CODY24 | Dr. Larry Farwell

Cognitive Neurodynamics

Automated Pain Severity Detection by Using Adaptive Band-Wise GNN Technique with

EEG Signals

--Manuscript Draft--

Manuscript Number:

CODY-D-24-00350

Full Title:

Automated Pain Severity Detection by Using Adaptive Band-Wise GNN Technique with

EEG Signals

Article Type:

Original Research

Keywords:

Pain severity classification, EEG, GNN, Connectivity, Adaptive hierarchical classifier

Abstract:

Recently, several studies concerning the richness of electroencephalogram (EEG) signals’ content to detect pain and classify pain severity have been made. In this study, a novel band-wise adaptive classifier is proposed based on successive brain graphs through five levels of pain. The EEG data was acquired from 44 healthy subjects while putting their hand in the cold water to feel pain continuously and increasingly over time till the intolerable stage that they got off their right hand. During the cold pressor test, participants reported their pain at five different levels while their EEG signals were captured by 32 silver electrodes. The data was decomposed into five frequency bands and for each band-wise channel, discriminative EEG features were calculated in two different feature sets. Afterward, three connectivity estimators were applied to determine the adjacency matrices for each frequency band and construct the

corresponding brain graphs. The constructed graphs along with feature nodes were fed into an adaptive graph convolutional neural network (GCNN) based hierarchical classifier which selects the most discriminative graphs at each node. The discriminability of each estimator was tested by the Kruskal-Wallis test. In each stage

of the proposed adaptive hierarchical tree, the most discriminative band-wise GCNN was selected and applied. Based on the proposed classifier, the highest accuracy reached 88.5% in classifying five levels of pain, which outperforms the counterparts.

Manuscript Click here to access/download;Manuscript;Nezam-Journal- sprng.pdf

Click here to view linked References

Automated Pain Severity Detection by Using

Adaptive Band-Wise GNN Technique with

EEG Signals

F. A. Alenizi1, T. Nezam2, R. Boostani3, R. J. Ismail4, M. Mohammadi5

1Elec. Eng. Dept., College of Eng., Prince Sattam Bin Abdulaziz University, Al-Kharrj 11942, Saudi Arabia.

2Elec. Eng. Faculty, Shiraz University of Technology, Shiraz, Iran.

3Biomedical Eng. Group, CSE & IT Dept., Electrical and Computer Engineering Faculty, Shiraz University, Shiraz, Iran.

4Computer Science Dept., Cihan University-Erbil, Erbil, Iraq.

5Information Technology Dept., College of Eng. And Computer Science, Lebanese French University, Kurdistan Region, Iraq.

Abstract

44 healthy subjects while putting their hand in the cold water to feel pain continuously and increasingly over time till the intolerable stage that they got off their right hand. During the cold pressor test, participants reported their pain at five different levels while their EEG signals were captured by 32 silver electrodes. The data was decomposed into five frequency bands and for each band-wise channel, discriminative EEG features were calculated in two different feature sets. Afterward, three connectivity estimators were applied to determine the adjacency matrices for each frequency band and construct the corresponding brain graphs. The constructed graphs along with feature nodes were fed into an adaptive graph convolutional neural network (GCNN) based hierarchical classifier which selects the most discriminative graphs at each node. The discriminability of each estimator was tested by the Kruskal-Wallis test. In each stage of the proposed adaptive hierarchical tree, the most discriminative band-wise GCNN was selected and applied. Based on the proposed classifier, the highest accuracy reached 88.5% in classifying five levels of pain, which outperforms the counterparts.

Keywords: Pain severity classification, EEG, GNN, Connectivity, Adaptive hierarchical classifier.

1. Introduction

Accurately determining the level of subjective pain intensity is crucial for the identification, prediction, and management of chronic pain conditions [1,2]. The current methods used to assess pain rely on self- report measures, which are not suitable for those who are unable to articulate their pain accurately or completely. This includes individuals with dementia [3,4], disorders of consciousness (such as coma) [3,5], cognitive impairments [3,6], non-verbal individuals (such as palliative care patients who are unable to communicate) [3,7], and children (such as infants and newborn populations) [3,8]. Moreover, due to subjective feelings of pain, it is difficult to measure it robustly [1,4]. Consequently, the intricacy of achieving precise pain assessment, especially among populations without the ability to communicate and report their pain level, underscores the need for enhanced automated evaluation techniques.

Recent efforts to alleviate the need for self-report techniques have made strides in understanding the biological indicators of pain severity through the utilization of neuroimaging [9,10]. Additionally, machine learning (ML) analysis of neuroimaging data facilitates the recognition of pain intensity biomarkers. ML encompasses algorithms that discern and comprehend patterns from data to generate predictions on new

inputs without explicit programming. This is accomplished through the utilization of optimization, statistical, and probabilistic methodologies [11–13].

