Recent approaches dealing with machine learning analyses have used a two-stage approach, consisting of feature selection followed by a predictive algorithm using a selected sets of variables [75,76,77,78,79]. Feature selection methods are used as the first stage to reduce irrelevant and redundant variables, which may otherwise add noise to the predictive models [75,76,77]. The advantages of feature selection include a better understanding of the data, reducing computation requirements, mitigating the effect of the curse of dimensionality, and also improving the predictor performance [76]. We applied binomial lasso regression [80,81,82] as a feature selection method [83,84], as implemented in R-package “glmnet”, to extract a subset of fMRI FC variables (N = 561) that held significant predictive value to discriminate AUD from the CTL group. Feature selection was implemented only for the rsFC variables, due to a high number of connections, most of which were deemed not relevant for our purpose of AUD classification. The method adapted in the current analysis is based on the Lasso method, as implemented in Fonti and Belitser [84]. The maximum number of output features “pmax” was
