Graph Clustering for Large-Scale Text-Mining of Brain Imaging Studies

This paper presents a Graph clustering method for identifying functionally similar key concepts for meta-analysis of brain imaging studies. We use an existing database of key concepts created by a large-scale automated text mining of brain imaging studies. The key concepts here refer to specific psychological terms of interest (for instance, 'decision', 'memory' etc) identified based on their frequency of occurrence (>1 in 1,000 words) in an individual article text of 5809 studies. The pair-wise distance between all 525 nodes was calculated using the Jaccard metric. A graph was created with 525 nodes representing the key concepts. An undirected edge was drawn from every node to the node with minimum distance. We present a clustering approach using a simple graph traversal to identify connected components so that every node belongs to exactly one cluster. The results from our clustering method reveal semantically related concepts confirming potential for further use in text-mining approaches for meta analysis of brain imaging studies.