One particular study revealed that the subjective sensation of pain was linked to alterations in the electroencephalogram (EEG) activity across various frequency bands, encompassing alpha-II, beta-I, beta- II, and theta [14]. Another study reported the effect of pain on the cortical oscillations in the theta, alpha, beta, and gamma frequency bands in different regions of the cortex, such as the frontal, central, parietal, temporal, and occipital regions [15-17]. Individuals with chronic pain often exhibit altered theta oscillations in their resting state EEG, as seen in fibromyalgia patients [18]. Furthermore, during pain, augmented theta oscillations are observed in the central and parietal lobs [19]. Alongside this, tonic pain stimulation leads to a decrease in the alpha band power and an increase in the beta band power, as mentioned in [15-17]. Studies have also shown that there is a decrease in the global alpha band power and an increase in beta band power in response to tonic cold pain stimulation [20]. Recently, it has been demonstrated that the peak alpha frequency can reliably predict pain sensitivity [21,22]. Hence, considering numerous investigations have been conducted to explore the association between pain intensity and EEG variations, it can be inferred as an indicator of the level of pain experienced.

The emergence of Graph Neural Networks (GNN) [23] has offered a concrete methodology for constructing neural networks that operate on graph domain information. Graph Convolutional Neural Networks (GCNN)[24] are an extension of CNNs designed for non-Euclidean domains and 3D shapes.

A few studies have investigated the utilization of GNNs in EEG classification problems. Researchers in [25], implied that GNNs are a better choice for EEG problems compared to CNN because it consider the complex functional neural connectivity between electrodes. In another study [26] GNNs were selected as a versatile tool for EEG processing because it capture functional connectivity between EEG channels, which is important for interpreting EEG-based motor imagery [26].

This study investigates a classification framework based on GCNN for automating the detection of pain severity. The framework utilizes the multivariate feature representation obtained through EEG signal-based brain connectivity analysis.

2. Related work

EEG is a tool of utmost importance owing to its numerous benefits, encompassing its expeditious and economical nature, its capability to capture brain activity with a temporal resolution of milliseconds, and its non-intrusive, secure, and user-friendly procedure [27]. EEG has the potential to be inferred as a biomarker of pain perception [28]. An increase in the delta and gamma bands during painful states is observed in several studies [28]. Researchers in [29] reported the energy of EEG in the Delta and Alpha frequency bands are the most important features. Previous studies have attempted to classify pain and no- pain states by estimating the coherence values in the Delta, Alpha, and Beta bands between the left and right brain lobes [30–35]. Alternatively, Zherdin and Schulz [36] adopted a multivariate pattern analysis approach to analyze EEG signals and subsequently utilized these characteristics as input for a support vector machines (SVM) classifier. Their findings, based on a sample size of 22 subjects and two states (pain and no-pain), demonstrated 83% classification accuracy, suggesting that further improvement is possible. In another research, EEG signal on nine subjects using the Cold Pressor Test (CPT) protocol was recorded and after feeding the EEG spectral features to a fuzzy-based SVM classifier, 96% accuracy was achieved between pain and no-pain states [37].

In a separate endeavor, Toliyat and Vatankhah [38] utilized the wavelet coherence transformation on the EEG signals of 13 individuals (recorded under the CPT protocol) and by employing a nonlinear SVM classifier, with the radial basis function (RBF) kernel, 95% accuracy in distinguishing between pain and no-pain states, was obtained.

Nezam et al. [39] performed their research on the EEGs of 24 individuals feeling pain in a CPT protocol. They applied ICA to remove EOG and EMG artifacts and proposed a hierarchical decision tree regarding different features such as fractal dimension, spectral entropy, Shannon entropy, and approximate entropy. The results were 83% and 62% accuracy for 2-level and 5-level pain classification, respectively. Therefore, it is necessary to develop a new category of methods to achieve higher accuracy in addressing multi-class pain problems. As more evidence, researchers in [40] have reported the significance of the alpha band during pain feelings.

EEG functional connectivity was investigated through the measurement of the synchronization and relation between various brain regions [41]. When someone feels pain, the brain's sensory area produces spikes that activate the related network of the brain [42]. To uncover the coactivation map of different brain regions in response to varying levels of pain, graph-based tools for structural and functional brain connectivity appear to be appropriate candidates [43].

Modarres Haghighi et al. [44] did their research on the assessment of the brain's oscillatory dynamics caused by pain by utilizing the characteristics of both local and global brain connectivity graphs and achieved 86% accuracy for the 5-class classification of pain. Although it is a good result for the problem, due to concentrating only on the alpha band features, the information on other bands is missed. While any individual may feel pain in any situation and state, investigating all frequency bands may help distinguish the severity of pain more precisely.

The utilization of GNNs [45] has emerged as a novel and underexplored deep learning-based approach for classification within the realm of neurophysiological investigation. The graph neural network GNNs [45] based algorithms place great importance on the fixed network structure that exists at a particular moment, which limits their capacity to detect the evolving alterations in brain network arrangements and successfully remove disturbances within brain nodes when undergoing the training phase [46]. GNNs effectively leverage the inherent topological structure of graphs, yielding commendable capabilities in terms of generalization, expressivity, and transferability [47]. To date, limited researchers have endeavored to employ GNN in EEG classification problems, showcasing distinct variations in terms of adopted graph representation techniques and model structures. In a study conducted by [48], a fusion model incorporating GCNN and Long Short-Term Memory (LSTM) has been proposed for emotion recognition, wherein the adjacency matrix has been computed through the utilization of a Gaussian kernel distance function.

