FC 107: Development and Validation of a Machine Learning-Based Virtual Biopsy System in Kidney Transplant Patients

BACKGROUND AND AIMS Graphical Abstract In kidney transplantation, day-zero biopsies are used to assess organ quality and discriminate lesions inherited from the donor or acquired after transplantation. However, many centres worldwide do not perform those biopsies which remain invasive, costly and may delay the transplant procedure. We aimed to develop and validate a noninvasive virtual biopsy system. METHOD A total of 17 centres were included from Europe, North America and Australia from 2000 to 2019. Candidate predictors were assessed following a prespecified protocol. Outcome measures were the day-zero biopsy lesions (Banff classification) including CV, AH, IFTA scores and % of sclerotic glomeruli. Six machine learning models were developed and their performances were assessed. RESULTS A total of 12 992 day-zero biopsies were included. Eleven parameters were used to build the classifiers, including donor age, kidney function, hypertension, BMI, proteinuria, diabetes, sex, donor type, cause of death and Hep-C status. The ensemble models (random forests, neural networks, gradient boosting, extreme gradient boosting tree, linear discriminant analysis, and naive Bayes) showed multi-AUC of 0.738, 0.817 and 0.788 for prediction of CV, AH and IFTA scores, and a good performance for predicting glomerulosclerosis (mean absolute error, MAE = 4.766). We confirmed the robustness and generalizability in multiple clinical scenarios and subpopulations and built an online interface for clinicians: https://transplant-pred/Virtual_Biopsy. CONCLUSION We developed and validated the first virtual biopsy system that enables the prediction of day-zero biopsy, based on routinely collected parameters. This can assist clinicians in assessing allograft quality, discrimination of donor derived versus acquired lesions after transplantation and prevent overdiagnosis of calcineurin inhibitor (CNI) toxicity.