Optimizing Preimplantation Kidney Transplant Biopsy Interpretation with Artificial Intelligence Assistance - Resubmission - 1

PROJECT SUMMARY Less-than-ideal kidneys recovered from deceased donors continue to be discarded at a rate of 40-50%, but evidence shows that many discarded kidneys could have yielded good outcomes and reduced the needlessly high mortality on the transplant waitlist. Ensuring appropriate utilization and preventing discards of usable

kidneys requires expert pathologist interpretations of preimplantation biopsies, as such expert biopsy interpretations strongly correlate with post-transplant outcomes above and beyond clinical characteristics. However, organ offers often occur in the middle of the night, and many transplant centers lack 24/7 availability

of an experienced kidney transplant pathologist, so expert interpretations are often unavailable in real time. Instead, clinical decision-making relies on the unreliable and inconsistent interpretations of on-call general pathologists. Preimplantation biopsy interpretation from non-experts carries a high risk of inappropriate discard;

in fact, biopsy findings are cited as justification for nearly 40% of discards. We hypothesize that applying modern artificial intelligence (AI) techniques, including a novel self-supervised deep learning framework that we have recently developed called Histomorphological Phenotype Learning to identify histopathologic clusters,

could ensure universal access to reliable preimplantation biopsy interpretation and reduce discards. This is feasible as digital imaging is becoming the standard amongst organ procurement organizations (OPOs), and AI-assisted histopathological interpretation has proven superior in other clinical scenarios.

Leveraging an array of >10,000 biopsy images shared from 4 OPOs, and externally validated with images from another 8 OPOs as well as an international cohort, we will compare self-supervised and expert-supervised “AI- pathology assisted” (AIPA) biopsy interpretations to those of an on-call pathologist (standard of care). We will

also compare AIPA to existing kidney biopsy scoring systems. Further, we will study stakeholder attitudes and catalog facilitators and barriers to implementation of AI-assisted preimplantation biopsy interpretation, to encourage rapid clinical adaptation to improved biopsy interpretations. To improve the clinical utility of

preimplantation biopsies and reduce inappropriate organ discard, we propose the development AIPA with the following aims: (1) to use self-supervised and expert-supervised learning to construct a comprehensive unbiased atlas of histopathology in preimplantation deceased donor kidney biopsies, (2) to develop and

externally validate AIPA-derived models and compare interpretations and association with clinically relevant post-transplant outcomes, and (3) to develop consensus for the implementation of AIPA in clinical practice. Our findings will be immediately clinically useful to kidney transplant providers and patients across the US in

evaluating the 50,000 kidneys recovered for potential transplantation per year. More reliable pre-transplant biopsy interpretations will reduce discards, facilitate more appropriate utilization, and improve the clinical utility of preimplantation biopsy for the increasing numbers of less-than-ideal kidneys being recovered.