Improving assessment in kidney transplantation by multitask general path model

Abstract

Background

Kidney transplantation is a pivotal intervention for individuals suffering from end-stage renal diseases, offering them the potential for restored health and an enhanced quality of life. However, the successful outcome of these transplantation procedures relies significantly on the careful matching of donor kidneys with compatible recipients. Unfortunately, the current kidney-matching process overlooks viability changes during preservation. The objective of this study is to investigate the potential for forecasting heterogeneous kidney viability using historical datasets to enhance kidney-matching decision-making.

Methods

We present a multitask general path model designed for continuous forecasting of kidney viability during preservation. This model quantifies likely viability trajectories of donor kidneys based on pathologist-provided biopsy scores during preservation, explicitly addressing both inter-kidney similarities and individual differences. To validate our model, we conducted viability assessments on six recently procured porcine kidneys and needle biopsy insertion experiments on phantoms, utilizing a leave-one-kidney-out cross-validation approach.

Results

Our proposed model consistently exhibited the lowest forecasting error (averaged root mean squared error, RMSEbegin=0.61 at the beginning and RMSEend<0.05 at the end of kidney preservation) when compared to widely-adopted benchmark models, including multitask learning (RMSEbegin=0.65, RMSEend=0.54), general path (RMSEbegin=0.58, RMSEend=0.49), and generalized linear models (RMSEbegin=0.59, RMSEend=0.56) in the kidney viability assessment study. Additionally, across all testing scenarios, the forecasting RMSE of our model rapidly diminished with minimal initial kidney samples during preservation. Similar patterns were observed from the needle biopsy insertion study.

Conclusions

In both validation studies, our model outperformed benchmark models and exhibited rapid learning with limited initial samples. This approach holds promise for enhancing kidney transplantation decision-making, including improving tissue extraction accuracy through needle biopsy data analysis. By implementing this model across various kidney assessment stages in transplantation, we aim to reduce kidney discards and benefit a larger number of patients.