Knowledge-Concept Diagnosis from fMRIs by Using a Space-Time Embedding Graph Convolutional Network

Diagnosing the contents of learning in brain activities is a long-standing research task in cognitive sciences. The current studies on cognitive diagnosis (CD) in education determine the status of knowledge concept (KC) based on the observed responses to test items. However, the learning process of KC in the brain is left with no touch. This paper proposes to solve the problem of knowledge-concept diagnosis (KCD) from fMRIs by identifying the concepts a student focuses on in learning activities. Using the graph convolutional network (GCN), we introduce the STEGCN approach composed of a spatial GCN for brain-graph structure, a temporal GCN for brain-activity sequence, and a fully connected network for KCD. To evaluate STEGCN, we acquired an fMRI dataset that was collected on five concepts when students were learning a computer course. The experiment results demonstrate that our proposed method yields better performance than traditional models, showing the effectiveness of STEGCN in concept classification. This study contributes to a new fMRI-based route for knowledge-concept diagnosis.