Professional Documents
Culture Documents
1 s2.0 S0010482522003985 Main
1 s2.0 S0010482522003985 Main
A R T I C L E I N F O A B S T R A C T
Keywords: Recently, Riemannian geometry-based pattern recognition has been widely employed to brain computer inter
EEG face (BCI) researches, providing new idea for emotion recognition based on electroencephalogram (EEG) signals.
Emotion recognition Although the symmetric positive definite (SPD) matrix manifold constructed from the traditional covariance
Riemannian manifold
matrix contains large amount of spatial information, these methods do not perform well to classify and recognize
SPD matrix
Dimensionality reduction
emotions, and the high dimensionality problem still unsolved. Therefore, this paper proposes a new strategy for
EEG emotion recognition utilizing Riemannian geometry with the aim of achieving better classification perfor
mance. The emotional EEG signals of 32 healthy subjects were from an open-source dataset (DEAP). The wavelet
packets were first applied to extract the time-frequency features of the EEG signals, and then the features were
used to construct the enhanced SPD matrix. A supervised dimensionality reduction algorithm was then designed
on the Riemannian manifold to reduce the high dimensionality of the SPD matrices, gather samples of the same
labels together, and separate samples of different labels as much as possible. Finally, the samples were mapped to
the tangent space, and the K-nearest neighbors (KNN), Random Forest (RF) and Support Vector Machine (SVM)
method were employed for classification. The proposed method achieved an average accuracy of 91.86%,
91.84% on the valence and arousal recognition tasks. Furthermore, we also obtained the superior accuracy of
86.71% on the four-class recognition task, demonstrated the superiority over state-of-the-art emotion recognition
methods.
1. Introduction low-cost, portable, and high temporal resolution EEG signals make it
possible to capture the emotion immediately once the participant re
Emotion is a complex psychological and physiological state [1], ceives the emotional stimulus [7].
which plays an important role in daily life, especially for interpersonal Lately, various classification strategies have been proposed for
communication. In recent years, with the continuous improvement of emotion recognition based on EEG signals [8]. These methods usually
human-computer communication requirements [2], many researchers extract features from the EEG signals and construct classifiers to group
began to focus on emotion recognition technology [3]. Through emotion them into different emotions. Conventional EEG feature extraction
recognition, we can establish an automatic system that can provide methods usually extract features from the time domain or frequency
correct feedback according to different emotions, so as to make domain of EEG signal, such as event-related potential (ERP) [9], statis
human-computer interaction more friendly [4]. In the past decades, tical features (fractal dimension and first-order difference) [10], power
many emotion recognition methods based on different spectral density, autoregressive coefficient, etc. In the construction of
non-physiological patterns have been proposed, such as facial expres classifiers, machine learning algorithms such as naïve Bayes, support
sion, gesture, and sounds [5]. Compared to the non-physiological sig vector machine (SVM), linear discriminant analysis (LDA) [11], and
nals, physiological signals [6] such as electroencephalogram (EEG), bagging ensemble [12] have been widely used. Subasi et al. [13]
electromyogram (EMG), and electrocardiogram (ECG) can be detected extracted features by tunable Q wavelet transform (TQWT), and
by wearable devices. Among these physiological signals, EEG signals employed a rotation forest ensemble classifier on emotion classification.
have been widely utilized for researches of emotion recognition. The J. Atkinson et al. [14] extracted features of statistical features,
https://doi.org/10.1016/j.compbiomed.2022.105606
Received 26 March 2022; Received in revised form 8 May 2022; Accepted 9 May 2022
Available online 13 May 2022
0010-4825/© 2022 Elsevier Ltd. All rights reserved.
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
2
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
projected to tangent space Tp (M) using the logarithmic mapping Logp (pi ) 3. Methods
as:
( − 1 − 1) 1 In this section, we first describe the construction of enhanced SPD
(3)
1
si = log p (pi ) = p2 log p 2 pi p 2 p2 matrices as new samples, and then, we propose a supervised dimension
reduction method on Riemannian manifold as well as an optimization
where si ∈ Tp (M) and p is the tangent point. The inverse mapping: si → pi , method. The framework of our method is shown in Fig. 2.
is defined by the exponential mapping Expp (si ):
( − 1 − 1) 1
(4) 3.1. Wavelet packet transform
1
pi = Expp (si ) = p2 exp p 2 si p 2 p2
2.4. Riemannian mean EEG signals have time-varying and non-stationary characteristics.
Only analyzing EEG signals int time domain or frequency domain can
The mean of Riemannian space is different from that of Euclidean not accurately reflect its characteristics [32]. Time-domain and
space. It is not arithmetic mean, but geometric mean [30]. The Rie frequency-domain characteristics need to be considered together.
mannian mean is defined as: Wavelet transform is a widely used time-frequency analysis tool [33,34].
