Emotion Quantification Along Valence Axis Using EEG Signals

Abstract

Emotion is an important feature of human personality that affects the quality of interaction between humans. It is fundamental to human experience and rational decisionmaking. Recently the researchers have shown great interest in detecting emotions automatically. A number of techniques have been employed for this purpose using channels such as voice and facial expressions. However, these channels are not very accurate and can be faked. The necessity of studying emotions has become necessary to fulfill the human need of brain computer interfacing. This task can be accomplished only if we can successfully classify emotions into its constituent classes. This task requires the accurate acquisition of physiological and EEG data. To achieve this objective a step is taken forward in the direction of classifying emotions into different classes from the EEG data. To accomplish the goal of classifying emotions into two classes Low Valence and High Valence i.e. unpleasant and pleasant along the valence axis, the EEG data available at enterface06 website collected using International Affective Picture System (IAPS) as a stimulus has been taken into consideration. For this, we required sophisticated features that have been extracted from the raw EEG data using tools like Power Spectral Density (PSD), Short Time Fourier Transform (STFT), and Event Related Potential (ERP). These features individually and in appropriate combinations were further used for emotion classification by using different classifiers such as Navie Bayes, Artificial Neural Network (ANN), Feedforward and Multilayer Layer Neural Network. The data available on the enterface 06 website was collected from five participants in three sessions, where all subjects were right handed males. In this, out of EEG raw data of 5 participants, the analysis has been performed on 3 participants namely P3, P4, and P5 and 7 electrodes namely Cz, F1, F2, FC1, FC2, Fz, and Pz. To reduce the data set, the EEG data for 3 participants has been down-sampled at sampling rate of 256Hz by using open software called EEG Lab. Using frequency domain, the maximum and minimum PSD values were determined according to different EEG frequency bands namely delta, theta, alpha, beta, and gamma Emotion is an important feature of human personality that affects the quality of interaction between humans. It is fundamental to human experience and rational decisionmaking. Recently the researchers have shown great interest in detecting emotions automatically. A number of techniques have been employed for this purpose using channels such as voice and facial expressions. However, these channels are not very accurate and can be faked. The necessity of studying emotions has become necessary to fulfill the human need of brain computer interfacing. This task can be accomplished only if we can successfully classify emotions into its constituent classes. This task requires the accurate acquisition of physiological and EEG data. To achieve this objective a step is taken forward in the direction of classifying emotions into different classes from the EEG data. To accomplish the goal of classifying emotions into two classes Low Valence and High Valence i.e. unpleasant and pleasant along the valence axis, the EEG data available at enterface06 website collected using International Affective Picture System (IAPS) as a stimulus has been taken into consideration. For this, we required sophisticated features that have been extracted from the raw EEG data using tools like Power Spectral Density (PSD), Short Time Fourier Transform (STFT), and Event Related Potential (ERP). These features individually and in appropriate combinations were further used for emotion classification by using different classifiers such as Navie Bayes, Artificial Neural Network (ANN), Feedforward and Multilayer Layer Neural Network. The data available on the enterface 06 website was collected from five participants in three sessions, where all subjects were right handed males. In this, out of EEG raw data of 5 participants, the analysis has been performed on 3 participants namely P3, P4, and P5 and 7 electrodes namely Cz, F1, F2, FC1, FC2, Fz, and Pz. To reduce the data set, the EEG data for 3 participants has been down-sampled at sampling rate of 256Hz by using open software called EEG Lab. Using frequency domain, the maximum and minimum PSD values were determined according to different EEG frequency bands namely delta, theta, alpha, beta, and gamma Short Time Fourier Transform (STFT) features were extracted using Fast Fourier Transform (FFT) function by applying the discrete FFT to the signal. In STFT we determined the mean in different EEG frequency bands. Another feature used for classification was Event Related Potential (ERP) wherein P100, N100, P200, N200, and P300 potentials were used as attributes for classification. On classification of STFT features with Naive Bayes, the accuracy determined was mere 51% where as the accuracy obtained with PSD and combination of PSD and STFT features with the same classifier was 56% and 64% respectively. The accuracy for ERP features remained at 56%. Since the accuracy achieved with these features using Naive Bayes classifier was very low another classifier with one hidden layer using Artificial Intelligence technique was implemented on MATLAB R2011a version. The accuracy obtained while using ERP as an attribute for all 3 participants stood at 76.59%. To further enhance the accuracy the work was performed on Multilayer 2-hidden layer network, which resulted in 100% accuracy on ERP features for two classes.