Abstract: TH-PO040
Machine Learning-Based Approach to Classifying Risk of Progression of Membranous Nephropathy to ESKD Using Electron Microscopy
Session Information
- AI, Digital Health, Data Science - I
November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Fermin, Jamie Ledesma, University of Florida Department of Electrical & Computer Engineering, Gainesville, Florida, United States
- Border, Samuel, University of Florida Department of Biomedical Engineering, Gainesville, Florida, United States
- Paul, Anindya S., University of Florida Department of Medicine - Quantitative Health Section, Gainesville, Florida, United States
- Heymann, Jurgen, National Institutes of Health Clinical Research Center - NIDDK Kidney Disease Section, Bethesda, Maryland, United States
- Rosenberg, Avi Z., Johns Hopkins School of Medicine Department of Pathology, Baltimore, Maryland, United States
- Waldman, Meryl, National Institutes of Health Clinical Research Center - NIDDK Kidney Disease Section, Bethesda, Maryland, United States
- Sarder, Pinaki, University of Florida Department of Medicine - Quantitative Health Section, Gainesville, Florida, United States
Background
Membranous nephropathy (MN) is an immune complex-mediated process that can result in nephrotic syndrome. Machine learning applications in medicine have emerged as powerful tools that can analyze large volumes of medical data (i.e., imaging, patient records) and extract patterns and relationships with diagnostic and prognostic insights. This work aims to use clinical and classical morphological image features to determine the risk of MN progression to end-stage kidney disease (ESKD).
Methods
Eighty MN cases from multiple institutions, each with available electron micrographs (EMs) of glomerular basement membrane (GBM) and corresponding baseline clinical metrics (albumin, eGFR, creatinine) were collated. An attention U-Net model was trained to perform multiclass segmentation of the GBM, electron-dense deposits, and background. Fifteen morphological features (i.e., total number of deposits, aspect ratio, circularity) were extracted from the predicted mask. Then, an artificial neural network (ANN) was trained to classify cases according to 3 risk tiers: low, intermediate, and high progression to ESKD. The same features were used to train an ANN to predict remission (no remission or complete remission).
Results
The segmentation model achieved a Dice score coefficient of 78%. Moreover, the ANN obtained classification accuracies of 73.3%, 67%, and 73.3% on the test set using clinical metrics alone, morphological features alone, and both clinical and morphological features, respectively. Another ANN trained to predict remission achieved 40% accuracy with both clinical and morphological features.
Conclusion
Clinical and morphological features can predict risk levels with reasonable accuracy. However, the same features are poor predictors of remission. Morphological features combined with clinical features do not improve risk classification accuracy. More meaningful and discriminative morphological features are needed to enhance predictive ability.
Funding
- NIDDK Support