An optimised ensemble for antibody-mediated rejection status prediction in kidney transplant patients

Antibody-mediated rejection (AMR) is one of the primary mechanisms of graft loss following organ transplantation. A key difficulty with AMR diagnosis is that symptoms typically manifest when the graft is already damaged beyond repair. Diagnosis is also complicated by differing interpretations of histological data by pathologists, highlighting the urgent need for more quantitative approaches. In this paper we propose an ensemble classifier approach to predicting AMR status from gene expression data. We employ two random oversampling techniques - Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Oversampling (ADASYN) - to address the class imbalance in the original data set, and use particle swarm optimisation (PSO) for the selection of the ensemble hyperparameters. Our results demonstrate that applying the PSO-optimised ensemble to the balanced data set provides better predictive performance than the ensemble alone, and represents an important step towards more accurate sub-clinical prediction of AMR status and improved patient risk stratification.