Abstract: SA-PO0016
Human-Artificial Intelligence (AI) Collaborative Assessment of Procurement Biopsies from Deceased Donors to Predict Post-Transplant Kidney Function: The ComPRePS AI Tool
Session Information
- Intelligent Imaging and Omics: Phenotyping and Risk Stratification
November 08, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Pardinhas, Clara, Universidade de Coimbra, Coimbra, Coimbra District, Portugal
- Paul, Anindya S., University of Florida, Gainesville, Florida, United States
- Rodrigues, Luis, Universidade de Coimbra, Coimbra, Coimbra District, Portugal
- Rubin, Jeremy, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Abdelazim, Haitham M, University of Florida, Gainesville, Florida, United States
- Gupta, Akshita, University of Florida, Gainesville, Florida, United States
- Border, Samuel, University of Florida, Gainesville, Florida, United States
- Sousa, Vitor Manuel Leitão de, Universidade de Coimbra, Coimbra, Coimbra District, Portugal
- Figueiredo, Arnaldo, Universidade de Coimbra, Coimbra, Coimbra District, Portugal
- Zee, Jarcy, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Sarder, Pinaki, University of Florida, Gainesville, Florida, United States
Background
Predicting post-transplant kidney function using procurement biopsies remains controversial due to limitations of current tools like the Kidney Donor Profile Index (KDPI) and histological scores (e.g., Banff, Remuzzi). These offer risk stratification but lack precision in forecasting post-transplant eGFR. We developed the Computational Renal Pathology Suite (ComPRePS: https://compreps.wiki/home), a human–AI cloud platform, to extract pathomics from kidney functional tissue units (FTUs) and applied machine learning (ML) to predict 12-month post-transplant eGFR.
Methods
We analyzed 139 procurement biopsies from 110 deceased donors intended for single-kidney transplantation. Samples were FFPE-processed, PAS-stained, digitized at 40x, and clinically annotated. ComPRePS segmented FTUs—glomeruli, tubules, arteries, and arterioles—with pathologist quality control. From these, 306 granular pathomic features were extracted, quantifying morphology, color, texture, and spatial patterns in eosinophilic, luminal, and nuclear subregions. Vascular fibrointimal sclerosis and IFTA were also segmented. Clinical and image data were integrated into lasso, ridge, elastic net, and random forest (RF) models using Minimum Redundancy Maximum Relevance (MRMR) feature selection. Models were internally validated via bootstrapping; performance was assessed by mean squared error (MSE).
Results
The RF model using 12 MRMR-selected features halved the MSE compared to KDPI-based regression. Key predictors included donor age, last eGFR, vascular fibrointimal thickening, glomerulosclerosis, and nuclear morphology. RF-predicted CKD stages aligned closely with observed outcomes, unlike KDPI which overclassified Stage 3a CKD.
Conclusion
Our human–AI approach to procurement biopsy analysis significantly improves post-transplant function prediction over existing tools. This integrative method may support better donor evaluation and organ allocation.