78. Machine Learning Classification of Obsessive-Compulsive Disorder Using Structural Neuroimaging Data: ENIGMA Working Group

Abstract

Background: Recent meta-analyses of structural neuroimaging data within ENIGMA consortia have reported significant but small group differences in brain structure between patients with obsessive-compulsive disorder (OCD) and healthy controls. We aimed to determine whether patients could be distinguished from controls at the individual level using multivariate pattern analysis, and to investigate whether clinical variables would influence classification performance. Method(s): Structural T1 images from 2,304 OCD patients and 2,066 healthy controls were obtained from 34 sites across the world. FreeSurfer was used to determine cortical thickness, cortical surface area and subcortical volumes. All 158 FreeSurfer variables were used as features for classification, with age and sex as covariates. Eight common machine learning algorithms were used, as well as different cross-validation (CV) procedures to determine the generalizability of the results. Additional classifications were performed using stratification for medication status, disease severity, chronicity, and age of onset. Result(s): Within-site classification performance with 10-fold CV varied greatly between sites, with area under the receiver-operator curve (AUC; range (SD)) 0.21(0.15)-0.89(0.08). Between-site classification performance was low with 10-fold CV stratified for site (0.57(0.02)-0.60(0.02)), and at chance level with leave-one-site-out CV (0.51(0.08)-0.54(0.08)). Classification performance improved considerably when accounting for medication status and age of onset. Conclusion(s): Between-site classification performance is poor and appears to result from heterogeneity between sites. These findings suggest that patients are difficult to distinguish from healthy controls using structural MRI data when combining multi-site data and not accounting for clinical diversity. In contrast, medication use has profound effects on brain structure and enables good single subject classification. Supported By: NWO/ZonMw Vidi 917.15.318 Keywords: Obsessive Compulsive Disorder (OCD), Structural MRI, Machine LearningCopyright © 2019