STRESS DETECTION USING NEW TIME-FREQUENCY DECOMPOSITION: PROGRESSIVE FOURIER TRANSFORM

Abstract

Introduction: Stress is a natural reaction to challenges encountered in everyday life. Chronic stress, which lasts for a long time, can negatively influence mental and physical health. Therefore, early detection and assessment of stress are crucial to reducing the risk of harm to an individual's well-being. Electroencephalograph (EEG) brain signals can be used to assess human stress levels. This research aims to investigate how EEG signals can detect stress and mental states using deep learning and feature extraction techniques. Method: Several feature extraction methods and classification models were developed in our proposed framework to extract and classify mental states’ features captured from the frequency domain and time-frequency domain of EEG signals. We proposed new feature decomposition approaches based on the progressive Fourier transform and the coordination of multiple brain areas working simultaneously. The performance of the proposed methods with different classification models were evaluated on publicly available EEG datasets. Results: Experiment results revealed that our proposed methods outperformed previous studies detecting stress and mental states. Using the AlexNet model to classify the three mental states (Concentrating, neutral, and relaxed), our novel proposed progressive Fourier transformation achieved the highest accuracy of 98.4%. Meanwhile, another model based on convolutional neural networks classified three stress levels (Stressed/Mild-stressed/Non-stressed) and two stress levels (Stressed/Non-stressed) with 96.6% and 96.3% accuracy, respectively, by concatenating the image representations of EEG extracted features from each channel. Conclusion: We believe our proposed framework can be used in real-world solutions to diagnose or treat patients suffering from stress and mental diseases by utilizing virtual reality (VR) simulation and exposure psychotherapy while using commercial EEG devices.

Date of Award	2022
Original language	American English
Awarding Institution	HBKU College of Science and Engineering