Multi-domain Feature Extraction Methods for Classification of Human Emotions from Electroencephalography (EEG) Signals

Over the past several decades’ research on emotions has gained huge popularity with contributions from many interdisciplinary and transdisciplinary researchers from areas like medicine, psychology, computer science, etc., contributing to the cause. Recognizing human emotions through the electroencephalogram signal is a trending topic nowadays. Neurological studies on multiple physiological datasets like DEAP, SEED, etc., have contributed to developing various applications for human emotion recognition. The main contribution of this study is to develop a supervised model that uses less data to predict human emotions without compromising much on the accuracy, for that we extracted multi-domain (time and frequency) features from EEG signal along with gamma frequency band for classification of emotions. The preprocessed open-source physiological EEG GAMEEMO dataset is taken for the experiment; the dataset was obtained by playing four different emotional computer-based video games with the subjects. In the first step of our study, we used Discrete Wavelet Transform to extract the gamma frequency from the GAMEEMO dataset. The second step is to extract Frequency Domain Features and Time Domain Features from the transformed data. In the third step, we used some supervised classifiers like Support Vector Machine, K-nearest neighbor, and Convolutional neural Networks for human emotion classification. Accuracy, F1-measure, Precision, Recall, and Kappa score are our performance metrics. Out of all the supervised learning models which are built, SVM gave the best results of 82.60 accuracy, 0.76 kappa score, 0.75 precision, 1.0 recall, and 0.86 F1-score. Instead of studying the entire signal, studying gamma band frequency can also give sufficient results for the study of emotions. The proposed methods can be extended to use multiple EEG signal datasets to classify human emotions in future studies.