Real-Time EEG-Based Emotion Recognition Using PSO-Optimized Support Vector

Machines for Enhanced Classification of Arousal and Valence

Niharika Gudikandulaa , Ravichander Janapatib , Rakesh Senguptac

a SR University, School of Engineering Ananthasagar, Hasanparthy, Warngal, 506371, Telangana, India
b SR University, School of Engineering Ananthasagar, Hasanparthy, Warangal, 506371, Telangana, India
c SR University, Center of Creative Cognition Ananthasagar, Hasanparthy, Warangal, 506371, Telangana, India

Abstract
People feel different emotions in different situations and at different stages of their daily practice: bored, calm, horrified and,
of course, funny. this case, an efficient function to recognize emotion is crucial for EEG data in order to qualify for depicting
emotions. The identification of the psychophysiological state of the people has been a major concern in many fields including
the adaptation ofSocial life to disabled people. In recent years, different physiological cues have been incorporated in emotion
recognition research because the requiredsuccess is not very high in those trials which are performed through basic methodologies
such as postures and facial gestures. evaluated utilizing the GAMEEMO dataset. For the purpose of this study, two emotions
are considered for the classification; Positive emotion and Negative emotion. PSObased deep learning method that can identify
human emotions by analyzing Electroencephalogram (EEG) signal is provided. underwent some preprocessing techniques like
bandpass filter and wavelet transform which is more sensitive to time frequency changes. Then feature extraction using Fast Fourier
Transform from which vectors were extracted. Last, the time frequency features extracted were used on one of the frequently used
machine learning technique called SUPPORT VECTOR MACHINES (SVM) having particle swam optimization classifier forThe
study employed the proposed method in GAMEEMO dataset to evaluate its effectiveness. This study considers two emotions,
namely‘Arousal’ and ‘Valence’ for classification Accuracy of LANV LAPV, HANV, HAPV . These experimental results represent
the complete works of the current study, proving the significance and efficiency of the proposed 99.37detection accuracy in the
proposed combination of SVM with PSO classifiers.
Keywords: EEG, GAMEEMO dataset, Particle Swarm Optimization, Support Vector Machine, Feature Extraction, Data
Augmentation, Emotion Recognition,

1. Introduction GSR, EMG, RT, and EEG, have been employed in the detection
of emotions in identifying a limited range of emotions. How-
Emotions are fundamental to changes in behavior and per- ever, it is evident that EEG data provide high temporal reso-
formance that signify learning, reflecting a sophisticated level lution, meaning-rich information that may be accessed using
of intelligence. To fully define emotion, it is essential to con- inexpensive, portable EEG equipment. Thus EEG is a commu-
sider three key aspects: The implicit or cognitive sense of emo- nication between the human mind and an external instrument
tion, the neurological and physiological mechanisms that un- may be created through an EEG-based system, which enables
derlie it, and the visible behavioral patterns, and the observ- the reading of physiological signals and the interpretation of
able expressive patterns, especially those manifested in facial specific elements of the individual’s mental state. These emo-
expressions [1]. Emotions are classified according to three tions can be captured during synaptic neuronal dendrite excita-
main theories: Plutchik’s, Ekman’s, and the James-Lange the- tions, BCI technology places electrodes on top of the head to
ory. Plutchik’s theory divides emotions into basic (e.g., joy, measure how the nervous system’s electricity reacts to current
fear, anger) and secondary (e.g., aggression, submission) cate- flow. EEG wearables with emotion identification might lead
gories. Ekman’s discrete model identifies The six primary feel- to breakthroughs in electronic learning, playing games, medi-
ings are sorrow, joy, annoyance, distaste, anxiety, and amaze- cal care, human-computer interface, and automobiles. The field
ment. The James-Lange theory categorizes emotions using a of Brain-Computer Interaction (BCI) originated in the 1970s,
dual-dimensional arousal valence paradigm, in which valence with early research focusing on the cortices of monkeys and
varies from negatively (unpleasant) to optimistic (pleasurable), rats at the University of Washington School of Medicine. Ini-
and arousal runs from minimal (passive) to excessive (active) tially, BCI research focused on medical diagnosis and brain
[2]. These emotions can be expressed by the human emotions function examination. Technological advancements have ex-
Gestures, speech, body language and physiological signals. But panded this scope to include BCI’s artificial intelligence. of
emotions can be accurately interpreted by physiological sig- facial expressions and eye movement, presence of AI, emotion
nals. Several physiological signals to be considered, like as recognition, and limb/hand movement detection. It seems that