Compared with the Fourier transform [35] and the short-time Fourier
∑ transform [36], the wavelet transform has the advantage of
n
DR (P1 , …, Pn ) = argmin δ2R (P, Pi ) (5)
P∈P(n) i=1 multi-resolution analysis and can reflect the local details of the signal in
multiple scales. However, wavelet transform only decomposes the
There is no closed-form expression to compute the mean. An efficient
low-frequency part downwards, and cannot satisfy high time resolution
iterative algorithm to compute the Riemannian mean of SPD matrices is
and high frequency resolution at the same time. The wavelet packet
given in Ref. [31].
transform (WPT) is an extension of the wavelet transform. It divides the
signal into high-frequency and low-frequency parts at multiple levels
through a set of filters.
which is different from the low frequency resolution of wavelet
Fig. 3. The enhanced SPD matrix sample. a) Contains covariance matrix blocks between frequency bands. b) The local diagonal blocks of enhanced SPD matrix.
3
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
transform at high frequencies [37]. WPT can subdivide the high fre on SPD matrix is proposed, which maps high dimensional SPD manifold
quency part of the signal, and is widely used in processing random to low dimensional SPD manifold, and can effectively separate different
signals. classes of samples.
u xi , sj = (9)
− e− δ (xi ,sj )/2σ , otherwise
2 2
These new samples X* are used to calculate the covariance matrix by
eq. (6), and finally we get the enhanced SPD matrix P* ∈ R4N×4N :
where δ is the Riemannian metric, σ is the width parameter of the kernel
P* =
1 [ * * T]
X (X ) function. Based on the above ideas, we propose the following objective
(T − 1) function:
⎛ ⎞
T T ∑
N ∑
c
( ) ( )
⎜ Xθ Xθ … Xθ Xγ ⎟ (8) F = argmin δ2 W T xi W, W T sj W u xi , sj
1 ⎜ ⎜⋮ ⋱ ⋮
⎟
⎟ i=1 j=1 (10)
=
(T − 1) ⎜
⎝
⎟
⎠ s.t.W T W = In
T T
Xγ Xθ ⋯ Xγ Xγ
In order to ensure that the new samples WT XW > 0, ∀X ∈ Sym+ , we
We selected EEG signals of some channels to form the enhanced SPD add the constraint WT W = In . There is no closed-form solution for eq
matrix, as shown in Fig. 3. (10), so we design an iterative optimization method to solve the
problem.
3.3. Dimensionality reduction
3.3.2. Optimization
The SPD matrix is usually in high-dimensional space. When the SPD Affine invariant metric can accurately calculate geodesic distance, so
matrix of size n × n is mapped to Euclidean space, the obtained vector AIRM metric is chosen in this paper. equation (10) will be rewritten as
dimensionality is as high as n(n + 1)/2. It will cause the curse of follows:
dimensionality [38]. Therefore, most researches usually choose SPD c ⃒⃒ ⃒⃒2
N ∑⃒⃒
matrix with smaller dimension, which will lead to the loss of many ∑ ⃒⃒ ( T −1 )⃒⃒⃒⃒ ( )
F = argmin T
⃒⃒ log W x WW sW ⃒⃒ u xi , sj (11)
useful features. In fact, this is also a very important reason to limit the i=1
⃒ ⃒ ⃒ ⃒
F
application of Riemannian geometry to multi-channel signals, such as j=1
emotional EEG signals. Sheung [39] proved that logMk ≈ Mk − I under the condition of
There are three main ways to reduce the dimensionality of SPD allowable error, and as the number of iterations increases, the error
matrix. The first method is to map SPD matrix to Euclidean space, and gradually decreases. Therefore, we can write
then use the common methods to reduce dimensionality, such as linear ( ) ( )
dimension reduction method PCA, nonlinear dimension reduction log W T x− 1 WW T sW ≈ W T x− 1 WW T s − I W
( ) (12)
method Isomap, LLE and so on. However, these methods do not provide ≈ W T log x− 1 WW T s W
explicit mapping, and the new samples need to be recalculated. The
The objective function of eq (11) can then be written as:
second is to map the SPD matrix to Hilbert space by using kernel method
and finding the kernel function suitable for Riemannian manifold. The
third is to map a high-dimensional SPD manifold to a low dimensional
SPD manifold, and the obtained low dimensional SPD matrix samples
can be directly used in the Riemannian space classification algorithm.
