Kidney Week

Abstract: SA-PO0015

Predicting Delayed Graft Function Using Multimodal Artificial Intelligence (AI) with Pathomics from Deceased Donor Biopsy and Clinical Data

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

• Rubin, Jeremy, University of Pennsylvania, Philadelphia, Pennsylvania, United States

Jeremy Rubin, MS

• Paul, Anindya S., University of Florida, Gainesville, Florida, United States

Anindya S. Paul, PhD

• Rodrigues, Luis, Universidade de Coimbra, Coimbra, Coimbra District, Portugal

Luis Rodrigues, MD

• Abdelazim, Haitham M, University of Florida, Gainesville, Florida, United States

Haitham M Abdelazim

• Gupta, Akshita, University of Florida, Gainesville, Florida, United States

Akshita Gupta, MS

• Border, Samuel, University of Florida, Gainesville, Florida, United States

Samuel Border, MS

• Pardinhas, Clara, Universidade de Coimbra, Coimbra, Coimbra District, Portugal

Clara Pardinhas

• Sousa, Vitor Manuel Leitão de, Universidade de Coimbra, Coimbra, Coimbra District, Portugal

Vitor Manuel Leitão de Sousa, PhD

• Figueiredo, Arnaldo, Universidade de Coimbra, Coimbra, Coimbra District, Portugal

Arnaldo Figueiredo, PhD

• Zee, Jarcy, University of Pennsylvania, Philadelphia, Pennsylvania, United States

Jarcy Zee, PhD

• Sarder, Pinaki, University of Florida, Gainesville, Florida, United States

Pinaki Sarder, PhD, FASN

Background

Delayed graft function (DGF) is an early indicator of kidney transplant injury and is associated with worse long-term outcomes. While ischemic injury and other donor-related pathology contribute, existing clinical tools lack precision in predicting DGF. We aimed to develop a predictive model combining donor clinical data with pathomic features extracted from procurement biopsies to improve early DGF risk stratification.

Methods

We analyzed procurement biopsies from 110 deceased donors and 139 kidney transplant recipients, with DGF occurring in 16.5%. Biopsies were FFPE-processed, PAS-stained, and digitized at 40x. Donors were 64.5% male, 100% white, with a mean age of 62.3 years; 68.2% died from stroke. KDPI averaged 69.5 (39.5% >85%). Using ComPRePS [https://compreps.wiki/home], glomeruli, tubules, arteries, arterioles, and key lesions (e.g., IFTA, vascular fibrointimal sclerosis) were segmented with expert validation. A total of 306 pathomic features were extracted, quantifying structure, size, and texture across eosinophilic, luminal, and nuclear subregions. We trained LASSO, ridge, elastic net, and random forest (RF) models using features selected via Minimum Redundancy Maximum Relevance (MRMR), and compared them to Kidney Donor Profile Index (KDPI) alone. Bootstrapping and five-fold cross-validation were used. Performance was evaluated using the concordance statistic (C-statistic).

Results

The top-performing model, RF with 26 MRMR-selected features, reached a C-statistic of 0.94 versus 0.5 for KDPI (left figure). ROC analysis on a hold-out set yielded an AUC of 0.81 (right figure). Key predictors included vascular injury markers (e.g., nuclear-to-object ratio, arteriolar distance transform, luminal homogeneity), glomerulosclerosis, and donor clinical factors.

Conclusion

Combining donor clinical and pathomic data enhances DGF prediction accuracy, outperforming KDPI alone. This approach may support better early risk stratification in kidney transplantation.

Funding

• NIDDK Support

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025aahppxk1