Classification of EEG Signals in Two Levels of Normal and Anxious Using Nonlinear Features

Anxiety is a natural reaction of humans to stress that occurs in the face of various factors. Anxiety is considered as a mental illness if it is excessive and uncontrollable in the form of fear and anxiety. Today, clinicians use certain criteria to diagnose anxiety disorders. This analytical-observational study was aimed at automatically classifying the two levels of anxious and normal by analyzing electroencephalogram signals. In this paper, the DASPS database was used, which contains a 14-channel electroencephalogram of 23 people (13 females and 10 males, mean age 30 years) during anxiety. Anxiety was presented in the form of flooding as actual exposure to the feared stimulus. Based on the results of the Self-Assessment Manikin, data were divided into two groups: (1) normal and low anxiety and (2) moderate and high anxiety. Approximate entropy, fractal dimension, and Lyapunov exponents were extracted from all channels as nonlinear properties. Maximum relevance and minimum redundancy were used to select the best feature to apply to the multilayer perceptron network. To evaluate the performance of the algorithm, different network structures were examined in terms of the number of features and neurons as well as different feature dimensions. Maximum accuracy, precision, f1-score, and sensitivity in 20 repetitions in all cases is equal to 100, and with an increasing number of neurons, the average accuracy increases. The best results were obtained for 5 features and 15 neurons, where the mean accuracy, precision, f1-score, and sensitivity for it were 80%, 92.75%, 84.15%, and 80.58%, respectively. The results of this paper indicated the capability of the proposed algorithm to distinguish anxious people from normal ones.