In a previous study, a regularized graph neural network (RGNN) was introduced to address the challenges associated with EEG-based emotion recognition. Also, a novel adjacency matrix was proposed, which combined a distance-based connectivity matrix with information from a selected number of global connections. This adjacency matrix was designed to capture long-distance interactions between the EEG channels [49].

In contrast to the aforementioned distance-based graph representation techniques, alternative approaches have been explored in other studies [50,51], in which the brain graph adjacency is derived from the functional connection strength between pairs of electrodes. Specifically, Pearson's Correlation [49] and Phase Locking Value (PLV) [51] were utilized as measures of functional connectivity for motor imagery and emotion recognition tasks, respectively.

All of the previous studies mentioned above involve the selection of a set of universal EEG features as channel properties (node features), leveraging domain-specific knowledge. However, it is worth noting that no similar investigation has been conducted thus far in the context of pain detection and severity classification. Our study considers all frequency bands involved in EEG and proposes a GCNN-based hierarchical classifier that adaptively assigns proper features and connectivity estimators of the most discriminative frequency bands in each level node.

The most highlighted contributions of this study are mentioned briefly below:

I. Expanding our previously collected dataset containing the EEG signals of 44 subjects through CPT II. First time applying GCNNs for classifying pain severity

III. Amalgamation of brain connectivity patterns and individual electrode-level features specific to pain, for feature extraction.

IV. Determining frequency-based connectivity values and resulting in different graph networks in each band

V. Designing a GCNN-based hierarchical classifier

VI. Applying an adaptive hierarchy by implementing the most discriminative GCNN in each node based on the Kruskal Wallis test.

3. Materials and methods

3.1. Data acquisition, accessibility, and ethics approval

In this particular investigation, a total of 44 right-handed volunteers who were deemed to be in good health, comprising 20 males and 22 females, with a mean age of 27 years and a range of 20 to 32 years, were recruited. It is important to note that no indication of pain, injury, or any other dermal issues on their hands was observed. Prior to the commencement of signal recording, each volunteer provided written consent. Additionally, they were allowed to withdraw from the experiment at any given time and for any reason.

The acquisition of EEG signals was carried out using 32 silver electrodes. A standard EEG cap was placed on the participants' heads, adhering to the 10-20 protocol. To maintain optimal electrode impedance, it was ensured that all electrodes had an impedance level below 5 Kilo-Ohms. The recorded signals were sampled at a rate of 250Hz and subsequently subjected to a Butterworth band pass filter with a 5th-order configuration. This filtering process involved removing baseline drift, linear trend, and high-frequency noises by setting the cut-off frequencies at 0.5Hz and 70Hz, respectively.

CPT was implemented as the stimulus for pain owing to its minimal side effects. Using EEG signal recording, each volunteer was seated on a comfortable chair in an environment devoid of any audio or visual stimuli. Initially, during the "no-pain" state, EEG signals were collected for 30 seconds. Subsequently, the subjects submerged one of their hands in cold water at a temperature of 1.7±0.2 degrees Celsius, and once they experienced pain, the corresponding EEG signals were categorized as the first classification of pain. All participants underwent the CPT procedure multiple times prior to the recording of EEG signals to obtain an accurate perception of pain levels. Due to the subjective nature of pain perception, the participants verbally communicated their pain intensity, and the recorded EEG signals were annotated accordingly. As time progressed, the perception of pain gradually escalated until the participants involuntarily got off their hands from the cold water.

It should be noted that Data archiving is not mandated but data will be made available on reasonable request.

The process of data acquisition of this study was approved by the local ethics committee of the Neurophysiology laboratory of Shiraz University of Medical Science to follow the world health organization mandatory fields (“Ethical principles for medical research involving human subjects,” 2001).

3.2. Proposed method

The study was done in a subject-based approach, which means each subject was analyzed separately and the preprocessing steps were done based on each individual’s EEG signal characteristics. Train and test sets were split as 12 subjects out of 44 were selected as test sets and the other 32 subjects were trained through a 5-fold cross-validation technique to improve the generalizability of the model.

Figure 1 shows the block diagram of the study. In the first step, the recorded EEG signal was decomposed into the five frequency bands e.g. Delta, theta, alpha, beta, and gamma bands. Then in each frequency band, brain connectivity features were calculated for each trial using preprocessed multichannel EEG data. Furthermore, specific EEG features that were relevant for distinguishing the multi-class pain dataset were computed for each individual trial. Subsequently, the connectivity relationships for each trial were determined using various connectivity parameters and visualized as a brain graph or network. In this graph, the handcrafted EEG features served as the node features, and the edges of the graph were determined based on the corresponding connectivity strength. The resulting graphs, along with their class labels, in each frequency band were then utilized as input to the GCNN model. In other words, five GCNNs were trained separately for each frequency band and as three different approaches are considered for calculating the connectivities, there are fifteen GCNN-trained models in total. It should be noted that the main classification task is done through an adaptive hierarchical decision tree-like classifier that passes different GCNN for each node of the tree based on its efficiency in classification results. A detailed explanation of this proposed methodology can be found in the following subsections.

