Abstract
Overcoming an initial psychotic episode does not always lead to recovery; relapses and subsequent psychotic episodes may happen afterward. Even if the characterization of psychotic disorders can be related to alterations in brain connectivity, clear identification of the brain areas for relapse is missing. Here, we leverage on whole-brain modeling linking anatomical structural information with functional activity as measured by MRI in 196 participants. Patients were classified into Stage II (first episode), IIIa (incomplete remission), IIIb (remission followed by one relapse), and IIIc (remission followed by several relapses), depending on the course of psychosis up to the time of the brain scan. From these data, a low-dimensional manifold reduction of the brain dynamics was obtained using deep learning variational autoencoders in which the different stages are represented, and a classification model can be trained to distinguish them. Then, a dimensionality analysis was performed to find the optimal dimension that allows the distinction between first episode and relapsing cases with high accuracy. Finally, perturbations were introduced in the model to reveal the brain regions associated with the absence of relapse, which could help predict which brain regions to target during therapy and assist the treatment of patients suffering from psychotic disorders.