Integrated multimodal analysis for high-accuracy anxiety disease subtype classification

Abstract

In this study, we propose a classification method for identifying subtypes of anxiety disorders (AD). A large dataset is built with 108 healthy controls and 179 subjects from four primary AD subtypes: generalized anxiety disorder (GAD), social anxiety disorder (SAD), panic disorder (PD), and specific phobia (SP). We calculate diverse multimodal data, including amplitude of low-frequency fluctuations, regional homogeneity, and voxel-based morphometry, and create brain gradient data to provide a comprehensive representation of these data. For subtype classification, we develop a hierarchical binary hypothesis testing (H-BHT) framework with a two-stage scheme. In the first stage, we use a traditional BHT method to identify AD individuals. In the second stage, we categorize AD subjects into different subtypes under multi-class hypotheses. Our experiments demonstrate that the gradient data outperforms single-modal data in subtype classification, achieving an impressive 97.9% accuracy. When performing a multivariate analysis of variance on the brain regions associated with the discriminative gradient data, it reveals significant biomarkers among the subtypes, including the insula, amygdala, orbital inferior frontal, middle frontal and anterior cingulate gyri. These regions are strongly correlated with emotion control, providing substantial support for the pathogenesis of existing AD subtypes and confirming the validity of our method.