Emotion Detection from EEG Signals using Machine Learning Techniques
Abstract
An Electroencephalograph (EEG) signal is the recorded brain activity through electrodes on the scalp. In the medical domain, EEG analysis is used to detect conditions such as brain tumors, seizures, epilepsy, and depression. Emotion detection from EEG signals has potential in various applications including marketing, workplace optimization, improvement of human-machine interfaces, and user experience. Recent studies apply different machine learning techniques to detect emotions such as k-nearest neighbors, support vector machine, convolutional and feed forward neural networks. However, the comparison of reported results from different studies is difficult as they use different datasets and evaluation techniques. Examples include a hold-out evaluation with random test set selection from random subjects, individual models or one global model, and various versions of cross-validation. Moreover, most studies have focused on extracting frequency-based features and then using those features for classification. In Contrast, this thesis focuses on automatically extracting features from time series data by directly applying deep learning algorithms, namely, recurrent and convolutional neural networks to classify emotions using EEG data. The results on the DEAP dataset varied from 55.1% to 95.9% in classification accuracy. For comparison purposes, these models were compared to a feed forward neural network model with two commonly used frequency-based features: power spectral density and differential entropy. Extracted features achieved accuracy ranging from 54.5% to 69.6%. Also, this thesis considers two evaluation scenarios based on each subject's data. Results show that the classification performance varies greatly depending on the evaluation technique and highlight the need for a consistent evaluation technique in order to compare different studies.