Loading...
Detecting Mental Stress using EEG and Deep Learning
Badr, Yara
Badr, Yara
Date
2023-04
Author
Type
Thesis
Degree
Citations
Altmetric:
Files
Description
A Master of Science thesis in Biomedical Engineering by Yara Badr entitled, “Detecting Mental Stress using EEG and Deep Learning”, submitted in April 2023. Thesis advisor is Dr. Hasan Al-Nashash and thesis co-advisors are Dr. Usman Tariq and Dr. Fares Al-Shargie. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
Stress plays a significant role in the development of mental, emotional, behavioral, and physical illnesses, and impairs focus, concentration, and performance. The objective of this thesis is to devise a novel approach to identify and alleviate stress levels using electroencephalogram (EEG) signals combined with deep learning techniques and binaural beats stimulation (BBs). The study involved an experiment under four different mental states: rest, control, stress, and stress mitigation. During the stress state, all participants performed Stroop Color-Word Task (SCWT) under time pressure. Meanwhile, in the stress mitigation, the participants performed the SCWT while listening to 16 Hz BBs. EEG, salivary cortisol, behavioral, and subjective measures were used to quantify stress levels, and a novel approach was proposed by merging Partial Directed Coherence (PDC) with Graph Convolutional Network (GCN). Two scenarios were investigated, one including all 45 participants, estimating them to have the same average baseline and detecting 4 mental states, and another where we divided the participants into two different baseline groups (each with 22 subjects) based on subjective data, thus ending up with 8 mental states. In scenario 1, we found that BBs increased target detection accuracy by 27.08% (p<0.001), while in scenario 2, BBs improved detection accuracy by 31.6% and 22.8% for groups 1 and 2, respectively. The improved detection accuracy could be attributed to the beta state induced by the 16 Hz wave. However, there was no significant change noticed for the Perceived Stress Scale (PSS-10) and cortisol. Nevertheless, using PSD topography, a shift in cortical activity back to the temporal region was observed during mitigation, signifying recovery of participants' mental activity and focus. The deep learning results showed that the GCN-PDC could discriminate between four distinct mental states with average accuracies of 99.59%, 99.40%, 99.26%, and 99.64% in alpha, beta, delta, and theta bands for scenario 1, and could classify between 8 mental states (low rest, high rest, low control, high control, low stress, high stress, low mitigation, and high mitigation) with average accuracies of 98.49%, 98.38%, 98.12%, and 98.49% in alpha, beta, delta, and theta bands for scenario 2.