Hierarchical supervised/unsupervised approach for subtype and redefine psychiatric disorders using resting state functional magnetic resonance imaging

Abstract

Information technologies such as deep learning and machine learning have made remarkable progress, and their effectiveness has been demonstrated in brain imaging research related to psychiatric disorders. For example, the application of supervised learning to resting‐state functional magnetic resonance imaging (fMRI) data has been used to identify psychiatric disorders based on their biological basis, and unsupervised learning has been used for subtyping of psychiatric disorders. However, there have been problems such as small effect sizes on the relationship between resting brain activity and cognitive functions, inter‐imaging‐site differences in brain imaging data, and development of the technologies based on DSM diagnoses. In this paper, I discuss these problems, the extent to which brain imaging studies have revealed psychiatric disorders, and what needs to be done now. I introduce our efforts to overcome these problems, constructing a large multi‐imaging‐site, multi‐disorder dataset, developing a novel harmonization technique to mitigate inter‐site differences in brain imaging data, and investigating subtypes of psychiatric disorders based on the brain circuit.