1
 Figure 1: Emotion Model

thanks to the developments of cost-effective microelectronics-

Some of these microcircuits allow BCI users to control various
activities through performing elaborate tasks. last generation
for self-decision making and better algorithmssuch as Machine
Learning (ML) has inspired researchers to use the ML in the
BCI SYSTEMS for enhancing and effective task performance.
ML is a crucial branch of statistics and computer science, utiliz-
ing computers to perform tasks without explicit programming.
It studies computer architectures and algorithms, learning from
observed facts. ML techniques are widely used in monitoring,
detection, classification, and other tasks. ML techniques play a
predominant role in data analysis in BCI systems, allowing use-
ful knowledge to be automatically obtained and applied to the
target task. As existing BCI systems are insufficient for learn-
ing, understanding, and interpreting complex brain activities,
the use of ML technology with BCIs can create successful inter-
actions between humans and machines can result from human Figure 2: All Features of EEG signal
emotions, and computer systems may eventually use artificial
emotional intelligence[3]. The rise in electronic device usage
has led to increased time spent on social media, online gaming,
and other digital activities. Unlike passive activities like watch-
ing films, playing computer games provides a more immersive signals to uncover patterns that convey emotional information,
and interactive experience, where individuals engage directly which classifiers then use to distinguish between emotional
with the environment and adopt the roles of game characters. states. The SVM-PSO approach, applied to EEG data from 28
This active participation has a significant impact on emotions, volunteers exposed to emotion-evoking visuals, demonstrates
making it crucial to study emotion recognition in gaming con- superior emotion recognition, classifying up to four distinct
texts. Understanding emotions during gameplay can offer in- emotions and advancing real-world emotion recognition sys-
sights into psychological well-being and enhance game design tems.
for better user experiences. This active participation signifi-
cantly impacts emotions, making it crucial to study emotion This paper is organized as follows: Section II offers back-
recognition in gaming contexts. To achieve this, we used four ground information on the theme of emotion classification and
types of games, preprocessed the data using a bandpass filter outlines the proposed analysis. Section III details the exper-
with wavelet denoising, and applied time-frequency feature ex- imental dataset used, along with the preprocessing steps and
traction methods. time-domain feature extraction methods employed. In Section
Toward this direction, the significance of using Particle IV, the performance of the SVM-PSO approach is presented
Swarm Optimization (PSO) for a user-independent EEG-based through experimental results, highlighting its effectiveness in
emotion recognition system. PSO, inspired by natural phenom- addressing EEG-based classification challenges and comparing
ena, efficiently optimizes problem-solving by simulating social it with other existing methods. Finally, Section V concludes the
animal behavior. In emotion recognition, PSO analyzes EEG paper.
2
 Table 1: Accuracy Comparison With Existing Emotion Recognition Methods Using GAMEEMO Dataset
Dataset Methods used Emotions Accuracy

[4] Multivariate multiscale modified-distribution en- Valence and Arousal 95.73% accuracy for va-
tropy (MM-mDistEn), combined with ANN lence and 96.78% for
model arousal 95.73% and 96.78%
[5] GoogLeNet-based deep learning method using ‘Positive’ and ‘Nega- 98.78% accuracy with
EEG signals, through Continuous Wavelet Trans- tive’ emotions SVM, 98.53% with k-NN,
form and 98.41% with ELM
98.78%
[6] PrimePatNet87, a hand-crafted network using arousal, valence 99% accuracy
prime pattern and TQWT techniques
[7] Spectral entropy, classifying with bidirectional Valence and Arousal 76.91% accuracy
LSTM
[8] Projection dictionary learning with the fast PDPL positive, neutral, and 69.89% accuracy
technique using a genetic algorithm negative
[9] General non-subject-based model and applies LAPV HAPV LANV 98.38% accuracy
Capsule Networks for binary and multi-class HANV
recognition
[10] LEDPat-Net19, an advanced emotion recognition arousal,dominance 99.29% accuracy
model with TQWT multilevel feature generation and valance
network
Proposed Bandpass filtering using wavelet transform with LAPV HAPV LANV 99.37 accuracy%
Method swarm optimization using SVM HANV

