Abstract: SA-PO0014
Automated Prediction of Drug Treatment Response in Lupus Nephritis Using Machine Learning
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
- Kasireddy, Harishwar Reddy, University of Florida, Gainesville, Florida, United States
- Demeke, Dawit S., University of Michigan Michigan Medicine, Ann Arbor, Michigan, United States
- La Rosa, Patricio Salvatore, Bayer Company, St. Louis, Missouri, United States
- Dai, Jessica, New York University, New York, New York, United States
- Nast, Cynthia C., Cedars-Sinai, Los Angeles, California, United States
- Rosenberg, Avi Z., Johns Hopkins University, Baltimore, Maryland, United States
- Fava, Andrea, Johns Hopkins University, Baltimore, Maryland, United States
- Petri, Michelle, Johns Hopkins University, Baltimore, Maryland, United States
- Mariani, Laura H., University of Michigan Michigan Medicine, Ann Arbor, Michigan, United States
- Buyon, Jill P., New York University, New York, New York, United States
- Sarder, Pinaki, University of Florida, Gainesville, Florida, United States
- Hodgin, Jeffrey B., University of Michigan Michigan Medicine, Ann Arbor, Michigan, United States
Group or Team Name
- Accelerating Medicines Partnership: RA/SLE Network (AMP RA/SLE).
Background
The prognostic utility of the current lupus nephritis (LN) classification is controversial, necessitating novel, reproducible methods. While LN patients are treated with a wide range of immunosuppressants, only a subset achieve a complete treatment response. To improve kidney outcomes, it is essential to develop more personalized treatment strategies through early identification of patients who will not respond to current treatments.
Methods
We analyzed periodic acid-Schiff stained whole slide images (WSIs) from kidney biopsies of 130 LN patients in the Accelerated Medicines Partnership, comparing 64 non-responders to 66 complete and partial responders after 52 weeks of immunosuppressive treatment. Non-globally sclerotic glomeruli (NGSG) were segmented and 72 pathomic features extracted per instance. Two modeling strategies were evaluated using 10-fold cross-validation for training: (1) XGBoost to average pathomic features to create mean vectors, and (2) Gated Attention-Based Multiple-Instance Learning (GAMIL) on instance-level features.
Results
GAMIL outperformed XGBoost with a higher F1 score (0.74 ± 0.06 vs. 0.63 ± 0.13), underscoring the benefit of preserving instance-level feature information for accurately identifying patients unlikely to respond to treatment at 52 weeks.
Conclusion
Machine learning models using kidney biopsy pathomic feature extraction enable pathologists to identify LN patients who are less likely to respond to standard immunosuppression, supporting personalized monitoring strategies to maximize the probability of complete response. Ultimately, integrating this approach into clinical decision-making could improve long-term kidney outcomes and help prevent progression to kidney failure.
We use a Detectron2 DeepLabv3+ CNN to segment non-globally sclerotic glomeruli (NGSG) and extract features. These pathomic descriptors train machine learning models (XGBoost or GAMIL) to predict therapeutic response in LN patients. This interpretable pipeline enables automated drug-response prediction directly from histopathology WSIs.
Funding
- NIDDK Support