In this paper, a supervised dimensionality reduction algorithm based
4
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
⃒⃒ ⃒⃒2
∑N ∑ c ( )⃒⃒⃒⃒ ( )⃒⃒⃒⃒
u xi , sj ⃒⃒log WT x− 1 WWT sW ⃒⃒
⃒⃒ ⃒⃒
4. Experiments and results
i=1 j=1
⃒⃒ ⃒F⃒2
∑ N ∑ c
( )⃒⃒⃒⃒ T ( − 1 T
) ⃒⃒⃒⃒ (13)
= u xi , sj ⃒⃒W log x WW s W⃒⃒
i=1 j=1
⃒⃒ ⃒⃒ In this part, we first introduce the public database used in this study.
( ( ) F)
∑∑ (
N c
) Then, we give a detailed description of the related experimental settings.
= u xi , sj Tr WT log x− 1 WWT s W Finally, the experimental results on DEAP database [41] are given and
i=1 j=1
( ) analyzed.
= Tr WT LW
N ∑
∑ c 4.1. DataSet
where L = u(xi , sj )log(x− 1 WWT s).
i=1 j=1
Obviously, this is a trace optimization problem [40], we can itera The data used in this paper was from DEAP database. It is a multi
tively optimize through eigenvalue decomposition. The pseudo code of modal dataset based on physiological signals to analyze emotional state,
Dimensionality Reduction is given in Algorithm 1. and it is one of the most widely used data sets in emotional computing
research. In the data collection experiments, 32 subjects participated in
Algorithm 1. Supervised Dimensionality Reduction on Riemannian 40 Music Video experiments with different emotion tendencies and each
manifold. experiment lasted for 1 min. After watching each video, self-assessment
Fig. 6. Performance comparison of each subject using different classifiers for valence.
makings (SAM) scale including four dimensions of pleasure, arousal, the dividing line, with arousal and valence above 5 being high arousal
dominance and liking were filled in for self-evaluation. The score range and positive valence, and below 5 being low arousal and negative
of each dimension was from 1 to 9. The data recorded 32 channels of valence. Therefore, we further segmented the four quadrants of the
EEG signals. In this paper, a preprocessed EEG dataset was used. The valence-arousal space according to the score as shown in Fig. 4.
sampling frequency of the original EEG signal is reduced to 128 Hz. The At the beginning of data collection, some subjects could not induce
electro-optic artifacts have been removed, and a band-pass filter of emotional state well, so the last 30s data was used as our experimental
4.0–45.0 Hz has been applied. data. It was reported that the duration of emotion was about 0.5–4s
[43], so we use the sliding window of 4s, overlap for 2s, and get 14
samples in one session. Neuroscience research shows that there are
4.2. Data processing differences between the left and right hemispheres of the brain, which
can regulate different emotional states [44]. The brain activities related
We analyzed DEAP data from the dimensions of valence and arousal, to emotions are mainly distributed in the temporal and frontal lobes
which have been divided into 9 levels [42]. A self-score of 5 was used as
6
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
Fig. 7. Performance comparison of each subject using different classifiers for arousal.
[45–47]. Therefore, using the difference in human brain response is very Support vector machine (SVM) classifiers are selected to identify
important for emotion recognition [48,49]. For our study, we used the different emotional states. For KNN, we search the best k value in the
seven symmetrical pairs of electrodes on the right (O2、P8、T8、FC6、 range of 3–10. For RF, the maximum tree depth was optimized in the
F4、F8、AF4) and left hemisphere (AF3、F7、F3、FC5、T7、P7、O1) range of 10–100, and the number of estimators was optimized in the
from the DEAP dataset. Fig. 5 illustrates the EEG electrodes placement. range of 100–500 with step of 100. For SVM, linear kernel is used as
Each subject’s EEG data was divided into 560 samples, four fre kernel function on LIBSVM software. We use the grid search method to
quency bands of each sample were extracted by wavelet packet. Follow search for the optimal parameter C in the range of 2− 10 − 210 . The pa
the method mentioned in Section 3 to generate the enhanced SPD matrix rameters of the above model are selected through 5-fold cross valida
as a new sample. The SPD matrix usually leads to high dimensional tion. We trained the model for each subject and used the mean results of
problems, especially enhanced SPD matrix embedded with more time- 5 times 5-fold cross-validation to access the model performance.
frequency information. The dimensionality reduction method we pro Accuracy (Acc) and F1-score are used to measure and evaluate the
posed can map a high-dimensional SPD manifold into a low-dimensional performance of our proposed method. The Acc is defined in formula
SPD manifold. We determined the best dimension for each subject (14):
through the gird search method.
