Development, application, and validation of a histological classification automation system for kidney allograft precision diagnostics

Over the last three decades, the Banff International histological classification has been the gold standard for allograft rejection diagnostics. However, in the era of precision medicine, the classification has become complex over time, leading to misclassifications with therapeutic consequences. We aimed to develop an integrated automated rejection classification system and demonstrate its clinical impact by validating its ability to improve rejection diagnoses and risk stratification of allograft outcomes.

A systematic review of all Banff classification rules until the latest version published in 2019 was performed. We translated all these rules into a coded algorithm covering every possible scenario (n > 450 billion). The algorithm was converted into decision trees and an application that assigns diagnoses. We then tested the Banff Automation System’s ability to reclassify rejection diagnoses compared with pathologists’ assessments from: 1) two prospective observational cohorts of adult kidney transplant recipients (Paris Transplant Group and University of Wisconsin-Madison); 2) one multicentric international cohort of pediatric kidney transplant recipients, and 3) two large clinical trials (KTD-Innov and EU-TRAIN), totaling 20 transplant referral centers in Europe and North America. ClinicalTrials.gov, NCT05306795.

We decoded and encoded all published Banff rules and related scenarios and translated them into an algorithm that fully automates the rejection diagnosing process. We devised a system that generates automatic diagnostic reports with biopsy-related diagnoses. We then tested the Banff Automation System to reclassify rejection-related diagnoses on 4,409 kidney transplant biopsies from 3,054 patients (2,076 biopsies in the Paris Transplant Group cohort, 555 in the University of Wisconsin-Madison’s cohort, 514 in the international pediatric cohort, 873 in the KTD-Innov trial, and 391 in the EU-TRAIN trial). In the adult kidney transplant biopsies, the Banff Automation System reclassified 83 of 279 (29.75%) antibody-mediated rejection (AMR) cases and 57 of 105 (54.29%) T-cell mediated rejection (TCMR) cases, while 237 of 3,239 (7.32%) biopsies with non-rejection-related diagnoses were reclassified as rejection. The rejection reclassification rates were 8 of 26 (30.77%) and 12 of 39 (30.77%) for AMR and TCMR, respectively, in the pediatric cohort. Finally, we found that the non-rejection diagnoses according to the pathologists which were reclassified as rejections by the Banff Automation System had similar outcomes as those of confirmed rejection cases.

We built the first comprehensive and fully integrated histological Banff Automation System and confirmed its ability to reclassify and stratify the risk of kidney transplant rejection. This decision support system might improve reproducibility, thereby reducing misclassifications; standardizing histology-based endpoints in clinical trials and posttransplant diagnostics; and optimizing therapeutic management.