2. Literature review
Recent studies have focused on extracting meaningful prop-
erties from EEG signals for emotion recognition. These studies
have proposed various algorithms that leverage different char-
acteristics of EEG signals to enhance emotion recognition. Au-
thor propose an emotion recognition system using multichan-
nel EEG and a novel entropy measure, multivariate multiscale
modified-distribution entropy (MM-mDistEn), combined with
an ANN model. Authors approach achieved 95.73% accuracy
for valence and 96.78% for arousal on the GAMEEMO dataset,
and 92.57% and 80.23% on the DEAP dataset, respectively [4].

This study proposes a GoogLeNet-based deep learning

method for emotion recognition using EEG signals, converted
to scalograms via Continuous Wavelet Transform. Testing
on the GAMEEMO dataset achieved 98.78% accuracy with
SVM, 98.53% with k-NN, and 98.41% with ELM, outperform-
ing existing methods by 21.85% [5]. To address the limita-
tions of deep models in EEG emotion classification, we pro-
pose PrimePatNet87, a hand-crafted network using prime pat-
tern and TQWT techniques. It achieved over 99% accuracy
on DEAP, DREAMER, and GAMEEMO datasets, demonstrat-
ing high performance with SVM and iterative majority vot-
ing [6]. This study analyzes emotions using EEG signals
from the GAMEEMO dataset. It involves calculating spectral
entropy, classifying with bidirectional LSTM, and evaluating Figure 3: Emotion Model
performance with accuracy, sensitivity, specificity, and ROC
curve. The method achieved 76.91% accuracy and a 90% ROC
value[7]. To address subject-specific variability in EEG emo-
3
 Figure 4: Workflow of EEG-based emotion recognition using the GAMEEMO database