TP + TN
Acc = (14)
TP + TN + FP + FN
4.3. Classification and evaluation metrics
where TP is the number of true positives, TN is the number of true
In this work, K-nearest neighbors (KNN), Random forest (RF) and
7
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
Fig. 8. Performance comparison of each subject using different classifiers for four-class.
negatives, FN is the number of false negatives, and FP is the number of Finally, KNN, RF and SVM were used to classify them in binary classi
false positives. fication of valence and arousal. The experimental results of different
F1-score combines the results of precision (Pre) and recall rete (Rec) subjects are shown in Fig. 6 – Fig. 7.
at the same time. Their definitions are as follows: For the binary classification of valence, as shown in Fig. 6, the
highest accuracy rate and F1-score for 32 subjects were 96.81% and
TP
Pre = (15) 96.82%, which were obtained by SVM. Overall average accuracies of all
TP + FP
subjects were calculated as 91.86% ± 3.09%, 85.22% ± 5.82%, 82.77%
TP ± 5.63% for SVM, KNN, RF, respectively. In addition, the average of F1-
Rec = (16) score of SVM, KNN, and RF were 90.34%, 82.32%, 80.77%, respectively.
TP + FN
From Fig. 7, the classification results on the binary classification of
2 × Pre × Rec arousal showed that the proposed method achieve the highest accuracy
F1 − score = (17)
Pre + Rec rate of 96.04% and the highest F1-score of 97.03% for 32 subjects. For
SVM, KNN and RF, the average of accuracies were 91.81% ± 3.03%,
4.4. Binary classification of valence and arousal 85.96% ± 5.51% and 83.75% ± 6.54%, the average of F1-score were
89.07% ± 4.36%, 81.63% ± 7.4%, 79.91% ± 7.19% respectively.
We took the geometric mean point of the training sample as the
tangent point to establish the tangent plane, and mapped the reduced-
dimensional matrix to the tangent space to obtain the feature vector.
8
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
Fig. 9. The enhanced SPD matrix in different valence-arousal states. Each enhanced SPD matrix contains covariance matrix blocks in θ, α, β, γ bands. a) HVHA: high-
valence/high-arousal b) LVHA: low-valence/low-arousal c) LVLA: low-valence/low-arousal d) HVLA: high-valence/low-arousal.
4.5. Four-class classification and the average F1-score were 85.60% ± 5.29%, 73.81% ± 8.13%,
72.82% ± 8.54% for SVM, KNN and RF, respectively.
For the four-class classification task, all data processing steps were
the same as the binary classification task. The difference was their labels, 5. Discussion
LALV, HALV, LAHV, and HAHV. The results of each subject are shown in
Fig. 8. In the four-class classification tasks, the highest accuracy we In this paper, we constructed the enhanced SPD matrices and
achieved was 97%, and the highest F1-score was 96.88%. The average designed a supervised dimensionality reduction on Riemannian mani
accuracies were 86.71% ± 4.98%, 77.12% ± 7.57%, 76.63 ± 7.3%%, fold. The enhanced SPD matrix, which contains more time-frequency
9
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
Table 1
Details from various previous researches.
