Stat Colloquium: Dr. Jeremy Rubin

Title: Statistical Methods for Variable Selection and Prediction with Pathomic Features

Abstract: The current gold standard for kidney disease diagnoses is manual visual assessment by pathologists of kidney biopsy tissue. However, these manual assessments are time-consuming, subjective, and lack reproducibility. The generation of many quantitative features from digital whole slide images (WSI) of biopsy tissue, an emerging field known as pathomics, may allow for the identification of novel, objective and comprehensive biomarkers of kidney disease as well as better prediction of kidney function outcomes. For each subject’s WSI, the same pathomic features are computed for each histologic object that has been identified and segmented by a deep learning model. We develop novel regression approaches to predict continuous outcomes of kidney function from these unbalanced, matrix-valued pathomic features and identify which features are most informative of the outcome. We illustrate that our approaches can identify the informative pathomic features in simulation studies and provide improved predictive accuracy of kidney function outcomes in real data analyses of image features from patients with glomerular disease. Additionally, we use the conformal prediction statistical framework to construct individual prediction intervals for continuous kidney function outcomes from the pathomic features. Through simulations and analysis of glomerular disease image feature data, we highlight that these prediction intervals reliably cover the true continuous kidney function outcomes. These statistical methods provide a suite of tools for discovering new pathomic biomarkers of kidney disease as well as predicting and quantifying the uncertainty of continuous kidney function outcomes.