The RF algorithm correctly identified group membership of 26 out of 30 individuals in each group, in classifying them into either AUD or CTL group, with an accuracy rate of 86.67% and the area under the curve of 93% (Figure 3). The OOB error or the misclassification rate was 13.33% (for each group). The model also identified 12 rsFC connections and two impulsivity scores (motor and non-planning) as significantly (p < 0.05) contributing to the classification (Table 3). Relative to the CTL individuals, AUD subjects showed a predominant pattern of hypoconnectivity (i.e., decreased rsFC in 9 out of 12 connections) across the major cortical and subcortical nodes of the reward network, in addition to three connections with hyperconnectivity in specific nodes (i.e., left nucleus accumbens–left posterior cingulate cortex (PCC), right pallidum–right PCC, and right hippocampus–left dorsolateral prefrontal cortex). AUD individuals also showed increased impulsivity in motor and non-planning categories. However, none of the neuropsychological variables were significant based on the p-value criterion.