Emotion recognition task Method Features and Classifier Accuracy (%)
Valence Arousal
Binary classification Gao et al. [50] Hjorth, PSD, Differential Entropy, Sample Entropy; SVM 80.52 75.22
Chen et al. [51] amplitude and PSD; SVM 87.07 86.98
Piho et al. [52] HOC, SF, PSE, HOS; SVM 87.11 84.84
An et al. [53] Differential Entropy, spatiotemporal features; CNN-SAE 89.49 90.76
Chen et al. [51] amplitude and PSD; CVCNN 88.76 85.57
Ganapathy et al. [54] spatial and temporal features; CNN 72 75
Our time-frequency and spatial features; SVM 91.86 91.81
Four-class Classification Kong et al. [55] the EEG phase synchronization 67.04
indexes (PSI); SRC
Kwon et al. [56] Wavelet transformed spectrogram of EEG, Short time zero Crossing rate of GRS; CNN 73.43
Mei et al. [57] Connection matrix of the brain structure; CNN 73.10
Our time-frequency and spatial features; SVM 86.71
and spatial information, was used as features. Fig. 9 shows the enhanced reduced by half, the accuracy is still higher than the SPD of the original
SPD matrix for different valence-arousal states, and each enhanced SPD size. As the matrix becomes smaller, the accuracy decreases slowly,
matrix contains the covariance matrix of multiple sub-bands. In the indicating that the reduced dimension matrix still retains most of the
high-valence/high-arousal state, channels from the frontal and temporal effective features. If it is classified in Riemannian space, the appropriate
lobes have strong effects in all four frequency bands. Channels in the matrix size can be selected according to the accuracy. If the classification
right hemisphere were more active in the low-valence/ow-arousal state. is carried out in Euclidean space, a smaller SPD matrix size can be
In the low-valence/high-arousal state, the frontal and temporal lobes selected, which can reduce the training time while maintaining a certain
from the left hemisphere were more active in the theta and alpha bands. accuracy.
In the high valence low arousal state, the interaction between the left By using the SVM classifier, we have obtained the best model per
and right hemispheres in the theta and alpha bands was significantly formance, which is significantly higher than the KNN and RF classifiers.
enhanced. The enhanced SPD matrices in different states capture the To verity the superiority of the proposed method, we compared the
features of different brain regions, which will make the SPD matrices latest methods with our proposed method. Table 1 details the features
distributed at different locations on the Riemannian manifold. So we can and classifiers used in the comparison researches. As shown in Table 1,
use the differences in the distribution of these SPD matrices to better using the SVM classifier [50–52], our proposed method improved at
identify different emotions. least 4.75% in binary classification of valence and at least 3.32% in
Moreover, the supervised dimensionality reduction algorithm we binary classification of arousal. We also compared with other methods
proposed effectively increased the separability of samples while that combined time or frequency domain features with spatial features
reducing the dimensionality. Sample labels were used to bring similar [51,53,54]. Compared with these methods, our method improved
samples closer to the center of the class, while samples of different types 2.37%–19.86% in binary classification of valence and 1.05%–16.81% in
were far away from each other. We compared the classical linear binary classification of arousal.
dimensionality reduction method PCA and the nonlinear dimensionality Our methods were also compared with other four-class emotion
reduction method Isomap with our method. The classification experi recognition methods and the specific details can be found in Table 1.
ments of valence, arousal and four-class on subject 1 were conducted Compared with these studies, our method is at least 13.28% higher than
through different dimensionality reduction methods. As shown in the comparative methods in four-class classification. Experimental re
Fig. 10, the accuracy of our dimensionality reduction method is signif sults indicated that our method could be well used in four-class emotion
icantly better than PCA and Isomap. In addition, we conducted a paired classification tasks. Because our method could gather multi-class sam
t-test on the results, and the p values obtained were less than 0.001, ples near multiple class centers and away from other class centers, which
indicating that the results were statistically significant. increased the separability of samples.
It is worth noting that our dimensionality reduction method maps the
SPD matrix to the low dimensional SPD matrix in Riemannian space, and 6. Conclusion
the samples obtained are suitable for both classifiers in Riemann space
and Euclidean space. Fig. 11 shows the classification accuracy for To conclude, we utilized the enhanced SPD matrix to improve the
different matrix sizes. The result show that when the matrix size is performance of EEG emotion recognition. We designed a supervised
10
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
dimensionality reduction method in Riemann space to reduce the techniques for automated accurate EEG emotion recognition, Comput. Biol. Med.
138 (2021), 104867.
dimensionality of the enhanced SPD matrix. While reducing the
[17] H. Wang, X. Wu, L. Yao, Identifying cortical brain directed connectivity networks
dimensionality, it can also reduce the intra-class distance and increase from high-density EEG for emotion recognition, IEEE Trans. Affect. Comput.
the inter-class distance between samples. In order to verify the effec (2020) 1, 1.
tiveness of our method, we used three classifiers to conducted the binary [18] L. Xu, M. Xu, T. Jung, D. Ming, Review of brain encoding and decoding
mechanisms for EEG-based brain–computer interface, Cognit. Neurodynamics 15
classification of valence/arousal and the four-class classification ex (2021) 569–584.
periments on DEAP. Compared with other state-of-the-art methods in [19] A. Barachant, S. Bonnet, M. Congedo, C. Jutten, Riemannian geometry applied to
recent years, the performance of our proposed method is superior. BCI classification, in: International Conference on Latent Variable Analysis and
Signal Separation, 2010, pp. 629–636.
