2.2.1 DEAP Public Dataset
The aim of the research done by Meng and Zhang (2020) was to study and recognise the anxiety faced by college students using deep learning method. Two datasets including the public DEAP dataset and their homemade dataset were used in their experiment for validation purpose. The DEAP data experiment comprised 32 participants whereas there were 20 participants involved in the homemade data experiment.
For the EEG data pre-processing in this study, filtering method was used to remove the artifacts. Then, Linear Discriminant Analysis (LDA), Convolutional Neural Network (CNN) and PCA algorithms were studied and compared for feature extraction. According to the results, CNN showed the highest accuracy among the three algorithms, followed by PCA and LDA. Since EEG signals are generally non- linear and non-Gaussian, in terms of dimensionality reduction it is harder to separate the different classes from the EEG data using PCA and LDA as there are still some correlations between the classes. Opposingly, CNN is useful in extracting representational abstract features from the raw EEG data.
Whereas for the anxiety classification, Takagi-Sugeno-Kang (TSK) fuzzy classifier had given the best results of 76.68% and 75.23% in identifying anxiety from the DEAP dataset and homemade dataset respectively. It was then followed by Radial Basis Functions Neural Networks (RBFNN), SVM, Random Forest and k-NN classifiers correspondingly.
On the other hand, Joshi and Ghongade (2021) proposed using EEG signals from DEAP database to develop an emotion detection system. Features such as differential entropy (DfE), rational asymmetry (RASM) and Linear Formulation of Differential Entropy (LF-DfE) were extracted from the EEG data. Then, artificial neural network (ANN) – MLP and recurrent neural network (RNN) – bidirectional LSTM (BiLSTM) classifiers were implemented and the performance of the features were compared with respect to each classifier.
As a result, the highest accuracies of 76% and 75.5% were obtained in classifying both levels of arousal and valence using LF-DfE features in the BiLSTM network. Therefore, anxiety which is generally indicated by LVHA can be detected using the emotion detection system proposed by Joshi and Ghongade (2021).
2.2.2 Stop Signal Task (SST) Personal Dataset
In the research done by Wang et al. (2019), 81 subjects were asked to participate in Stop Signal Task (SST) for anxiety induction purpose whereby conflict processes were stimulated in the brain. State-Trait Anxiety Inventory (STAI) score for each participant was evaluated. Generally, the STAI score is evaluated based on a 4-point scale with a total of 20 questions portraying a set of stressful situations. The total score of this report is 80 and high anxiety is often indicated by a higher STAI score ranging from 45 to 80. EEG signals were then collected at that time and CNN architectures were used to predict the occurrence of anxiety. In their work, 3D CNN was proposed with the aim to improve the performance of the system using 2D CNN. 3D CNN architecture was designed such that the EEG input was reshaped into a 3D tensor as shown in Figure 2.7.
Figure 2.7: 3D tensor of EEG input (Wang et al., 2019)
The combination of temporal and spatial features in the 3D CNN architecture had proved to enhance the accurateness in identifying anxiety. Hence, Layer-wise Relevance Propagation (LRP) was applied in the CNN network whereby the predicted results were propagated backward in the network to verify the correlated EEG components which are related to anxiety. With that, LRP was able to project the evaluated STAI score back to the input and generate brain topology heat maps as shown in Figure 2.8 representing the criticalness of anxious traits on different regions of the brain exhibited by the participants.
Figure 2.8: Brain topology heat maps based on STAI scores (Wang et al., 2019)
According to Wang et al. (2019), Goal-Conflict-Specific-Rhythmicity (GCSR) is an early biomarker used for anxiety detection. It is usually measured during the SST using F8 electrode at the right frontal cortex of the brain. Since the frontal lobe is the part of the brain that helps regulate one’s emotions and GCSR is positively correlated to the STAI score, hence this study was conducted to examine the variation of GCSR at F8 position based on STAI score. From the third row in Figure 2.8, it was observed that participants 3, 4 and 13 which are of high anxiety (higher STAI score) exhibited stronger GCSR values (more positive, indicated by red colour) at their right frontal regions respectively. Whereas for participants 1, 11 and 12 in the first row of Figure 2.8 which are of no or low anxiety, the extracted GCSR values from EEG signals collected at F8 position were less or negatively correlated to anxiety (indicated by green or blue colour). Hence, different anxious states can be detected based on the GCSR values calculated.
2.2.3 Changhai Hospital Personal Dataset
According to Xie et al. (2020), CNN algorithm is useful in handling data and performing various tasks which include EEG classification. Therefore, prior to studying anxiety and depression, the EEG dataset used in this work was obtained from Changhai Hospital in Shanghai. This dataset reflects the EEG signals of the patients suffering from anxiety, whereas the EEG data for normal people were collected separately.
The EEG data were pre-processed using EEGLAB to remove artifacts and then passed through Butterworth 6th order filter for further filtering. Next, Phase Lag Index (PLI) was applied to calculate the functional connectivity matrix of the pre-processed data. After that, this matrix was used to train different deep learning models such as Deep Belief Network (DBN), LDA and CNN for anxiety classification and the results were compared. The highest accuracy of 67.67% was obtained using Brain Networks (functional connectivity features) with CNN classifier.
