to enhance choices for the best live-donor selection. This methodology may improve the outcomes of kidney paired exchange schemes. In general terms we show how the new AI and ML tools can have a role in developing effective and equitable healthcare.

Keywords: Transplant outcomes; deceased kidney donor; machine learning; personalized medicine.

Figure 1.

Hierarchy of cohort selection.

Figure 2.

Actual survival probabilities were calculated using Kaplan–Meier survival product estimator; then, the average of these were calculated at each time point. The average actual survival probabilities were plotted against the equivalent predicted probabilities at different time points.

Figure 3.

DCA showing the net benefit of the UK-LTOP predictions versus ‘treat all (assuming all will lose their graft and need intervention)’ and versus ‘treat none (assuming none will lose their graft and no intervention needed)’.