tion recognition, we employed projection dictionary learning LANV and HANV zones (representing negative emotions)
with the fast PDPL technique and optimized parameters us- were examined separately. The process involved averaging
ing a genetic algorithm. Our method achieved 69.89% accu- the EEG signals within the positive and negative emotion cat-
racy on SEED, 24.11% on MPED, and 64.34% for two-class egories to create a unified EEG signal for each category. A
GAMEEMO[8]. band-pass filter is applied to the EEG signals to isolate the de-
Emotion recognition using EEG signals, influenced by aural sired frequency bands,followed by wavelet denoising using the
and visual stimuli, is challenging due to internal psychological ’db4’ wavelet and an 11-level decomposition. Wavelet trans-
emotions. This study introduces a general non-subject-based form is widely utilized for representing signals in the time-
model and applies Capsule Networks for binary and multi-class frequency domain. It decomposes a time-domain signal into
recognition. The method achieved a 10% improvement in ac- wavelet coefficients using a mother wavelet function, achieved
curacy over existing studies using the GAMEEMO dataset[9]. through shifting and dilation of the mother wavelet. When a
This work presents LEDPatNet19, an advanced emotion recog- noisy version of a signal is available, the challenge is to re-
nition model using EEG signals. It integrates a multilevel fea- store the original information. Traditional wavelet denoising
ture generation network with TQWT, statistical, and nonlinear modifies the wavelet coefficients based on their local proper-
feature extraction. Evaluated on GAMEEMO and DREAMER ties and then inverts the transformation to obtain a clean signal.
datasets, it achieved accuracies of 99.29% and 94.58% for dif- Wavelet denoising is a popular method for signal processing
ferent emotion categories, demonstrating its effectiveness[10]. and involves three key steps. First, the input signal is decom-
posed into wavelet coefficients. Next, these coefficients are ad-
justed based on a thresholding technique. Finally, the modified
3. Preprocessing
coefficients are used in the inverse transform to reconstruct the
Preprocessing of EEG signals is essential for noise removal signal, free from noise. In our research, we demonstrated effec-
and can be achieved through methods like band-pass filtering, tive wavelet-based EEG denoising using universal and statisti-
wavelet transform and Fourier transform. These techniques en- cal threshold functions.
hance the signal-to-noise ratio (SNR), reduce computational In our study EEG signals were filtered using a 5th-order
cost, and improve the accuracy of emotions. In this study, band-pass filter with cutoff frequencies of 0.05 Hz and 45 Hz.
the GAMEEMO dataset, consisting of EEG signals collected The filtering process can be represented as:
at a 128 Hz sampling rate, was filtered and denoised to remove
N−1
noise and artifacts. Four frequency bands were targeted: 8-8.57 X
Hz, 9.5-10.5 Hz, 11.5-12.5 Hz, and 14.5-15.5 Hz, correspond- y(t) = h[n] · x(t − n) (1)
n=0
ing to different emotional states.
In the GAMEEEMO dataset analysis for both binary and y(t) is the filtered signal. x(t) is the original signal. h[n] is the
multi-class classification, EEG signals from the LAPV and impulse response of the band-pass filter. N is the filter order.
HAPV zones (representing positive emotions) and from the Wavelet denoising is used to remove artifacts from the EEG
4
 ity reduction of the EEG data while preserving the essential
information contained in the EEG signals. FE was conducted
on the cleaned EEG signals by segmenting them into epochs
of 16 seconds each. Features in both the time and frequency
domain were retrieved. For time-domain features, our study
used mean, variance, skew, and kurtosis; for frequency-domain
features, Power Spectral Density (PSD) and Fast Fourier Trans-
form (FFT) were applied. These features capture the essential
characteristics of the EEG signals for subsequent classification.
The Fast Fourier Transform (FFT) is used to extract fre-
quency domain features:
N−1
2π f n
X
X(k) = x(n) · e− j N (4)
n=0
X( f ) signifies the frequency characteristics of the signal. x[n]
is the signal representation of time-domain. N is the number of
points in the EEG signal.
Welch’s PSD estimation method:
K
Figure 5: All Features of EEG signal 1 X
P xx ( f ) = |Xk ( f )|2 (5)
K k=1
signals. The denoising process typically involves thresholding P xx ( f ) is the power spectral density. Xk ( f ) is the Fourier
the wavelet coefficients can represented as: transform of the k-th segment of the signal. K is the number
of segments.
c j if |c j | > λ
(
ĉ j = (2) Time-Domain Features
0 if |c j | ≤ λ
N
c j are the wavelet coe f f icients.λ is the threshold level. ĉ j 1 X
µ= xi (6)
are the denoised wavelet coefficients. N i=1
N
An EEG epoch was defined by segmenting the raw EEG 1 X
σ2 = (xi − µ)2 (7)
signal into 4-second intervals for each channel, with the sub- N − 1 i=1
sequent epoch obtained by shifting the window by 1 second.
N
After segmentation, features were computed independently for 1 X (xi − µ)3
γ1 = (8)
each epoch and channel, focusing on commonly used features. N i=1 σ3
Time-domain features (TDFs) were calculated directly from
N
the raw EEG signal, while frequency domain features were 1 X (xi − µ)4
extracted from the power spectral density (PSD) and time- γ2 = −3 (9)
N i=1 σ4
frequency domain features (TFDFs) were computed from dis-
crete wavelet transform (DWT) coefficients[11]. With respect to segmentation, Data augmentation (DA) refers
The challenge of EEG channel selection remains an active to the process of creating new samples to supplement an exist-
area of research, as using multiple channels can sometimes be ing data set by modifying the existing samples. This approach
redundant. In our study we are focusing on channel selec- might enhance the classification accuracy and its stability, espe-
tion, spatial averaging of features from the left and right hemi- cially when dealing with EEG data. The main goal of DA is to
spheres. Epochs are created by dividing the EEG signal into decrease the sensitive to such transformations and make classi-
fixed-length segments. Each segment (epoch) can be repre- fiers less biased by making them more invariant to such trans-
sented as: formations and therefore allowing the model to perform better
on unseen datasets. To improve the quality and stability for the
Epochk = {xt0 , xt1 , . . . , xtn } (3) heterogeneity to work smoothly for the training dataset, data
augmentation strategies were used. Gaussian noise was added
k denotes the epoch number.
to the EEG signals to simulate variability, and time-shifting and
t0 , t1 , . . . , tn are the time indices within the segment.
amplitude scaling were applied to introduce temporal and am-
plitude variations. This augmentation process quadrupled the
4. Feature Extraction dataset size, providing a richer training set for the classifier.
Adding Gaussian noise to the signal:
EEG signals are inherently complex and rich in information.
In feature extraction,The intention is to achieve dimensional- xaug (t) = x(t) + N(0, σ2 ) (10)
5
Figure 6: Caption for Figure 1