2.2.4 Healthy Brain Network (HBN) Public Dataset
The EEG data collected in HBN were mainly from children and adolescents. 92 participants were involved in this experiment and the anxiety levels in them were rated using the Screen for Child Anxiety Related Disorders (SCARED) scale. SCARED questionnaire has a total of 41 items with a 3-point scale each (see Figure A3 & A4, Appendix A). A total score of 25 and above indicates the possibility of a child experiencing a form of anxiety disorder. According to Zhang et al. (2020), the aim of his study is to analyse the correlation of eye movements of the subjects and EEG signals for anxiety detection during resting state.
The EEG signals acquired were fed into a filter and then pre-processed using PCA algorithm. Augmented Lagrange Multipliers method (ALM) was also applied here for problem reduction purpose and to remove the artifacts found in the signals.
After that, power spectrum features (PSD) were extracted from the EEG data using TEAP toolbox available in MATLAB or Octave software whereas features from the eye movements were captured using an interval algorithm (Zhang et al., 2020).
From the analytical results, the fusion of both EEG and eye movement features using Group Sparse Canonical Correlation Analysis (GSCCA) had proved to increase the accuracy of classifying anxiety compared to unimodal EEG. Besides that, gamma band in EEG was found to have the highest accuracy (k-NN: 78.57%; SVM: 82.70%) in detecting anxiety compared to other frequency bands in the fusion model.
2.2.5 Social Anxiety Disorder (SAD) Study using Deep Learning
In the research done by Al-Ezzi et al. (2021), deep learning models were used for SAD identification and to classify the seriousness of SAD based on the neural brain information extracted in the Default Mode Network (DMN).
89 subjects were asked to evaluate their self-assessment of Social Interaction Anxiety Scale (SIAS). A SIAS questionnaire comprises 20 items with a 4-point scale each (see Figure A5, Appendix A). Usually, the total score of 60 and above indicates a severe SAD risk. After that, the EEG data of the participants were recorded during resting state as DMN is believed to be active at this condition. For signal filtering, FIR band-pass filter was applied to define the frequency range whereas the artifacts removal was carried out using BESA software.
According to Al-Ezzi et al. (2021), effective connectivity features were extracted from the EEG signals and quantified using Partial Directed Coherence (PDC) algorithm to predict SAD. Apart from that, deep learning models such as Long short- term memory (LSTM) and CNN were implemented to classify SAD. As a result, the highest accuracy was obtained for all levels of severity with an average percentage of 93% using the combination of both CNN and LSTM models.
2.2.6 Trait Anxiety Detection using MUSE EEG Headband
A study about trait anxiety was presented by Arsalan and Majid (2021) based on the EEG recordings of 65 subjects during resting condition. This dataset is publicly available to promote future improvements. In general, trait anxiety is a type of anxiety that shows part of a person’s personality and it does not need to be intentionally induced by certain stressful conditions or danger. For instance, trait anxiety can be described as more of a personal thought or feeling that arises due to some situations faced in daily routines.
In this experimental setup, MUSE EEG headband with a built-in noise removal circuit was used. Then, analysis of variance (ANOVA) was applied on the raw EEG data by studying the strength of the energy at different intervals of frequencies (power spectral density) to select a significant set of channels before proceeding to feature extraction.
The features extracted from the EEG signals include maximum absolute value, signal energy, signal sum, etc. (Arsalan & Majid, 2021). Wrapper method was then applied for feature selection such that the features which can produce the highest classification accuracy based on the classifiers implemented were chosen.
In this work, several classification techniques had been proposed and compared.
The training of the models was done using Weka 3.8 tool. As a result, Random Forest classifier yielded the highest accuracy of 87.69% and 83.07% for both two-class and three-class anxiety classification respectively, followed by Multilayer Perceptron (MLP) and Logistic Regression (LR). Besides that, the smaller feature vector length proposed also contributed to a better anxiety classification accuracy.
Table 2.2: Summary of previous research on EEG-based anxiety detection using different datasets
Researcher Type of anxiety study/ dataset
Feature extraction
Classification method
Best accuracy result Baghdadi,
et al.
DASPS RMS k-NN 67.00%
All features SSAE 83.50%
Shikha, et al.
DASPS Time domain
features
Decision Tree
70.25%
All features SSAE + RFECV
83.93%
Meng &
Zhang
DEAP All features TSK fuzzy system
76.68%
Joshi &
Ghongade
DEAP LF-DfE BiLSTM 76.00%
Wang et al.
SST Spatial and
temporal features
3D-CNN -
Xie et al. Changhai Hospital Functional connectivity
CNN 67.67%
Zhang et al.
HBN PSD + eye
movement features
SVM 82.70%
Al-Ezzi et al.
SAD Effective
connectivity
CNN + LSTM
93.00 %
Arsalan &
Majid
Trait anxiety Selected features using
Wrapper method
Random Forest
87.69%