Causal Learning and Knowledge Fusion Mechanism for Brain Functional Network Classification

Abstract

Current studies have shown that the classification of human brain functional networks (BFN) is a reliable way to diagnose and predict brain diseases. However, a great challenge for current traditional machine learning methods and deep learning methods is their poor performance or lack of interpretability. To alleviate this problem, we propose a novel causal learning and knowledge fusion mechanism for brain functional network classification, named CLKF. The proposed mechanism first extracts causal relationships among brain regions from functional magnetic resonance imaging (fMRI) data using partial correlation and conditional mutual information, and obtains the relationships between BFN and labels by Gaussian kernel density estimation. Then, it fuses these two types of relationships as knowledge to aid in the classification of brain functional networks. The experimental results on the simulated resting-state fMRI dataset show that the proposed mechanism can effectively learn the causal relationships among brain regions. The results on the real resting-state fMRI dataset demonstrate that our mechanism can not only improve the classification performance of both traditional machine learning and deep learning methods but also provide an interpretation of the results obtained by deep learning methods. These findings suggest that the proposed mechanism has good potential in practical medical applications.