Figure 7: Caption for Figure 2

6
 Algorithm 1: Pseudo-Code of SVM-PSO Algorithm
1. Initialize the swarm of particles S and corresponding velocity vectors.
2. For each iteration t = 1 to max iterations do:
3. For each particle i = 1 to S (total particles) do:
4. Evaluate the fitness of each particle based on SVM classification accuracy on a validation set.
5. For each dimension d = 1 to the number of parameters (2 for C and γ) do:
6. Update the velocity of each particle i using:
The particle’s own best position (pbest).
The global best position (gbest) in the swarm.
7. Update the position of particle i using the updated velocity.
8. End For (dimension update loop)
9. If the fitness of particle i is better than its pbest, update pbest.
10. End For (particle loop)
11. If any particle’s fitness is better than the global best (gbest), update gbest.
12. End For (iteration loop)
13. Train the final SVM model using the best hyperparameters C and γ obtained from gbest.
14. Terminate the process if the stopping criterion is met .

Iteration f-count Best Stall

f(x) Mean f(x) Iterations
0 50 0.02721 0.1216 0
1 100 0.02471 0.3624 0
2 150 0.02283 0.106 0
3 200 0.02127 0.09142 0
4 250 0.01783 0.09686 1
5 300 0.01783 0.06921 1
6 350 0.01783 0.05688 2
7 400 0.01783 0.06898 3
8 450 0.01783 0.04266 4
9 500 0.01783 0.06975 5
10 550 0.01783 0.04315 6
11 600 0.01783 0.0291 7
12 650 0.01783 0.03437 8
13 700 0.01658 0.02783 1
14 750 0.01658 0.02879 1
15 800 0.01658 0.04934 2
16 850 0.01658 0.04337 3
17 900 0.01658 0.02934 4
18 950 0.01658 0.03055 5 Figure 10: Electrode Placement of 14-channel EMOTIV EPOC+ of 10-20 sys-
tem
19 1000 0.01658 0.0299 6
20 1050 0.01658 0.03025 7
21 1100 0.01658 0.03205 8
22 1150 0.01658 0.02958 9 xaug (t) = x(t + ∆t) (11)
23 1200 0.01658 0.02917 10
∆t is the shift amount.
24 1250 0.01658 0.04261 11
N(0, σ2 ) represents Gaussian noise with mean 0 and variance
25 1300 0.01658 0.04261 12
σ .
2
26 1350 0.01658 0.04913 13
The wavelet transform, which uses fixed-size windows for
27 1400 0.01658 0.04333 14
frequency analysis, employs variable-size windows for time-
28 1450 0.01658 0.04997 15
frequency decomposition. It uses shorter windows for high
29 1500 0.01658 0.02953 16
frequencies and longer windows for low frequencies, matching
30 1550 0.01658 0.03032 17
real-world signal characteristics better. The orthogonal discrete
31 1600 0.01658 0.03047 18
wavelet transform is generally not time shift invariant; different
32 1650 0.01658 0.02941 19
shifts in the input can produce different de-compositions, po-
tentially limiting its effectiveness in some applications. Shifting
the signal by a certain number of samples:
7
 We examined the amplitude factor c, which parametrizes the 6. Classification
spatial scale ratios. Our findings indicated that variations in
the amplitude scaling can significantly influence the observed The features extracted in the third stage of the proposed
differences in EEG spectra. Specifically, we discovered that method were classified into emotion categories. To determine
when the amplitude factor was adjusted, it impacted how the the most effective machine learning technique for achieving op-
spatial scale ratios interacted with the frequency components, timal results, we evaluated three popular classification meth-
revealing patterns consistent with real data reported in previous ods: k-Nearest Neighbors (k-NN), Convolutional Neural Net-
studies. This highlights the importance of amplitude scaling in works (CNN), and Support Vector Machines (SVM) with Par-
accurately modeling and interpreting EEG signals. Scaling the ticle Swarm Optimization (PSO).
amplitude of the signal: The SVM is a powerful classifier known for its simplicity
and effectiveness. It works by finding hyperplanes that maxi-
xaug (t) = α · x(t) (12) mize the margin between different classes, enhancing general-
ization. The success of SVM in classification tasks is closely
α is the scaling factor. tied to the choice of hyperparameters. PSO provides an auto-
mated and efficient way to tune these parameters, which is cru-
cial for achieving high accuracy and reliability in classification
tasks, especially with complex datasets like EEG signals. EEG
5. GAMEEMO Dataset data is highly nonlinear and complex. SVMs with RBF kernels
are well-suited for such tasks, but their effectiveness depends on
the correct setting of parameters. PSO helps find these settings,
