Adolescent alcohol use is linked to disruptions in age-appropriate cortical thinning: an unsupervised machine learning approach

Cortical thickness changes dramatically during development and is associated with adolescent drinking. However, previous findings have been inconsistent and limited by region-of-interest approaches that are underpowered because they do not conform to the underlying spatially heterogeneous effects of alcohol. In this study, adolescents (n = 657; 12-22 years at baseline) from the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) study who endorsed little to no alcohol use at baseline were assessed with structural magnetic resonance imaging and followed longitudinally at four yearly intervals. Seven unique spatial patterns of covarying cortical thickness were obtained from the baseline scans by applying an unsupervised machine learning method called non-negative matrix factorization (NMF). The cortical thickness maps of all participants' longitudinal scans were projected onto vertex-level cortical patterns to obtain participant-specific coefficients for each pattern. Linear mixed-effects models were fit to each pattern to investigate longitudinal effects of alcohol consumption on cortical thickness. We found in six NMF-derived cortical thickness patterns, the longitudinal rate of decline in no/low drinkers was similar for all age cohorts. Among moderate drinkers the decline was faster in the younger adolescent cohort and slower in the older cohort. Among heavy drinkers the decline was fastest in the younger cohort and slowest in the older cohort. The findings suggested that unsupervised machine learning successfully delineated spatially coordinated patterns of vertex-level cortical thickness variation that are unconstrained by neuroanatomical features. Age-appropriate cortical thinning is more rapid in younger adolescent drinkers and slower in older adolescent drinkers, an effect that is strongest among heavy drinkers.

Automated summary

This study used unsupervised machine learning to successfully identify spatial patterns of cortical thickness variation at the vertex level, which are not constrained by neuroanatomical features. The findings suggest that age-appropriate cortical thinning is faster in younger drinkers and slower in older drinkers, with the strongest effect observed in heavy drinkers.