Counterfactual inference graph network for disease prediction

Abstract

Graph convolutional networks are widely used as computational models that integrate the data of image and non-image modalities in the medical diagnostic domain, especially while exploring population-based disease prediction. However, existing methods use associative combinations of non-image information while constructing population graphs, which are often interfered while capturing the relation between subjects, and the causal relationship between these variables is ignored. In this study, we present a counterfactual inference graph network for improving disease prediction accuracy and generalization by constructing a learnable population relational graph structure. Specifically, nodes can be regarded as contexts, and edge (non-)existence and node classification can be regarded as treatments and outcomes, respectively. We enhance graph learning by predicting both, the observed factual classifications and counterfactual classifications. To provide an estimate of the prediction uncertainty associated with relational graphs, a novel concept called edge uncertainty was proposed. Through extensive experiments on six publicly available datasets on four types of diseases: autism spectrum disorder (ASD), Alzheimer’s disease (AD), thoracic diseases, and ocular diseases, we demonstrate that our proposed method can significantly and consistently improve the prediction performance compared with baseline approaches.