Random Forest Classification of Alcohol Use Disorder Using EEG Source Functional Connectivity, Neuropsychological Functioning, and Impulsivity Measures.

Seed regions of the default mode network (DMN) consisting of 6 regions in the left hemisphere (blue beads) and 6 homologous regions in the right hemisphere (red beads) as listed in Table 2. Axial (top), coronal (front), and sagittal (left) views are shown. [PCC–Posterior cingulate cortex; ACC–Anterior cingulate cortex; IPL–Inferior parietal lobule; PFC–Prefrontal cortex; LTC–Lateral temporal cortex; PHG–Parahippocampal gyrus; L–Left; R–Right; A–Anterior; P–Posterior; S–Superior; I–Inferior].

The multi-way importance plot showing the top significant variables (labelled and marked with black circles) that contributed to the classification of alcohol use disorder from control individuals based on the measures Gini decrease, number of trees, and p-value. All 4 impulsivity scores, 29 FC connections, and 3 neuropsychological variables were significant (circled and labelled red and gray dots). Impulsivity scores stood on top in the importance list. Note that the variables that were not significant or important (purple dots) are not highlighted. [Abbreviations in the FC variable labels: BIS–Barratt Impulsivity Scale; MI–Motor impulsivity; NP–Non-planning; AI–Attentional impulsivity; Tot–Total; Span_Fw–Span forward; TotCor_Fw–Total correct forward; TotCorAvgTime_Fw–Total correct average time forward; TotCor_Bw–Total correct backward; De–Delta; Th–Theta; Al–Alpha; Be–Beta; Ga–Gamma; s1-s12–default mode network (DMN) seeds 1-12 as listed in Table 2].

The distribution of minimal depth among the trees of the forest for the significant variables is shown in different colors for each level of minimal depth. The mean minimal depth in the distribution for each variable is marked by a vertical black bar overlapped by a value label inside a box. Lower mean minimal depth of a functional connectivity variable represents higher number of observations (participants) categorized in a specific group on the basis of the variable (i.e., better classification). The top significant variables (29 functional connectivity connections, 3 neuropsychological scores and all 4 impulsivity scores) followed the same rank in the plot as in Table 3, which is ordered based on p-values. [Abbreviations in the FC variable labels: De–Delta; Th–Theta; Al–Alpha; Be–Beta; Ga–Gamma; s1-s12–default mode network (DMN) seeds 1-12 as listed in Table 2].

Illustration of rankings of variables based on any of the two Random Forest (RF) parameters of importance (panels in the lower triangle of the grid showing distribution of rankings of all predictor variables with black dots along a blue trend line) as well as correlation coefficient across rankings of any two parameters (panels in the upper triangle of the grid). It is shown that all RF parameters of importance were found to have very high correlations among each other, suggesting high reliability of each of these parameters to rank the importance of variables for group classification.

Panels (a–e): Significant default mode network (DMN) connections within each frequency band, which contributed to the Random Forest classification of alcohol use disorder (AUD) from control (CTL) individuals based on the importance parameters, viz., p-value (p < 0.05), number of trees, and Gini index as listed in Table 3. Panel (f): Common connections across different frequency bands. The three images within each panel represent three views, respectively: axial or top view, coronal or front view, and sagittal or left view. The blue and red beads represent left and right-sided DMN nodes/seeds respectively, and the connections represented in orange/red and cyan/blue lines indicate hyper- and hypo-connectivity, respectively, in AUD compared to CTL group. Note that one of the four common connections, L.ACC–R.PHG, although shown here in orange/red (panel f), showed hypoconnectivity in beta band (panel d) and hyperconnectivity in gamma bands (panel e).

Significant default mode network (DMN) connections based on a stricter threshold (p ≤ 0.001) are shown. The three images within each panel representing specific frequency represent these three views, respectively: axial or top view, coronal or front view, and sagittal or left view. The blue and red beads represent left and right-sided DMN nodes/seeds respectively, and the connections represented in orange and cyan lines indicate hyper- and hypo-connectivity, respectively, in alcohol use disorder (AUD) compared to control group.

Correlation matrix showing associations among the top significant variables based on explorative (descriptive) correlational analysis for the interpretative purpose. Values within each cell represents bivariate Pearson correlation between the variable on its vertical axis and the variable on its horizontal axis. Correlation values are color coded [red/pink shades represent negative r-values, blue/cyan shades indicate positive r-values, darker color represent higher magnitude] and significant correlations (before Bonferroni correction) have been marked with asterisks [*p < 0.05; **p < 0.01; and ***p < 0.001]. The correlations that survived Bonferroni correction (r ≥ 0.50 and p < 0.00005) have been marked with a filled circle at the end of triple asterisks (***●). Abbreviations in the variable labels: De–Delta; Th–Theta; Al–Alpha; Be–Beta; Ga–Gamma; s1-s12–default mode network seeds 1-12 as listed in Table 2.