Validating RenalAgent: An Artificially Intelligent Autonomous Agent for Renal Pathology

Background and aims: Histopathological assessment of renal allografts using the Banff classification remains the diagnostic cornerstone for transplant dysfunction, yet inter-observer variability, global pathologist shortages, and inconsistent application undermine diagnostic standardisation worldwide. Existing computational pathology tools address isolated tasks but require programming expertise, specialised hardware, and typically fail to generalise beyond their training institution. We aimed to develop and internationally validate the first fully autonomous AI agent for end-to-end renal allograft pathology, longitudinal graft surveillance, and prognostication.
Methods: RenalAgent orchestrates vision-language, panoptic segmentation, and classification models through a central conductor language model. The system autonomously processes H&E, PAS, MT, Jones silver, C4d immunohistochemistry, and BK-stained whole-slide images to generate 16 Banff lesion scores, rejection subtype classifications, continuous histomorphometric measures, and structured diagnostic reports. Validation was conducted across five international cohorts: Mass General Brigham (MGB; n=15,000 biopsies), the Australian Chronic Allograft Dysfunction cohort (AUSCAD; n=2,300), the Genomics of Chronic Allograft Rejection cohort (GOCAR; n=619), and sporadic clinical cases from European and Asian centres. A prospective reader study evaluated diagnostic accuracy and report quality alongside expert transplant pathologists in real-time clinical workflows.

Results: Across 15,000 MGB biopsies, RenalAgent achieved high concordance with expert pathologists (mean quadratic-weighted kappa=0.73 across 16 Banff scores) and near-perfect rejection subtype classification (ABMR AUC=0.96; TCMR AUC=0.97; IFTA AUC=0.98). Performance was maintained without retraining in AUSCAD and across European and Asian validation cases. In the GOCAR cohort, longitudinal evaluation demonstrated RenalAgent's capacity to track rejection trajectory and chronic injury progression across serial biopsies. Integration of C4d and BK immunohistochemistry expanded assessment to antibody-mediated rejection activity and viral nephropathy surveillance. The prospective reader study confirmed strong concordance between RenalAgent-generated reports and expert pathologist assessments.

Conclusions: RenalAgent is the first autonomous AI system validated for end-to-end renal allograft pathology across five international centres spanning four continents. By eliminating technical barriers and demonstrating prognostic utility through longitudinal graft surveillance, RenalAgent represents a paradigm shift toward autonomous diagnostic agents in transplantation, with direct implications for global diagnostic standardisation, equitable access to specialist pathology, and personalised post-transplant management.