Primary Submission Category: Mediation
Calculating Mediation Effects of High Dimensional Radiomic Data Between Exposure and Outcome
Authors: Emily Mastej, Debashis Ghosh,
Presenting Author: Emily Mastej*
Radiomics involves the mathematical extraction of quantitative features from medical images. While there is a wide range of radiomic-based prediction research, there is an issue with the clinical translation of these methods as well as their explainability due to the “black box” nature of deep learning algorithms. One solution to gain understanding of the mechanistic pathway between radiomic features and outcomes is to build causal inference models, specifically mediation models. We propose a downstream radiomics analysis method that uses high dimensional mediation to explore the causal pathway between an exposure and an outcome through a radiomic mediator. This method takes an exposure, radiomic features, and an outcome and finds principal directions of mediation (PDMs) or weighted groups of radiomic features that independently mediate the indirect effect of the exposure on the outcome with the largest indirect effect being mediated by the first PDM. We applied our method to T2 MRI radiomic data obtained from 203 subjects with either a glioblastoma or a glioma. Using IDH gene mutation status as the exposure and survival outcomes as the phenotype of interest, we used our method to find groups of radiomic features that were principal directions of mediation. Original tumor shape sphericity and wavelet HHL first order median were both found to be highly involved radiomic features in the first and second PDMs.