Currently, our method needs to map the SPD matrix to the Euclidean [20] A. Barachant, S. Bonnet, M. Congedo, C. Jutten, Multiclass brain–computer
space for classification, which increases the computation complexity. In interface classification by Riemannian geometry, IEEE (Inst. Electr. Electron. Eng.)
the next work, we will try to design a classifier that can be directly Trans. Biomed. Eng. 59 (2011) 920–928.
[21] Y. Wang, S. Qiu, X. Ma, H. He, A prototype-based SPD matrix network for domain
applied in Riemannian space to better distinguish the emotion states. In adaptation EEG emotion recognition, Pattern Recogn. 110 (2021), 107626.
addition, emotional EEG signals have large individual differences [22] F. Tang, M. Fan, P. Tiňo, Generalized learning riemannian space quantization: a
among different subjects, and the Riemannian metric has a strong case study on riemannian manifold of spd matrices, IEEE Transact. Neural
Networks Learn. Syst. 32 (2020) 281–292.
generalization ability. Therefore, in future work, we will explore the [23] J. Xu, M. Grosse-Wentrup, V. Jayaram, Tangent space spatial filters for
common emotional patterns among different subjects under the Rie interpretable and efficient Riemannian classification, J. Neural. Eng. 17 (2020),
mannian geometry framework and improve the performance of emotion 026043.
[24] F. Wu, A. Gong, H. Li, L. Zhao, W. Zhang, Y. Fu, A new subject-specific
recognition across subjects.
discriminative and multi-scale filter bank tangent space mapping method for
recognition of multiclass motor imagery, Front. Hum. Neurosci. 15 (2021) 104.
Declaration of competing interest [25] M. Congedo, A. Barachant, R. Bhatia, Riemannian geometry for EEG-based brain-
computer interfaces; a primer and a review, Brain Comput. Interface 4 (2017)
155–174.
The authors declared that they have no conflicts of interest to this [26] T. Zhang, W. Zheng, Z. Cui, Y. Zong, Y. Li, Spatial–temporal recurrent neural
work. network for emotion recognition, IEEE Trans. Cybern. 49 (2018) 839–847.
[27] A.P. Shirokov, Structures on differentiable manifolds, in: R.V. Gamkrelidze (Ed.),
Progress in Mathematics: Algebra and Geometry, Springer US, Boston, MA, 1971,
Acknowledgments pp. 137–207.
[28] W. Förstner, B. Moonen, "A metric for covariance matrices." Geodesy-the Challenge
This work was supported by the National Nature Science Foundation of the 3rd Millennium, Springer, Berlin, Heidelberg, 2003, pp. 299–309.
[29] X. Pennec, P. Fillard, N. Ayache, A Riemannian framework for tensor computing,
of China (61971168, 61871427), the Open Research Projects of Zhijiang Int. J. Comput. Vis. 66 (2006) 41–66.
Lab (No. 2021MC0AB04), the Zhejiang Provincial Key R&D Program of [30] T. Ando, C. Li, R. Mathias, Geometric means, Lin. Algebra Appl. 385 (2004)
China (No.2021C03031) and the Zhejiang Provincial Natural Science 305–334.
[31] P.T. Fletcher, S. Joshi, Principal geodesic analysis on symmetric spaces: statistics of
Foundation of China (LZ22F010003). diffusion tensors, in: Computer Vision and Mathematical Methods in Medical and
Biomedical Image Analysis, Springer, 2004, pp. 87–98.
References [32] A.S. Al-Fahoum, A.A. Al-Fraihat, Methods of EEG signal features extraction using
linear analysis in frequency and time-frequency domains, Int. Sch. Res. Notices
2014 (2014) 1–7, 02-13 2014.
[1] J.J. Gross, Emotion Regulation: Conceptual and Empirical Foundations, Handbook
[33] M.R. Islam, M. Ahmad, Wavelet analysis based classification of emotion from EEG
of Emotion Regulation, 2014, pp. 3–20.
signal, in: 2019 International Conference on Electrical, Computer and
[2] J.A. Russell, A circumplex model of affect, J. Pers. Soc. Psychol. 39 (1980) 1161.
Communication Engineering (ECCE), 2019, pp. 1–6.
[3] R. Cowie, E. Douglas-Cowie, N. Tsapatsoulis, G. Votsis, S. Kollias, W. Fellenz, J.