The GAMEEMO dataset[12] is a comprehensive emotion ensuring that the SVM can handle the nonlinearity in the data
recognition database based on EEG signals, created by Alakus effectively. Traditional optimization techniques like grid search
et al. at Firat University, Department of Software Engineer- or manual tuning are time-consuming and may not yield the
ing. In a game, emotions are primarily driven by the quad- best results. PSO automates this process, exploring a wider pa-
rants of the Arousal-Valence model. Negative emotions like rameter space more efficiently. PSO is a optimization strategy
’angry’ and ’bored’ are on the left side of the valence dimen- derived from social life of birds that is useful in managing pop-
sion, while positive emotions like ’happy’ and ’calm’ are on the ulation based systems. In this study, PSO was used to fine-tune
right. The upper arousal dimension reflects high emotional or the SVM parameters, specifically the regularization parameter
behavioral arousal, and the lower part indicates apathy. There- C and the kernel scale σ. The objective function minimized by
fore, emotions are classified into four quadrants: bored, calm, PSO was the classification error obtained through 5-fold cross-
angry, and happy. As players interact with the game, they en- validation. The optimized SVM classifier was trained on the
gage in high mental activity, processing images, forming narra- augmented dataset and evaluated on a separate test set. The
tives, and evaluating situations. Initially, emotions are not fixed model’s performance was assessed using accuracy, precision,
to any specific quadrant. It involves 28 participants, aged be- recall, and F1 score metrics, ensuring a comprehensive analysis
tween 20 and 27 years, who were subjected to four computer of its effectiveness in various classification scenarios.
games designed to elicit four basic emotions: boredom, calm- PSO is a population of optimization solution derived from
ness, horror, and amusement. Each game session lasted for 5 the behavior of a swarm of particles. Each particle contains a
minutes, resulting in a total of 20 minutes of EEG data per par- potential solution in the search space and changes its position
ticipant. The EEG signals were recorded using a 14-channel according to the previous best it has found and the best of all
EMOTIV EPOC+ Mobile EEG device, capturing 38,252 sam- particles found. PSO seeks to find the optimal values of hy-
ples from each participant per game. Based on our findings, perparameters C and γ for SVM by minimizing classification
it appears that electrodes located in the frontal (Af3-Af4), oc- error. Mathematically, the velocity of a particle i at time t + 1 is
cipital (O1-O2), and both hemispheres of the brain might po- updated using the following equation:
tentially aid with categorization. with the temporal (T7-T8) re-
gions. We note that trends could be ultimately discovered in
id + xid = vid + xid + c1 r1 (pid − xid ) + c2 r2 (pgd − xid ) (13)
vt+1 t+1 t t t t t t
any frequency range; yet, the Alpha (8-Bands designated as 13
Hz, Delta (1-4 Hz), and Theta (4-8 Hz), in that sequence, are
most strongly associated with different types of emotions. In the Particle Swarm Optimization (PSO) algorithm, vi(d) in-
Additionally, each subject’s ratings of the games are provided dicates the velocity of the particle i in the d-th dimension at
in .pdf format, using the Self-Assessment Manikin (SAM) form some iteration t. The position of each particle depends on per-
to assess arousal and valence levels. The games are labeled sonal experience and that of the other particles in its swarm;
as G1, G2, G3, and G4, representing different emotional stim- pbesti is the best position encountered by particle i, while gbest
uli. The GAMEEMO dataset offers an alternative to traditional is the best position encountered by any particle in the swarm.
EEG data, showcasing the capabilities of portable and wear- Specifically, two acceleration constants are included in the al-
able EEG devices in emotion recognition studies. This dataset gorithm; c1 that refers to the tendency of cognitive learning, and
is valuable for researchers aiming to explore and improve emo- c2 that is derived from the concept of social learning. These
tion recognition techniques using EEG data. constants are then multiplied by r1 and r2 , random numbers in
8
 [t].5
Figure 11: PSO+SVM with 99.37% Accuracy Figure 13: KNN with 71.36% Accuracy