Fig.1. Outline of the proposed study.

3.2.1. Proposed hierarchical classifier

After constructing the graphs, the Kruskal-Wallis analysis was performed to consider the separability of different classes. Different estimators were applied to each level and the outcome of the test provides an

adaptive hierarchical tree that applies the most discriminative estimator to the corresponding GCNN in each stage of the classification. Figure 3 shows the most discriminative estimators in each classification level.

Fig. 3. The candidate estimators in each node

As shown in Fig. 3, the most discriminative estimator in the first node is the PC estimator in the delta band which classifies five classes into No-pain, Low pain versus Medium, High, and Intolerable ones. The next chosen estimator is GC in the alpha band which classifies No-pain from low pain class. PC in the alpha band is the discriminator of Medium versus High-Intolerable levels and finally, High and Intolerable classes are classified by calculating GC in the gamma band. It is worth noting that PLI could not reach the criteria in comparison with PC and GC at any node.

The following Figures demonstrate samples of the changes in each node’s derived network. Changes in PC of the delta band during pain increases are illustrated in Fig. 4 in both network connectivities and adjacency matrices illustration. The modification in different levels of pain is evident in the Figures. It should be noted that to give a better visualization, the edges above the mean of all other connectivity values are drawn in the Figures.

Fig. 4. Brain graphs and corresponding adjacency matrices of PC estimator in delta band for five levels of pain from left to right.

In Fig. 5 the modifications in the most discriminative band-wise estimators which were applied to the corresponding nodes in the adaptive hierarchical classifier are illustrated. Fig. 5.A. is an illustration of network changes in the Alpha band based on the GC estimator which is applied as the base of the GCNN in the no-pain state vs. low-pain node. Fig. 5.B shows the network changes in the Alpha band based on the PC estimator which is applied in the median-pain vs. High/intolerable-pain node. Fig. 5.C. is the network changes in the Gamma band based on GC estimator which is determined as the GCNN constructor for classifying high-pain state vs. intolerable-pain state.

Vs.

(A) No-pain vs. Low-pain (B) Median-pain vs. High/Intolerable-pain

Vs.

Fig. 5. A. Brain graph network changes in the Alpha band based on GC estimator in no-pain state vs. low-pain state (right to left). B. Network changes in the Alpha band based on PC estimator in median-pain vs. High/intolerable-pain states (right to left). C. Network changes in the Gamma band based on GC estimator in of high-pain state vs. intolerable-pain state (right to left).

As another input to GCNN, two pain-specific discriminative feature sets were considered and applied as the node features. The first set of features was a combination of time domain and nonlinear features derived from each channel and the second set consisted of time-frequency features. In Table 1, a detailed list of each set is shown.

Table1. Two feature sets as GNN’s nodes features.

Feature set

Feature types

Features names

Time domain and nonlinear features

mean, SD, sample entropy, Shannon entropy, fractal dimension

Time-frequency features

min. value, max. value, mean, SD, power value, absolute mean of wavelet coefficients

3.3. Connectivity estimators employed

Within the realm of brain connectivity estimation, our study has chosen three estimators of distinct modalities. A concise overview of each estimator and the rationale behind their selection is provided below.

3.3.1. Time-domain FC estimator: Partial Correlation (PC)

Historically, the utilization of bi-variate time-domain cross-correlation measures, specifically Pearson's correlation coefficient, has been customary in deducing connectivity between two distinct areas of the brain. Due to the potential influence of a neighboring third region on the correlation between the activities of the two brain regions, the concept of partial correlation, which evaluates the connectivity between two regions while mitigating the impact of the third region, has been taken into account [52]. The calculation of partial correlation is presented in (1).

????|? ????????

???|? =

?(1?????2)(1?????2)

(1)

where in this context, the symbol "???" represents the measure of cross-correlation between EEG signals of every two electrodes, denoted as ? and ?. Additionally, "???|?" signifies the cross-correlation between the same electrodes ? and ?, but in the presence of a third channel ?. When considering an EEG device with a total of ? electrodes, it is possible to establish (??2) partial correlations between every two channels ? and

?, taking into account the other (??2) channels as the third channel. Due to the influence of volume conduction, whereby the correlation between two channels may be influenced by any of the remaining (?

? 2) channels, this study examines the minimum value among these (??2) potential partial correlation values. This minimum value is regarded as the measure of partial correlation interaction strength between the channels ? and ?.

