Alcohol use disorder detection using EEG Signal features and flexible analytical wavelet transform

Abstract

The frequent excessive drinking of alcohol severely affects the neuronal composition and working of the brain and consequently developed Alcohol Use Disorder (AUD). Subjects suffering from AUD are prone to various diseases, psychological and cognitive issues if not identified and treated timely. Electroencephalogram (EEG) signals are used to record the internal structure and activity of the brain. The manual screening of EEG signals for AUD detection is complicated for practitioners because EEG signals are recorded in microvolts (μv) and consists of the inherent internal complexity of the brain. Therefore, an automated computer-aided diagnosis (CAD) is used for assisting the medical practitioner in AUD screening process. The recorded EEG signals of a subject are nonlinear and oscillatory, and CAD methods examine these signals in their frequency sub-bands. In this paper, flexible analytical wavelets transform (FAWT) based machine learning models are proposed for automated alcoholism detection. In the proposed methodology, EEG signals are decomposed into approximate and detailed wavelet coefficients using FAWT. The statistical features such as mean, standard deviation, kurtosis, skewness, and Shannon entropy are extracted from the selected wavelet coefficients. The features are fed to the various machine learning models including Least Square -Support Vector Machine (LS-SVM), Support Vector Machine (SVM) and Naïve Bayes learners for training. The training and testing are performed using 10-fold cross-validation. The performance of models is evaluated using all essential measures such as accuracy, sensitivity, specificity, F-measure, precision, Matthews correlation coefficient (MCC) and ROC. The results suggest that LS-SVM using polynomial kernel performed best with accuracy 99.17%, Sensitivity 99.17%, and Specificity as 99.44% using 10-fold cross-validation technique.