Live-Donor Kidney Transplant Outcome Prediction (L-TOP) using artificial intelligence

Area of Science:

• Nephrology
• Transplantation
• Artificial Intelligence
• Medical Informatics

Background:

• Accurate prediction of live-donor kidney transplant outcomes is crucial for clinical decisions and donor selection.
• Current prediction models lack sufficient discriminative and calibration power, necessitating improved risk stratification tools.

Purpose of the Study:

• To evaluate the efficacy of various artificial intelligence (AI) algorithms in enhancing the risk stratification index for live-donor kidney transplantation.
• To compare the performance of AI models against existing prediction tools like the Living Kidney Donor Profile Index (LKDPI).

Main Methods:

• Analysis of pre-transplant variables from 66,914 live-donor kidney transplants using data from the United Network of Organ Sharing database.
• Randomized data into training (80%) and test (20%) sets, focusing on death-censored graft survival as the primary outcome.
• Evaluation of four machine learning models using discrimination metrics (CTD, AUC) and calibration (IBS), with decision-curve analysis for clinical utility assessment.

Main Results:

• The deep Cox mixture model demonstrated superior discriminative performance (AUC 0.70 at 5 years) and good calibration (IBS 0.09).
• This AI model achieved a time-dependent concordance index (CTD) of 0.70 at 5 years, significantly outperforming the LKDPI (CTD 0.56).
• Decision-curve analysis indicated a net clinical benefit for the AI model compared to LKDPI-based strategies.

Conclusions:

• The AI-based deep Cox mixture model, Live-Donor Kidney Transplant Outcome Prediction, surpasses current models in predicting graft survival.
• This advanced model offers potential for optimizing live-donor selection and improving decisions in kidney transplantation.
• The model could be instrumental in enhancing outcomes within paired exchange programs.