????? = min[???|?] , ?? ? ?, ? (2)

3.3.2. . Phase-domain FC estimator: Phase Lag Index (PLI)

The utilization of phase synchronization techniques [53] in the examination of connectivity is grounded on the concept that if two neural masses are functionally linked, then the phase disparity of the resulting electrical signals from those regions will remain relatively constant. Within recent EEG literature, Phase Locking Value (PLV) [54] and Phase Lag Index (PLI) [55] have emerged as two widely used estimators for phase connectivity. Due to its limited susceptibility to volume conduction effects [55] in comparison with PLV, PLI has been selected for our sensor-based investigation of connectivity. When dealing with two sensor signals, the computation of PLI is accomplished using equation (3).

PLI = |??1 ???

??? (??? [??(?????)??]) |

(3)

where the instantaneous phases ??? and ??? are derived through the Hilbert transform on the original time series. The length of the original signals is denoted by ?, and the ??? function is considered. The ?????? function is employed to eliminate the phase lag of 0 and ??, as these values could potentially arise from the influence of volume conduction.

3.3.3. The EC estimator in the time domain: Granger Causality (GC)

The fundamental assumption underlying the concept of Granger causality is the temporal precedence of the cause over the effect, which aids in predicting the effect [56]. The measure of Granger causality in the time domain [57] serves as a directed indicator of the causal influence [58] between two brain regions, wherein one region influences the activation of the other region and vice versa. Granger causality is quantified by evaluating the auto-regressive (AR) and bi-variate auto-regressive (BVAR) statistics of the given time-series data. The computation of Granger causality between signals ??? and ??? was performed as

described below [59].

??? = ???

1 A?? ???????? + e???

1 A?? ???????? + ??? 1

B? ????????

+ ???

(4)

(5)

GC is computed as,

???????????????(e???)

????????? ???????????????(?? )

where the time lag between observations, denoted as ?, and the maximum lag, denoted as ??, are crucial factors in the analysis. The AR coefficients represented by ??, ??, and ??, play a significant role in this process. Additionally, the prediction errors are denoted by ?? and ??.

3.4. Brain Connectivity network construction

The brain connectivity network is constructed using a graph, denoted as ?? = {?, E, A}. The set of nodes or vertices is represented by V, the set of edges is represented by E, and the adjacency matrix of the graph is represented by A (? ? R?×?, where ? is the total number of nodes). In the sensor space studies, the positions of the electrodes are considered as the nodes of the graph. The connectivity strength between two electrodes represents the edge between the corresponding nodes. The weights of the edges are determined by the corresponding connection strengths [60].

In this analysis, the edges of the sparse network were determined by considering connections with a strength value greater than the average strength of all connections. It is worth noting that several thresholds were determined and the optimal one was identified to disregard the weaker connections. Then this matrix was binarized with values equal to 1 for the entities greater than the threshold and other entities equal to zero.

GNN models generally learn to transform the initial feature representation of the nodes, denoted as matrix

?, and produce a transformed representation output ?? ? R?×??. For each GNN layer ?, the feature update can be expressed as shown in equation (7),

??+1 = f (?? , A) (7)

where ? takes on values from 0 to ?? (??: the total number of layers), ?0 is the input representation ?, ?? is the output ??, and ? represents the function to be learned. One possible implementation of ? is given by equation (8),

??+1 = ??(??? ??) (8)

where ?? is the weight matrix and ?? denotes a non-linear transformation. In this case, the features of all neighboring nodes, including itself, were aggregated using a weighted sum, followed by the application of a non-linear function.

The computation of (8) repeatedly across all GNN layers has the potential to result in elevated values for the entries of ?? . To address this concern, the GCNN [61] approach made modifications to the feature transformation of (8) using (9).

??+1 = ??(??1/2???1/2?? ??) (9)

To mitigate the possibility of ?? reaching excessively high values, the adjacency matrix ? was normalized by the diagonal degree matrix D. A detailed theoretical analysis of GCNN, rooted in spectral theory, can

be found in the original studies [61,62]. The entire graph's transformed feature representations were then consolidated by one or more fully connected layer(s) [63].

3.5. GCNN implementation details

In our conducted experiments, we entailed the utilization of the adjacency matrix and node features to construct a graph signal representation of the EEG signals, acting as the input for the classification model. The input was then processed through multiple graph convolution layers, interspersed with non-linear activation functions. Subsequently, a layer performs the pooling operation to integrate and convert the final node representations into a cohesive graph embedding. This embedding was further propagated through a fully connected layer, culminating in the acquisition of the final class output via a softmax activation function. During the training process, a binary-cross entropy loss function and the 'Adam' optimizer were employed.

The model employed two graph convolution layers. Hyper-parameter tuning was done through a grid- search algorithm and the most efficient values of each parameter were selected and applied which is mentioned in the following. The activation function utilized in the graph convolution layers was the Rectified Linear Unit (ReLU). The drop-out rate applied to the fully connected layers was set to 0.2. The learning rate specified for the model was 0.001. The batch size employed during training was 32. The optimizer utilized for the model was 'Adam'. Figure 2 shows the applied GNN network architecture.

Fig. 2. GCNN architecture.

