Identification of Binge Drinkers via Convolutional Neural Network and Support Vector Machine

Studies have described neural and psychosocial markers of binge drinking. Deep learning would address how well these markers distinguish binge and non-binge drinkers. We examined the data of 180 hinge and 282 non-hinge drinking young adults from the Human Connectome Project. We randomly selected 90% of the subjects as training sample to build convolutional neural network (CNN) and support vector machine (SVM) models, and evaluated their performance in the remaining 10%. Imaging data were processed with published routines. 2D/3D-CNN of gray matter volumes (GAM) exhibited an area under the curve (AUC) of 0.802/0.812 and SVM of psychosocial measures, GMVs and cortical thickness each exhibited an AUC of 0.883, 0.746 and 0.589 in the classification of binge and non-binge drinkers. Among the psychosocial measures, rule breaking behavior score showed the greatest difference and contributed most significantly to the classification in SVM model. Among the GMVs, left cerebellum showed the greatest difference in GMV and contributed most significantly' to the classification in SVM model. These findings show that, associated with subtle cerebral volumetric differences, young adult binge drinking is best predicted by psychosocial measures.