[34] Z. Mohammadi, J. Frounchi, M. Amiri, Wavelet-based emotion recognition system
G. Taylor, Emotion recognition in human-computer interaction, IEEE Signal
using EEG signal, Neural Comput. Appl. 28 (2017) 1985–1990.
Process. Mag. 18 (2001) 32–80.
[35] J.W. Cooley, J.W. Tukey, An algorithm for the machine calculation of complex
[4] T.J. Brigham, Merging technology and emotions: introduction to affective
Fourier series, Math. Comput. 19 (1965) 297–301.
computing, Med. Ref. Serv. Q. 36 (2017) 399–407.
[36] J. Allen, Short term spectral analysis, synthesis, and modification by discrete
[5] Z. Zeng, M. Pantic, G.I. Roisman, T.S. Huang, A survey of affect recognition
Fourier transform, IEEE Trans. Acoust. Speech Signal Process. 25 (1977) 235–238.
methods: audio, visual, and spontaneous expressions, IEEE Trans. Pattern Anal.
[37] S. Chinara, Automatic classification methods for detecting drowsiness using
Mach. Intell. 31 (2008) 39–58.
wavelet packet transform extracted time-domain features from single-channel EEG
[6] R.A. Calvo, S. D’Mello, Affect detection: an interdisciplinary review of models,
signal, J. Neurosci. Methods 347 (2021) 108927.
methods, and their applications, IEEE Trans. Affect Comput. 1 (2010) 18–37.
[38] M. Congedo, P.L.C. Rodrigues, F. Bouchard, A. Barachant, C. Jutten, A closed-form
[7] C. Li, W. Tao, J. Cheng, Y. Liu, X. Chen, Robust multichannel eeg compressed
unsupervised geometry-aware dimensionality reduction method in the Riemannian
sensing in the presence of mixed noise, IEEE Sensor. J. 19 (2019) 10574–10583.
Manifold of SPD matrices, in: 2017 39th Annual International Conference of the
[8] M.R. Islam, M.A. Moni, M.M. Islam, M. Rashed-Al-Mahfuz, M.S. Islam, M.K. Hasan,
IEEE Engineering in Medicine and Biology Society (EMBC), 2017, pp. 3198–3201.
M.S. Hossain, M. Ahmad, S. Uddin, A. Azad, Emotion recognition from EEG signal
[39] S.H. Cheng, N.J. Higham, C.S. Kenney, A.J. Laub, Approximating the logarithm of a
focusing on deep learning and shallow learning techniques, IEEE Access 9 (2021)
matrix to specified accuracy, SIAM J. Matrix Anal. Appl. 22 (2001) 1112–1125.
94601–94624.
[40] E. Kokiopoulou, J. Chen, Y. Saad, Trace optimization and eigenproblems in
[9] C.A. Frantzidis, C. Bratsas, C.L. Papadelis, E. Konstantinidis, C. Pappas, P.
dimension reduction methods, Numer. Lin. Algebra Appl. 18 (2011) 565–602.
D. Bamidis, Toward emotion aware computing: an integrated approach using
[41] S. Koelstra, C. Muhl, M. Soleymani, J. Lee, A. Yazdani, T. Ebrahimi, T. Pun,
multichannel neurophysiological recordings and affective visual stimuli, IEEE
A. Nijholt, I. Patras, Deap: a database for emotion analysis; using physiological
Trans. Inf. Technol. Biomed. 14 (2010) 589–597.
signals, IEEE Trans. Affect. Comput. 3 (2011) 18–31.
[10] S. Tripathi, S. Acharya, R.D. Sharma, S. Mittal, S. Bhattacharya, Using deep and
[42] J.A. Russell, A. Mehrabian, Evidence for a three-factor theory of emotions, J. Res.
convolutional neural networks for accurate emotion classification on deap dataset,
Pers. 11 (1977) 273–294.
in: Twenty-ninth IAAI Conference, 2017, pp. 4746–4752.
[43] V. Gupta, M.D. Chopda, R.B. Pachori, Cross-subject emotion recognition using
[11] R. Nawaz, K.H. Cheah, H. Nisar, V.V. Yap, Comparison of different feature
flexible analytic wavelet transform from EEG signals, IEEE Sensor. J. 19 (2018)
extraction methods for EEG-based emotion recognition, Biocybern. Biomed. Eng.
2266–2274.
40 (2020) 910–926.
