EEG based emotion recognition system using MFDFA as feature extractor

TLDR: An effective classifier named Support Vector Machine (SVM) is introduced to categorize the EEG feature space related to various emotional states into their respective classes and the result reveal that frontal, temporal and parietal regions of the brain are relevant to positive emotion recognition and frontal andParietal regions are activated in case of negative emotion identification.

Abstract: Emotion is a complex set of interactions among subjective and objective factors governed by neural/hormonal systems resulting in the arousal of feelings and generate cognitive processes, activate physiological changes such as behavior. Emotion recognition can be correctly done by EEG signals. Electroencephalogram (EEG) is the direct reflection of the activities of hundreds and millions of neurons residing within the brain. Different emotion states create distinct EEG signals in different brain regions. Therefore EEG provides reliable technique to identify the underlying emotion information. This paper proposes a novel approach to recognize users emotions from electroencephalogram (EEG) signals. Audio signals are used as stimuli to elicit positive and negative emotions of subjects. For eight healthy subjects, EEG signals are acquired using seven channels of an EEG amplifier. The result reveal that frontal, temporal and parietal regions of the brain are relevant to positive emotion recognition and frontal and parietal regions are activated in case of negative emotion identification. After proper signal processing of the raw EEG, for the whole frequency bands the features are extracted from each channel of the EEG signals by Multifractral Detrended Fluctuation Analysis (MFDFA) method. We introduce an effective classifier named Support Vector Machine (SVM) to categorize the EEG feature space related to various emotional states into their respective classes. Next, we compare Support Vector Machine (SVM) with various other methods like Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA) and K Nearest Neighbor (KNN). The average classification accuracy of SVM for positive emotions on the whole frequency bands is 84.50%, while the accuracy of QDA is 76.50% and with LDA 75.25% and KNN is only 69.625% whereas, for negative emotions it is 82.50%, while for QDA is 72.375% and with LDA 65.125% and KNN is only 70.50%.