S. Joshua Swamidass, MD, PhD
Associate Professor of Pathology & Immunology
Washington University School of Medicine in St. Louis
Modeling Kidney Transplant Outcome From the Donor Biopsy Using a Deep Learning Derived Measure of Percent Global Glomerulosclerosis
Background: Donor kidney biopsies are commonly performed prior to transplantation to determine organ quality and risk of graft failure. The ability of the biopsy to predict graft outcomes has shown mixed results in prior studies. Current standard of care depends on manual visual assessment of biopsy frozen sections by on-call pathologists who do not have subspecialty training in kidney pathology. The commonly reported histologic chronic damage finding, %global glomerulosclerosis, shows suboptimal reproducibility, accuracy, and precision. It is our hypothesis that variability in quantitation of %global glomerulosclerosis is a major reason why biopsy results have shown variable correlations with graft outcomes. We developed an automated deep learning algorithm to rapidly identify and classify glomeruli in donor kidney biopsy frozen section whole-slide images with better performance than on-call pathologists. Pathologists can utilize the algorithm through a HIPAA-compliant, cloud-based web application called Trusted Kidney. Trusted Kidney provides a dynamic, easy-to-use interface. The purpose of this study is to evaluate rapid automated analysis of donor kidney biopsies to elucidate relationships between automated quantitation of %global glomerulosclerosis and clinical outcomes.
Methods: Kidney biopsy whole slide images obtained from deceased donors between April 2015 and January 2021 from the Washington University WUPAX digital database were used in this study. Kidneys that were transplanted into recipients >18 yrs of age, were without missing metadata, had 25 or more glomeruli, and >1 year follow-up data were included. Patients with dual organ transplants, history of previous transplant, and ABO incompatible transplants were excluded for a total of 1437 slides from 724 kidneys (# male=440, # female=284). These were uploaded to the Trusted Kidney website for automated detection and classification of glomeruli for computation of % global glomerulosclerosis. Glomeruli counts for individual kidneys were pooled from multiple sections if more than one section was present. Recipient outcome data including delayed graft function, graft survival, creatinine, glomerular filtration rate, mortality, and hospitalization were extracted from the UNOS database.
Results: The distribution of % global glomerulosclerosis for the cases in this study was approximately log-normal, so log transformation was used to meet the normality assumption of statistical tests. Mean log(% global glomerulosclerosis) correlated with graft survival (p=0.0372) and kidney function (glomerular filtration rate) (p=0.0135). There was no significant correlation between %global glomerulosclerosis and delayed graft function, creatinine, mortality, or hospitalization.
Conclusions: This is the first large scale study evaluating a web-based deep learning application to correlate %global glomerulosclerosis in donor kidney biopsies with transplant recipient clinical outcomes. Mean intraoperative %global glomerulosclerosis was found to significantly correlate with graft survival and kidney function on univariate analysis. Application of this automated technique is anticipated to enable analysis of sample sets much larger than is practicable by human evaluation alone, and to provide new opportunities for organ outcome prediction from biopsy evaluation.
S. Joshua Swamidass MD PhD is an associate professor of Laboratory and Genomic Medicine at Washington University in St. Louis. His group uses machine learning to solve problems at the intersection of medicine, biology, and chemistry.