[44] K.M. Alfano, C.R. Cimino, Alteration of expected hemispheric asymmetries:
[12] K.S. Kamble, J. Sengupta, Ensemble machine learning-based affective computing
valence and arousal effects in neuropsychological models of emotion, Brain Cognit.
for emotion recognition using dual-decomposed EEG signals, IEEE Sensor. J. 22 (3)
66 (2008) 213–220.
(1 Feb.1, 2022) 2496–2507.
[45] B. Bogert, T. Numminen-Kontti, B. Gold, M. Sams, J. Numminen, I. Burunat,
[13] A. Subasi, T. Tuncer, S. Dogan, D. Tanko, U. Sakoglu, EEG-based emotion
J. Lampinen, E. Brattico, Hidden sources of joy, fear, and sadness: explicit versus
recognition using tunable Q wavelet transform and rotation forest ensemble
implicit neural processing of musical emotions, Neuropsychologia 89 (2016)
classifier, Biomed. Signal Process Control 68 (2021), 102648.
393–402.
[14] J. Atkinson, D. Campos, Improving BCI-based emotion recognition by combining
[46] J.P. Ramirez-Mahaluf, J. Perramon, B. Otal, P. Villoslada, A. Compte, Subgenual
EEG feature selection and kernel classifiers, Expert Syst. Appl. 47 (2016) 35–41.
anterior cingulate cortex controls sadness-induced modulations of cognitive and
[15] T. Tuncer, S. Dogan, A. Subasi, A new fractal pattern feature generation function
emotional network hubs, Sci. Rep. 8 (2018) 1–16.
based emotion recognition method using EEG, Chaos, Solit. Fractals 144 (2021),
[47] N. Zhuang, L. Jiang, B. Yan, L. Tong, J. Shu, K. Yang, D. Yao, P. Xu, Y. Zeng, Neural
110671.
mechanism of affective perception: evidence from phase and causality analysis in
[16] A. Dogan, M. Akay, P.D. Barua, M. Baygin, S. Dogan, T. Tuncer, A.H. Dogru, U.
the cerebral cortex, Neuroscience 461 (2021) 44–56.
R. Acharya, PrimePatNet87: prime pattern and tunable q-factor wavelet transform
11
Y. Gao et al. Computers in Biology and Medicine 146 (2022) 105606
[48] D. Huang, S. Chen, C. Liu, L. Zheng, Z. Tian, D. Jiang, Differences first in [53] Y. An, S. Hu, X. Duan, L. Zhao, C. Xie, Y. Zhao, EEG emotion recognition based on
asymmetric brain: a bi-hemisphere discrepancy convolutional neural network for 3D feature fusion and convolutional autoencoder, Front. Comput. Neurosci. (2021)
EEG emotion recognition, Neurocomputing 448 (2021) 140–151. 83.
[49] J. Zhang, M. Chen, S. Zhao, S. Hu, Z. Shi, Y. Cao, ReliefF-based EEG sensor [54] N. Ganapathy, Y.R. Veeranki, R. Swaminathan, Convolutional neural network
selection methods for emotion recognition, Sensors 16 (2016) 1558. based emotion classification using electrodermal activity signals and time-
[50] Q. Gao, C. Wang, Z. Wang, X. Song, E. Dong, Y. Song, EEG based emotion frequency features, Expert Syst. Appl. 159 (2020), 113571.
recognition using fusion feature extraction method, Multimed. Tool. Appl. 79 [55] W. Kong, X. Song, J. Sun, Emotion recognition based on sparse representation of
(2020) 27057–27074. phase synchronization features, Multimed. Tool. Appl. 80 (2021) 21203–21217.
[51] J.X. Chen, P.W. Zhang, Z.J. Mao, Y.F. Huang, D.M. Jiang, Y.N. Zhang, Accurate [56] Y. Kwon, S. Shin, S. Kim, Electroencephalography based fusion two-dimensional
EEG-based emotion recognition on combined features using deep convolutional (2D)-convolution neural networks (CNN) model for emotion recognition system,
neural networks, IEEE Access 7 (2019) 44317–44328. Sensors 18 (2018) 1383.
[52] L. Piho, T. Tjahjadi, A mutual information based adaptive windowing of [57] H. Mei, X. Xu, EEG-based emotion classification using convolutional neural
informative EEG for emotion recognition, IEEE Trans. Affect. Comput. 11 (2020) network, in: 2017 International Conference on Security, Pattern Analysis, and
722–735. Cybernetics (SPAC), 2017, pp. 130–135.
12