Abstract: TH-OR019
Deep Learning Assessment of Chronic Injury in T0 Kidney Biopsies Predicts Donor Kidney Function and Long-Term Graft Survival
Session Information
- Artificial Intelligence in Kidney Care: From Pathology to Prediction
October 22, 2026 | Location: Room 711, Convention Center
Abstract Time: 04:30 PM - 04:40 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Gupta, Akshita, University of Florida Department of Health Outcomes & Biomedical Informatics, Gainesville, Florida, United States
- La Rosa, Patricio Salvatore, Bayer CropScience LLC, St. Louis, Missouri, United States
- Zee, Jarcy, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Rosenberg, Avi Z., Johns Hopkins University Department of Pathology, Baltimore, Maryland, United States
- Tomaszewski, John E., University at Buffalo Department of Pathology and Anatomical Sciences, Buffalo, New York, United States
- Clapp, William L., University of Florida Department of Pathology Immunology and Laboratory Medicine, Gainesville, Florida, United States
- Stegall, Mark D., Mayo Clinic Minnesota Department of Surgery, Rochester, Minnesota, United States
- Park, Walter, Mayo Clinic Minnesota Department of Surgery, Rochester, Minnesota, United States
- Rodrigues, Luis, Centro Hospitalar e Universitario de Coimbra EPE Campus dos Hospitais da Universidade de Coimbra, Coimbra, Coimbra District, Portugal
- Paul, Anindya S., University of Florida Department of Medicine, Gainesville, Florida, United States
- Naik, Abhijit S., University of Michigan Division of Nephrology, Ann Arbor, Michigan, United States
- Jen, Kuang-Yu, University of California Davis Department of Pathology Microbiology and Immunology, Davis, California, United States
- Sarder, Pinaki, University of Florida Department of Medicine, Gainesville, Florida, United States
Background
As the demand for donor kidneys exceeds organ availability, there is an increasing need for tools determining if donor biopsy findings truly predict post-transplant outcomes. Donor biopsy assessment at transplantation (T0) is subjective and weakly predictive of long-term outcomes. Computational pathology offers an opportunity to extract relevant features from routine histology. We developed a chronicity-gated attention-based multiple instance learning (ABMIL) framework that summarizes chronic injury patterns across T0 whole-slide images to predict donor terminal eGFR and further evaluated the association of this learned representation with long-term recipient outcome death-censored graft failure (DCGF).
Methods
T0 biopsies from two sites (n=667) were used for 5-fold cross-validation. Donor terminal eGFR was classified into 3 classes. We evaluated image-only, clinical-only (donor demographics and medical history), and combined image + clinical models. External validation was performed on an independent cohort (n=169). Long term transplant outcomes were evaluated using death-censored graft failure (DCGF) free survival analysis.
Results
On external validation, image-only achieved AUROC 0.81±0.01 for donor terminal eGFR classification, comparable to the image+clinical model (0.80±0.02). Given comparable performance, subsequent DCGF survival analyses were performed using the image-only model. Patients with high predicted eGFR had significantly lower DCGF hazard (HR=0.68, log-rank p=9.94 x 10-5).
Conclusion
Our deep learning model focused on chronic injury patterns in T0 biopsies can accurately classify donor kidney function and stratify long-term graft survival on external validation. This approach may provide an objective and interpretable tool for kidney transplant assessment and could be extended to evaluate additional injury patterns and transplant outcomes in future studies.
Funding
- NIDDK Support