Alcoholism detection using support vector machines and centered correntropy features of brain EEG signals

The regular excessive consumption of alcohol may leads to alcohol use disorder (AUD). AUD is considered a serious health issue which may hamper physical health and social life of a patient if not detected and treated timely. The screening of AUD patients using physiological characteristics is very difficult task for the doctors. Therefore, brain electroencephalograms (EEGs) signal analysis is popularly used to accurately detect the AUD disorder (or the person is alcoholic or normal). Manual analyses of EEG signals are complicated and time-consuming as it is recorded in microvolt (μv). Therefore, computer-aided diagnosis (CAD) is used by a neurologist to analyze the EEG signals from their frequency sub-bands. The EEG signals recording obtained from the subject are nonlinear and unstable with respect to time. In this paper, support vector machine techniques are used with nonlinear parametric signals in time-frequency domain on features extracted from EEG signals. The features are extracted with the help of continuous wavelet transform, Tuned Q wavelet transform (TQWT) is used for decomposition of the signals. The decomposed subband, Centered Correntropy (CC) features are extracted and used to find out the minute changes in the nonlinear signal with lag time delay, which is very similar to the autocorrelation of the signal. Then these features are reduced by applying principal component analysis (PCA), which is then passed to least squares support vector machine (LS-SVM) for classification between alcoholic and normal EEG signals. Training of the data is done with ten-fold cross-validation to increase the accuracy. Our proposed work is with three different types of kernel functions like linear, RBF and Polynomial kernel. Comparing the result, we found that the RBF kernel gives the highest performance having accuracy is 97.06%, sensitivity is 97.45%, specificity is 97.666%, and Matthews's correlation coefficient is 0.94 by varying the Q parameter of wavelet transform from 3 to 8.