Non-local diffusion-based biomarkers in patients with cocaine use disorder

Abstract

Cocaine use disorder (CUD) is widely known to result in neurological reconfiguration which can be observed via local diffusivity characteristics of the brain. These changes can be highly correlated while simultaneously variable across patients with different comorbidities or histories of substance use. This implies that more complex neuroimage analysis tools may be necessary to better detect specific biomarkers that vary across these dimensions. We investigated white and gray matter integrity using voxel-based diktiometry (VBD) on whole brain diffusion tensor images (DTI) across a database of CUD patients and healthy controls using a purely data-driven approach. These VBD maps reveal significant cortical and subcortical differences that are indicative of these neural modifications including the insula, cerebellum, ventricles, thalamo-cortical radiations, and corpus callosum bundles. In order to disambiguate these results and investigate the heterogeneity of CUD, the VBD maps have been decomposed into five decorrelated biomarkers: one in the region surrounding the left insula, one implicating the corpus callosum, two concentrated in the left cerebellum, and the last concerning a proximal region of the interhemispheric fissure which serve as potential biomarkers playing a role in CUD. These decorrelated biomarkers have themselves been correlated with consumption patterns and psychiatric and borderline personality disorder scores on the CUD patient group. This preliminary approach to using machine learning techniques to both detect and disambiguate complex non-linear patterns shows promise for better understanding complex and heterogeneous disorders such as CUD.