Automated Alcoholism Detection Using Fourier-Bessel Series Expansion Based Empirical Wavelet Transform

In this paper, the Fourier-Bessel series expansion based empirical wavelet transform (FBSE-EWT) is proposed for automated alcoholism detection using electroencephalogram (EEG) signals. The FBSE-EWT is applied to decompose EEG signals into narrow sub-band signals using a boundary detection approach. The accumulated line length, log energy entropy, and norm entropy features are extracted from different frequency scales of narrow sub-band signals. A total of twenty features are extracted from each attribute and out of which ten features are from low to high frequency sub-band signals and other ten features are from high to low frequency sub-band signals. In order to reduce the classification model complexity, the most significant features are selected using feature selection techniques. Six feature ranking methods such as Relief-F, <inline-formula> <tex-math notation="LaTeX">${t}$ </tex-math></inline-formula>-test, Chi-test, relief attribute evaluation, correlation attribute evaluation, and gain ratio are used to select the most common features based on the majority voting technique. Experiments are performed by considering top ranked 5, 10, 15, and 20 features and classification methods such as least square support vector machine (LS-SVM), support vector machine (SVM), and <inline-formula> <tex-math notation="LaTeX">${k}$ </tex-math></inline-formula> nearest neighbor (<inline-formula> <tex-math notation="LaTeX">${k}$ </tex-math></inline-formula>-NN) classifiers. The training and testing is done using leave-one out cross-validation (LOOCV) in order to avoid over-fitting. The performances of classifiers are evaluated using accuracy, sensitivity, and specificity measures. The results suggest that LS-SVM with radial basis function (RBF) kernel achieves a highest average accuracy of 98.8%, sensitivity of 98.3%, and specificity of 99.1% with top 20 significant features.