Using deep learning to identify brain networks mediating cognitive and motor impairments in alcohol use disorder

Alcohol Use Disorder (AUD), with a lifetime prevalence of 29.1% in the U.S., is associated with functional impairment affecting visuospatial working memory, executive functions, and motor control. The objective of this study was to distinguish people with AUD from controls on the basis of functional brain and neuropsychological measures that would contribute to identifying mechanisms of AUD-related dysfunction. A data-driven, deep-learning framework jointly analyzed 6105 region-to-region connections from resting-state functional MRI and 16 cognitive and motor performance scores. The deep learning method first derived 16 brain networks aligned with neuropsychological functions and then combined them into 14 functional units. After determining the most important functional unit for diagnostic classification, mediation analysis identified the neural pathways of that unit through which AUD affects neuropsychological performance. The Temporal Attention Network (TAN) fully mediated the effect of AUD diagnosis on spatial working memory (Visual Span). TAN also fully mediated the effects of AUD on visually guided attention, set-shifting, and motor performance (Trail Making Test), which, in parallel was mediated by a second network, the Sensorimotor Network (SMN). In conclusion, selective and dissociable brain functional and neuropsychological relationships differentiated individuals with AUD from controls. These relations, which were identified with deep learning technology and replicated on an independent dataset of people with HIV (with or without AUD comorbidity), provide support for brain functional substrates of commonly observed, AUD-related neuropsychological deficits.