Random forest based classification of alcohol dependence patients and healthy controls using resting state MRI

Abstract

Currently, classification of alcohol use disorder (AUD) is made on clinical grounds; however, robust evidence shows that chronic alcohol use leads to neurochemical and neurocircuitry adaptations. Identifications of the neuronal networks that are affected by alcohol would provide a more systematic way of diagnosis and provide novel insights into the pathophysiology of AUD. In this study, we identified network-level brain features of AUD, and further quantified resting-state within-network, and between-network connectivity features in a multivariate fashion that are classifying AUD, thus providing additional information about how each network contributes to alcoholism. Resting-state fMRI were collected from 92 individuals (46 controls and 46 AUDs). Probabilistic Independent Component Analysis (PICA) was used to extract brain functional networks and their corresponding time-course for AUD and controls. Both within-network connectivity for each network and between-network connectivity for each pair of networks were used as features. Random forest was applied for pattern classification. The results showed that within-networks features were able to identify AUD and control with 87.0% accuracy and 90.5% precision, respectively. Networks that were most informative included Executive Control Networks (ECN), and Reward Network (RN). The between-network features achieved 67.4% accuracy and 70.0% precision. The between-network connectivity between RN-Default Mode Network (DMN) and RN-ECN contribute the most to the prediction. In conclusion, within-network functional connectivity offered maximal information for AUD classification, when compared with between-network connectivity. Further, our results suggest that connectivity within the ECN and RN are informative in classifying AUD. Our findings suggest that machine-learning algorithms provide an alternative technique to quantify large-scale network differences and offer new insights into the identification of potential biomarkers for the clinical diagnosis of AUD.

Keywords: Alcohol dependence; Functional connectivity; ICA; Resting state; fMRI.

Figure 1

Spatial maps of 32 independent components grouped into 5 categories (Different colors represent different network sub-components in each category for A to D): A: default-model networks (3 components). B: Visual networks (7 components) C: Sensorimotor networks (6 components) D: executive control networks (12 components) E: all other networks (4 components) including salience network (Green), subcortical network (Blue), auditory network (Violet), cerebellar (Yellow)

Figure 2

Between-network functional connectivity features: correlation coefficients matrices between each pair of networks in AUD (left) and controls (right)

Figure 3

The procedure of feature extraction and pattern classification

Figure 4

Within-network features maximally contributing to classification performance including ECN1 (Violet), ECN2 (Blue), RN (Cyan)

Figure 5

The stability of within-network features that have the strongest predicting power (ECN1, RN and ECN2). The mean and standard error of feature importance of 92 random forests by using leave-one-out cross-validation (LOOCV) were calculated and presented in the figure.

Figure 6

The stability of between-network features that have the strongest predicting power. The prediction accuracy was computed 92 times by using leave-one-out cross-validation (LOOCV). The mean and standard error of feature importance of 92 random forests by using leave-one-out cross-validation (LOOCV) were presented in the figure.