4. Results and Discussion

In this section, an analysis is conducted on the efficacy of the proposed framework by utilizing the multi- variable feature representations derived from three brain connectivity parameters in conjunction with specific individual electrodes' feature sets. The effectiveness of this framework is evaluated through the

computation of a hybrid feature representation, which comprises two steps: the computation of the brain connectivity network structure and the computation of the individual EEG electrodes' features. Both of these steps are subsequently discussed.

The formation of the brain connectivity network involves the exploration of three distinct connectivity estimators, namely Partial Correlation, Phase Lag Index, and Granger Causality, to construct a suitable and differentiating brain connectivity network. For each connectivity estimator, a comprehensive and weighted adjacency matrix is obtained from the multichannel EEG data of each trial. To create a sparse and cost- effective brain network, an average thresholding scheme is employed, wherein the complete networks undergo a thresholding process. Here, the thresholding was performed by eliminating the weak and insignificant connections from the complete network.

The goal of organizing two different feature sets was to obtain the influence of each set on the ultimate result. Therefore, the classification was performed three times, first with F1 features, second with F2, and third with all F1 and F2 together. As shown in Table 2, applying both F1 and F2 feature sets outperforms the other two states.

Table 2. The influence of the node’s feature set.

Node

Feature

Accuracy(%)

Specificity(%)

Sensitivity(%)

88.1

82.3

90.1

87.2

81.5

89.2

F1&F2

88.5

82.6

90.7

4.1. Ablation study of the proposed method

To make sure about the performance of the proposed method, the classification problem was solved by performing GCNN without considering the decision tree structure. The results approved the efficiency of the proposed method to a considerable margin and are shown in Table 3.

Table 3. Ablation study.

Classification Structure

Accuracy(%)

Specificity(%)

Sensitivity(%)

5-class GNN with PC

81.7

76.2

82.9

5-class GNN with GC

81.5

75.8

82.6

4.2. Results of the Proposed Method

In Table 4, the values of accuracy, specificity, and sensitivity in each of the four stages of the hierarchical classifier are depicted in each class.

Table 4. Classification results of the pain classes at different nodes of the proposed hierarchical classifier.

Classification level

Accuracy(%)

Specificity(%)

Sensitivity(%)

No pain, low/med, high, intolerable

95.8

92.7

96.5

Low/ no-pain

82.6

74.3

85.5

Med/ High, intolerable

84.2

75.2

87.6

High/ intolerable

91.4

88.5

93.2

The confusion matrix which is interpreted as an illustration of the proposed model performance is summarized in table 5.

Table 5. Confusion matrix

Actual

Predicted

No- pain

Low

Med

High

Intolerable

No-pain

Low

Med

High

Intolerable

To compare the results of the proposed method to state-of-the-art schemes, a few studies are reported in Table 6. The method of [36] was multivariate pattern analysis which resulted in 83% accuracy in classifying pain vs no-pain class. [38] applied classification by calculating approximate entropy, fractal dimension and Lyapunov exponent as the features and achieved 89% accuracy for the detection of pain. Wavelet higher order spectral is the feature used in [64] in detecting pain. Researchers in [39] did the study in both two- class and five-class classifications of pain by considering fractal dimension, approximate, Shannon and spectral entropies as the features. In [44], they used connectivity features in the alpha band and 86% accuracy was achieved for five classes of pain.

Table 6. Comparison of the proposed method with the state-of-the-art methods.

Reference

Number of classes

Accuracy(%)

[36]

2-class

[38]

2-class

[64]

2-class

[39]

2-class

[39]

5-class

[44]

5-class

Proposed method

5-Class

88.5

4.3. Discussion

Graph Neural Networks (GNNs) are highly advantageous for EEG analysis due to their ability to capture the intricate functional neural connectivity between various electrode locations, thereby facilitating a more precise interpretation of EEG signals [65]. GNNs possess the capability to depict the input EEG signal as a graph, with the nodes symbolizing the EEG channels, and employ a trainable weighted adjacency matrix to acquire the optimal connections between nodes [66, 67]. This methodology enables the extraction of spatio-temporal characteristics from the EEG signal, thereby enhancing the classification performance [67, 68]. Furthermore, GNNs provide neuroscientific interpretability and explainability to deep learning techniques tailored specifically for EEG-based classification problems.

Besides, using of a hierarchical classifier while applying frequency-band-wise features at each node, resulted in an an adaptive and more efficient classification. Also, the computation cost was decreased, too.

As a suggestion for future work, applying the proposed method over a bigger dataset and also proposing a method for classifying chronic pain may be good candidates.

4.4. Conclusion

In this study, an adaptive novel band-wise GCNN-based classifier was proposed to classify five levels of pain. In contrary to a few other studies done on the multi-level classification of pain, which were focused only on the alpha band, this study investigates all frequency bands in the analysis and classification. This will give the model the ability to perform well for the patients in coma or unconscious situations. Also, the adaptive manner of the proposed hierarchical classifier makes the model calculate and apply the most discriminative estimators that decrease the computation cost of the model for inference. To the best of our knowledge, this is the first time to benefit from the GCNN classifier for classifying different levels of pain and by considering the pain-specific features, as feature nodes, comprehensive information about the brain network in each pain level is extracted.