[t].5
Figure 12: CNN with 62.31% Accuracy

the specified range of the interval [0, 1]. Besides, w determines

the impact of previous velocities in the balance between explo-
ration and exploitation phases, during the search.
The position of the particle i is updated using:

xi(t+1) = xi(t) + v(t+1)

i

This iterative process allows particles to explore the search

space and converge towards the global optimal solution.
Furthermore, we implemented k-NN and SVM classifiers to
examine emotions derived from EEG data. Next, we used the
same approach to alternative classifiers cnn and knn.Initial pre-
processing of the EEG data was followed by feature reduction
using the fft feature extraction technique. Classification of the
reduced features was performed using k-NN aCNN nd SVM.
Our KNN yielded a score of 71.36%, whereas our CNN score
was 62.31% and PSO+SVM achieved an accuracy of 99.37
The assessment of the classifier’s performance involves such
aspects as performance indicators, estimating errors, and carry-
ing out the significance testing . Performance measures and Figure 14: All Features of EEG signal
error estimation determine the fulfillment rate of the classi-
9
 accuracy by the lower values shown in Table 2. The Kappa
factor(KF) measures the agreement between observed and ex-
pected classifications while adjusting for the agreement that
could occur by chance. Cohen introduced the Kappa factor as a
metric to quantify the level of agreement between two raters or
classifiers.
P0 − Pe
K= (19)
1 − Pe
P0 is the observed agreement denoted,
TN + TP
P0 = (20)
T N + T P + FP + FN
Expected agreement is categorized based on probability by
using the notation P. The sum of all possible probabilities of a
random agreement is the total random agreement probability.
they respond with either ‘Yes’ or ‘No’, that is,

Pe = PY ES + PNO (21)

Figure 15: Comparison of performance metrics across different articles and

FP + T P FN + T P
datasets. PY ES = ∗ (22)
T N + T P + FP + FN T N + T P + FP + FN

