UNSUPERVISED DISCOVERY OF MENTAL DISORDER FACTORS USING MRI

Abstract

Even though there are classification criteria for each mental disorder, significant heterogeneity still exists within a disorder. Consequently, many studies have defined mental disorder subtypes, where each participant belonged to a single subtype. However, such categorical subtypes might obscure continuous individual variation, thus motivating a dimensional approach.

In this thesis, we develop two hierarchical Bayesian models, MMLDA and polar-LDA, which allow an individual to express one or more factors (categorical subtypes) to varying degree (continuous), thus potentially reconciling dimensional and categorical models of mental disorders.

Importantly, we demonstrated that the Alzheimer's factors are associated with distinct patterns of brain atrophy, cognitive deficits and tau deposition, while the autism factors are associated with distinct brain functional connectivity patterns and behavioral symptoms. The resulting individual-specific factor loadings can be interpreted as an individualized subtype diagnosis of an individual, representing a small, but crucial step toward precision medicine.