Acknowledgment: This study is supported via funding from Prince Sattam bin Abdulaziz University, project number (PSAU/2023/R/1445).

References

1. Bendinger, T. & Plunkett, N. Measurement in pain medicine. BJA Educ. 16, 310–315 (2016).

2. Fillingim, R. B., Loeser, J. D., Baron, R. & Edwards, R. R. Assessment of chronic pain: Domains, methods, and mechanisms.

J.Pain 17, T10–T20 (2016).

3. Herr, K., Coyne, P. J., McCaffery, M., Manworren, R. & Merkel, S. Pain assessment in the patient unable to self-report: Position statement with clinical practice recommendations. Pain Manag. Nurs. 12, 230–250 (2011).

4. Breivik, H. et al. Assessment of pain. Br. J. Anaesth. 101, 17–24 (2008).

5. Schnakers, C. & Zasler, N. D. Pain assessment and management in disorders of consciousness. Curr. Opin. Neurol. 20, 620–

626 (2007).

6. Arbour, C. & Gelinas, C. Behavioral and physiologic indicators of pain in nonverbal patients with a traumatic brain injury: An integrative review. Pain Manag. Nurs. 15, 506–518 (2014).

7. McGuire, D. B., Kaiser, K. S., Haisfield-Wolfe, M. E. & Iyamu, F. Pain assessment in noncommunicative adult palliative care patients. Nurs. Clin. N. Am. 51, 397–431 (2016).

8. Witt, N., Coynor, S., Edwards, C. & Bradshaw, H. A guide to pain assessment and management in the neonate. Curr. Emerg.

Hosp. Med. Rep. 4, 1–10 (2016).

9. van der Miesen, M. M., Lindquist, M. A. & Wager, T. D. Neuroimaging-based biomarkers for pain. PAIN Rep. 4, e751 (2019).

10. Mari, T. et al. Systematic review of the effectiveness of machine learning algorithms for classifying pain intensity, phenotype or treatment outcomes using electroencephalogram data. J. Pain 23, 349–369 (2022).

11. Samuel, A. L. Some studies in machine learning using the game of checkers. IBM J. Res. Dev. 3, 210–229 (1959).

12. Vu, M.-A.T. et al. A shared vision for machine learning in neuroscience. J. Neuroscience 38, 1601–1607 (2018).

13. Jordan, M. I. & Mitchell, T. M. Machine learning: Trends, perspectives, and prospects. Science (80-) 349, 255–260 (2015). [14] Bunk, S. F., Lautenbacher, S., Rüsseler, J., Müller, K., Schultz, J., & Kunz, M.. (2018). Does EEG activity during painful

stimulation mirror more closely the noxious stimulus intensity or the subjective pain sensation. 35. https://doi.org/10.1080/08990220.2018.1521790

15. Ploner, M., Sorg, C. & Gross, J. Brain rhythms of pain. Trends Cogn. Sci. 21, 100–110 (2017).

16. Afrasiabi, S., Boostani, R., Masnadi-Shirazi, M.A., A Physiological-Inspired Classification Strategy to Classify Five Levels of Pain. 26th National and 4th International Iranian Conference on Biomedical Engineering (ICBME), Tehran (Iran), pp. 1-6,

2019.

17. Afrasiabi, S., Boostani, R., Masnadi-Shirazi, M.A. Differentiation of pain levels by deploying various electroencephalogram synchronization features and a dynamic ensemble selection mechanism. Physiological Measurement 41 (11), 115004, 2020.

18. Fallon, N., Chiu, Y., Nurmikko, T. & Stancak, A. Altered theta oscillations in resting EEG of fibromyalgia syndrome patients. Eur. J. Pain 22, 49–57 (2018).

19. Michail, G., Dresel, C., Witkovsky, V., Stankewitz, A. & Schulz, E. Neuronal oscillations in various frequency bands differ between pain and touch. Front. Hum. Neurosci. 10, 182 (2016).

20. Shao, S., Shen, K., Yu, K., Wilder-Smith, E. P. V. & Li, X. Frequency-domain EEG source analysis for acute tonic cold pain perception. Clin. Neurophysiol. 123, 2042–2049 (2012).

21. Furman, A. J. et al. Cerebral peak alpha frequency predicts individual differences in pain sensitivity. Neuroimage 167, 203–

210(2018).

22. Furman, A. J. et al. Sensorimotor peak alpha frequency is a reliable biomarker of prolonged pain sensitivity. Cereb. Cortex

30, 6069–6082 (2020).

[23] F. Scarselli, M. Gori, A.C. Tsoi, M. Hagenbuchner, G. Monfardini, The graph neural network model, IEEE Trans. Neural

Netw. 20 (1) (2008) 61–80.

[24] M. Defferrard, X. Bresson, P. Vandergheynst, Convolutional neural networks on graphs with fast localized spectral filtering, Adv. Neural Inf. Process. Syst. 29 (2016) 3844–3852.

