Kidney Week

Abstract: SA-PO0051

Interpretable Multiple Instance Learning (MIL) of Glomerular Pathomics for Kidney Function Prediction: SHAP (SHapley Additive exPlanations) Analysis

Session Information

• Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
  October 24, 2026 | Location: Exhibit Hall A, 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

• Deng, Ruining, Weill Cornell Medicine, New York, New York, United States

Ruining Deng, PhD

• Shi, Tianyu, Sichuan University, Chengdu, Sichuan, China

Tianyu Shi

• Yang, Yuechen, Vanderbilt University, Nashville, Tennessee, United States

Yuechen Yang

• Salvatore, Steven, Weill Cornell Medicine, New York, New York, United States

Steven Salvatore, MD

• Seshan, Surya V., Weill Cornell Medicine, New York, New York, United States

Surya V. Seshan, MBBS

• Sabuncu, Mert, Weill Cornell Medicine, New York, New York, United States

Mert Sabuncu, PhD

• Yang, Yihe, Weill Cornell Medicine, New York, New York, United States

Yihe Yang, MD, MS

Background

MIL-based glomerular pathomics correlates more strongly with eGFR than human assessment by preserving morphologic heterogeneity across individual glomeruli. To improve interpretability, SHAP analysis was applied to identify the morphologic features driving model predictions.

Methods

162 whole-slide images of non-neoplastic kidney parenchyma (an individual patient each) were included (Fig.1). Glomeruli and glomerular lesions were segmented with a fine-tuned YOLOv12 model and classified into five classes:viable, GGS, SGS, ischemic, and atubular. A 314-dimensional glomerular pathomics feature vector including fine- and coarse-texture descriptors was extracted(PySpatial). 3 MIL models were trained to predict pre-op eGFR(class-ratio, pathomics, and combined). Model performance was evaluated by 5-fold cross-validation(R²,RMSE,MAE,Pearson r,and pooled R²).SHAP analysis identified the contribution of each features and feature families.

Results

SHAP analysis showed that a relatively small subset of pathomics features disproportionately contributed to eGFR prediction. The most influential features were related to glomerular texture heterogeneity, intensity distribution, and structural morphology, with texture descriptors contributing more strongly than size-based measurements. At the feature-family level, texture and intensity-distribution families dominated attribution rankings, while raw size/shape descriptors and lesion-class proportions contributed less. Instance-level attribution demonstrated that predictions relied on distributed morphologic patterns across multiple glomeruli rather than isolated lesion categories, suggesting that subtle architectural variation encodes functional information.

Conclusion

SHAP identified glomerular texture and morphologic heterogeneity as key drivers of eGFR prediction beyond conventional lesion assessment. Explainable computational pathology may improve discovery of quantitative kidney biomarkers.