Predicting Kidney Post-Transplantation Function from Optical Coherence Tomography Images Using Machine Learning Approaches

Abstract

Delayed graft function (DGF) is a major post-transplant complication in kidneyrecipients, particularly from deceased donors. This study investigates the effectiveness of machine learning approaches in predicting DGF using texture features extracted from optical coherence tomography (OCT) images, supplemented with Kidney Donor Profile Index (KDPI) scores. To address significant class imbalance between immediate graft function (IGF) and DGF, three strategies were evaluated: threshold optimization using the GHOST algorithm, data balancing via SMOTE-Tomek, and cost-sensitive learning (CSL). We demonstrate that classifiers trained on KDPI scores alone under- performed compared to those trained OCT images derived texture features. Classifiers trained on KDPI alone showed moderate improvements with GHOST, but incorporat- ing OCT-derived texture features significantly enhanced model performance across all classifiers. These findings underscore the utility of OCT imaging in assessing kidney allograft quality and predicting post-transplant outcomes, highlighting the potential of machine learning classifiers to estimate the risk of delayed graft function (DGF) in deceased donor kidneys prior to transplantation.