Behavior Cognition Association Analysis with Heterogeneous Multimodal Causal Discovery

Describing the association between external behavior and internal cognition is significant for revealing the pathogenesis of cognitive diseases (e.g., dementia). Existing causal discovery methods perform well on homogeneous data, but struggle to handle heterogeneous multimodal data, due to the differences in data structure, distribution, and representation. Therefore, we propose a Heterogeneous Multimodal Causal Discovery method for Behavior Cognition Association Analysis (HMCD-BCA). This method employs Convolutional Auto-Encoder Network to extract features of heterogeneous multimodal data and map the multimodal data to a unified feature space. Based on consistent features, classical causal discovery algorithms are used to calculate causal associations between behavioral data and cognitive data. We validate our method on both simulated dataset and real behavior-cognition bimodal dataset. The experimental results demonstrate the effectiveness of our method in inferring causal associations between behavior and cognition.