The use of whole-brain models and variational autoencoders for the low-dimensional representation of psychosis and its perturbational landscape

Abstract

Psychosis can be described as an alteration in brain connectivity that leads to an impairment of cognition and the speed at which the information gets processed, what causes a diversity of psychiatric symptoms. This symptomatology is characterized by changes in the brain activity in certain areas, which can be detected by Functional Magnetic Resonance Imaging (fMRI) as it registers changes in the brain associated with blood flow, and this allows us to measure brain activity and connectivity between regions. Furthermore, the state of these alterations may differ between patients depending on the severity of their condition and the number of episodes they have had or may suffer. This study focuses on the use of the connectivity and structural information extracted from fMRIs and a whole-brain model to generate synthetic data with enough resemblance to the original dataset cases to train a Variational Autoencoder architecture for the creation of a low dimensional space in which the cases where patients have had one psychotic episode (remitting) or multiple (relapsing) are represented, and therefore a classification model can be trained to distinguish them. A dimensionality analysis has been performed to find the most optimal dimension of this space that allow us to distinguish between remitting and relapsing cases with high enough accuracy. Moreover, perturbations were introduced in the original model to generate new data which was reclassified in the low dimensional space to find which alterations could produce changes in the classification of the psychotic stage.