Kidney Week

Abstract: PUB072

Tervaert-Guided Integration of Histology and Clinical Data for Diabetic Kidney Disease Distinction

Session Information

• Publication Only

Category: Diabetic Kidney Disease

• 701 Diabetic Kidney Disease: Basic

Authors

• Afsari, Fatemeh, University of Florida, Gainesville, Florida, United States

Fatemeh Afsari, PhD

• Rosenberg, Avi Z., Johns Hopkins University, Baltimore, Maryland, United States

Avi Z. Rosenberg, MD, PhD

• Eadon, Michael T., Indiana University, Bloomington, Indiana, United States

Michael T. Eadon, MD, FASN

• Clapp, William L., University of Florida, Gainesville, Florida, United States

William L. Clapp, MD

• Han, Seung Seok, Seoul National University, Gwanak-gu, Seoul, Korea (the Republic of)

Seung Seok Han, MD, PhD

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

Pinaki Sarder, PhD, FASN

Background

Diabetic nephropathy (DN) and focal segmental glomerulosclerosis (FSGS) cause proteinuria but require different treatments. DN is managed by glycemic and blood pressure control; FSGS may need immunosuppression or steroids. Biopsy distinction is tough. We developed a diagnostic index using PAS-stained WSIs and clinical data to distinguish pure DN from DN with overlapping podocytopathy.

Methods

We analyzed 88 PAS-stained WSIs (DN, n=78; DN+podocytopathy, n=10), using 7.8:1 oversampling to address class imbalance. High-res images were segmented to extract glomerular morphometrics (PAS-positive matrix, lumina, nuclei), and clinical data (proteinuria, eGFR, HbA1c). Expert-labeled glomerular patches trained a vision transformer to assign Tervaert stages. DN scores, image features, and clinical data informed an SVM/XGBoost index targeting nodularity vs mesangiopathy and proteinuria vs podocyte injury.

Results

XGBoost achieved 95% accuracy with high precision, recall, and F1-score (Fig. 1), supporting integration of image features, clinical data, and DN scores. ROC curves were plotted (Fig. 2). SVM showed an AUC of 0.77 (moderate); XGBoost reached 0.93, indicating strong separability.

Conclusion

This transformer-based index distinguished pure DN from DN with podocytopathy via XGBoost, highlighting deep learning–driven histopathological biomarkers. Multi-center validation may enhance DKD care.

Fig 1. Performance metrics for SVM and XGBoost classifiers

Fig 2. ROC curves and AUC values for SVM and XGBoost classifiers

Funding

• Other NIH Support

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025x6w9e2fd