fier function.Of all the performance measures, the most recom-

mended is the confusion matrix in combination with accuracy FN + T N FP + T N
PNO = ∗ (23)
and the error rating, which can be directly obtained from the T N + T P + FP + FN T N + T P + FP + FN
confusion matrix besides the other measures like recall, speci-
ficity, precision, RMSE, Cohen’s kappa (k) and F-measure. 7. Results and Discussion
In the process of determining the effectiveness of implemen-
tation of the proposed method for detecting emotions of a hu- The proposed PSO-SVM model demonstrated superior per-
man based on the analysis of the EEG signal, which are called formance in classifying EEG signals compared to traditional
True Positives or TP if they are accurately identified. whereby methods. The accuracy achieved was significantly higher, and
those that are wrongly classified as positive are referred to as the confusion matrix revealed improved classification across all
False. Likewise, inaccurate negative affectivity implies that one emotional states. The precision, recall, and F1 scores, in ad-
is experiencing negative emotions that are not warranted. clas- dition also confirmed the stability of the model, and as for the
sified as negatives are True Negatives (TN), andthat are actually ability to process. This is due to the complex and noisy nature
positive but are classified as negative are referred to as False of EEG data.
Negatives (FN).
TP + TN 8. Conclusion
Accuracy = (15)
TP + TN + FP + FN
The True Positives (TP) Accurate Negative Analysis(TN) In- This study presents a comprehensive approach to EEG sig-
correct Positive Results(FP) True Negative Results(FP) nal classification using the GAMEEMO dataset. By integrat-
ing advanced preprocessing techniques, feature extraction, data
TP
Precision = (16) augmentation, and PSO for SVM optimization, the proposed
TP + FP methodology achieves high accuracy in emotion recognition
TP tasks. The success of the PSO-SVM model underscores the po-
Recall = (17) tential of optimization techniques in enhancing the performance
TP + FN
of machine learning models in BCI applications.
Precision × Recall
F1 score = 2 × (18)
Precision + Recall 9. Future Work
Root Mean Square Error (RMSE) are often utilized in re-
gression analysis, is the square root of the arithmetic mean of Future research could explore the application of deep learn-
the squared difference between the predicted value and the ob- ing techniques in conjunction with PSO for EEG signal classifi-
served value. Is a measure of the accuracy of a model by show- cation. Additionally, the effectiveness of the proposed method-
ing how close the predicted outcomes of a specific model are ology could be tested on other EEG datasets to validate its gen-
to the actual outcomes. RMSE gives an indication of the best eralizability across different BCI applications
10
 Figure 16: Figure A Figure 19: Figure C

Figure 20: Figure D

Figure 17: Figure B
Figure 21: This is another long description showing multiple figures.
Figure 18: This is a description that’s too long to fit on one line.

Acknowledgements [6] A. Dogan, M. Akay, P. D. Barua, M. Baygin, S. Dogan, T. Tuncer, A. H.

Dogru, U. R. Acharya, Primepatnet87: prime pattern and tunable q-factor
Thanks to ... wavelet transform techniques for automated accurate eeg emotion recog-
nition, Computers in Biology and Medicine 138 (2021) 104867.
[7] T. Alakus, I. Turkoglu, Emotion recognition with deep learning using
gameemo data set, Electronics Letters 56 (25) (2020) 1364–1367.
References [8] J. Su, J. Zhu, T. Song, H. Chang, Subject-independent eeg emotion recog-
nition based on genetically optimized projection dictionary pair learning,
[1] C. E. Izard, Human emotions, Springer Science & Business Media, 2013. Brain Sciences 13 (7) (2023) 977.
[2] A. Abdulrahman, M. Baykara, A comprehensive review for emotion de- [9] S. Toraman, Ö. O. Dursun, Gameemo-capsnet: Emotion recognition from
tection based on eeg signals: Challenges, applications, and open issues., single-channel eeg signals using the 1d capsule networks., Traitement du
Traitement du Signal 38 (4) (2021). Signal 38 (6) (2021).
[3] S. Rasheed, A review of the role of machine learning techniques towards [10] T. Tuncer, S. Dogan, A. Subasi, Ledpatnet19: Automated emotion recog-
brain–computer interface applications, Machine Learning and Knowledge nition model based on nonlinear led pattern feature extraction function
Extraction 3 (4) (2021) 835–862. using eeg signals, Cognitive Neurodynamics (2022) 1–12.
[4] S. T. Aung, M. Hassan, M. Brady, Z. I. Mannan, S. Azam, A. Karim, [11] P. Boonyakitanont, A. Lek-Uthai, K. Chomtho, J. Songsiri, A review
S. Zaman, Y. Wongsawat, Entropy-based emotion recognition from mul- of feature extraction and performance evaluation in epileptic seizure de-
tichannel eeg signals using artificial neural network, Computational intel- tection using eeg, Biomedical Signal Processing and Control 57 (2020)
ligence and neuroscience 2022 (1) (2022) 6000989. 101702.
[5] M. Aslan, Cnn based efficient approach for emotion recognition, Jour- [12] T. B. Alakus, M. Gonen, I. Turkoglu, Database for an emotion recogni-
nal of King Saud University-Computer and Information Sciences 34 (9) tion system based on eeg signals and various computer games–gameemo,
(2022) 7335–7346.

11
Biomedical Signal Processing and Control 60 (2020) 101951.

12