EEG Based Diagnosis of Stress and Mental Illness Using Feature Selection Methods and Classifiers

Abstract

Detecting mental stress levels through EEG signal analysis poses several challenges. Classical newlineapproaches have proven to be inadequate in providing accurate estimations, while neural network newlineapproaches show promising results. However, irregular data shapes and multiparametric newlinedependencies in EEG signals hinder the performance of convolutional neural networks (CNNs). newlineTo address these challenges, this research introduces a novel approach that utilizes image-based newlinespectrogram processing techniques to overcome shape irregularities and enhance the convergence newlinecapabilities of the network models. Specifically, 2D images obtained from azimuthal projections newlineof EEG signals are processed using a CNN for stress detection. The efficacy of this approach is newlineevaluated by varying the number of convolutional layers in the model, resulting in an impressive newlineaccuracy of 93% in mental stress detection using the DASPS EEG dataset. newlineMoreover, this thesis presents StressNet, a distinctive architecture that combines a 2D CNN and a newlinelong short-term memory (LSTM) network to detect stress states in EEG signals. The proposed newlinemethod decomposes the signals into alpha, beta, and theta components, generating azimuthal newlineprojection-based images. These images are then fed into the CNN for feature extraction, followed newlineby the LSTM for modelling the temporal dynamics. The StressNet model achieves exceptional newlineaccuracy, surpassing human stress detection rates with an impressive 97.8% accuracy on DEEP newlineand SEED datasets. newlineFurthermore, this work encompasses an extensive EEG dataset collected during various types of newlinestimulation, including mathematical problem-solving, horror video watching, mental challenge newlinetests, and relaxed state recordings. This dataset serves as a valuable resource for studying the neural newlinecorrelates of stress and developing stress detection models based on EEG data. The Azimuthal newlineProjection and StressNet models presented in this thesis provide powerful tools for stress newlinemanagement, healthcare applications, and workplace safety.