[25] Demir, A., Koike-Akino, T., Wang, Y., Haruna, M., & Erdogmus, D. (2021, November). EEG-GNN: Graph neural networks

for classification of electroencephalogram (EEG) signals. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) (pp. 1061-1067). IEEE.

[26] Vivek, S., B., V., Adarsh, A., Gubbi, J., Muralidharan, K., Ramakrishnan, R. K., & Pal, A. (2022). ST-GNN for EEG Motor

Imagery Classification. https://doi.org/10.1109/BHI56158.2022.9926806

[27] Tagging EEG features within exam reports to quickly generate databases for research purposes. (2023). Computer Methods and Programs in Biomedicine, Volume 242, December 2023, 107836

[28] Zis, P., Liampas, A., Artemiadis, A., Tsalamandris, G., Neophytou, P., Unwin, Z., Kimiskidis, V. K., Hadjigeorgiou, G. M., Varrassi, G., Zhao, Y., & Sarrigiannis, P. G. (2022). EEG Recordings as Biomarkers of Pain Perception: Where Do We Stand

and Where to Go? Pain and therapy, 11(2), 369–380.

[29] S. Shao, K. Shen, K. Yu, E.P. Wilder-Smith, X. Li, Frequency-domain EEG source analysis for acute tonic cold pain perception, Clin. Neurophysiol. 123 (10) (2012) 2042–2049.

[30] A.C. Chen, P. Rappelsberger, O. Filz, Topology of EEG coherence changes may reflect differential neural network activation in cold and pain perception, Brain Topogr. 11 (2) (1998) 125–132.

[31] S. Ferracuti, S. Seri, D. Mattia, G. Cruccu, Quantitative EEG modifications during the cold water pressor test: hemispheric

and hand differences, Int. J. Psychophysiol. 17 (3) (1994) 261–268.

[32] A.C. Chen, S.F. Dworkin, J. Haug, J. Gehrig, Topographic brain measures of human pain and pain responsivity, Pain 37 (2) (1989) 129–141.

[33] D. Fattahi, B. Nasihatkon, R. Boostani, A general framework to estimate spatial and spatio-spectral filters for EEG signal

classification, Neurocomputing 119 (2013) 165–174.

[34] M. Gram, C. Graversen, S.S. Olesen, A.M. Drewes, Dynamic spectral indices of the electroencephalogram provide new insights into tonic pain, Clin. Neurophysiol. 126 (4) (2015) 763–771.

[35] W. Peng, C. Babiloni, Y. Mao, Y. Hu, Subjective pain perception mediated by alpha rhythms, Biol. Psychol. 109 (2015) 41–

150.

[36] E. Schulz, A. Zherdin, Decoding and individual‘s sensitivity to pain from the multivariate analysis of EEG data, Cereb

Cortex 22 (5) (2012) 1118–1123.

[37] P. Panavaranan, Y. Wongsawat, EEG-based pain estimation via fuzzy logic polynominal kernel support vector machine, in:

6th Biomedical Engineering International Conference, 2013, pp. 1–4.

[38] M. Vatankhah, A. Toliyat, Pain level measurement using discrete wavelet transform, Int. J. Eng. Technol. 8 (5) (2016) 380. [7] L.J. Hadjileontiadis, EEG-based tonic cold pain characterization using

[39] T. Nezam, R. Boostani, V. Abootalebi, K. Rastegar, A novel classification strategy to distinguish five levels of pain using the EEG signal features, IEEE Trans. Affective Comput. (2018) 1–9.

[40] S. Afrasiabi, R. Boostani, M.A. Masnadi-Shirazi, T. Nezam, An EEG based hierarchical classification strategy to differentiate five intensities of pain, Expert Syst. Appl. 180 (2021) 115010180.

[41] Kav?i?, A., Demsar, J., Georgiev, D., Pe?ari? Megli?, N., & Soltirovska Salamon, A.. (2023). EEG functional connectivity after perinatal stroke.. https://doi.org/10.1093/cercor/bhad255

[42] M.P. Van Den Heuvel, H.E.H. Pol, Exploring the brain network: a review on resting state fMRI functional connectivity, Eur. Neuropsycho-Pharmacol. 20 (8) (2010) 519–534.

[43] E.A. Necka, I.S. Lee, A. Kucyi, J.C. Cheng, Q. Yu, L.Y. Atlas, Applications of dynamic functional connectivity to pain and its modulation, Pain Reports 4 (4) (2019).

[44] P. Modares-Haghighi., R, Boostani., M, Nami., S, Sanei. (2021). Quantification of pain severity using EEG-based functional connectivity. Biomedical Signal Processing and Control, doi: 10.1016/J.BSPC.2021.102840

[45] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, S.Y. Philip, A comprehensive survey on graph neural networks, IEEE Trans. Neural Netw. Learn. Syst. 32 (1) (2020) 4–24.

[46] Brain network classification based on dynamic graph attention information bottleneck. (2023, December 16). Brain network classification based on dynamic graph attention information bottleneck. Computer Methods and Programs in Biomedicine.

[47] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, M. Sun, Graph neural networks: A review of methods and applications, AI Open 1 (2020) 57–81.

[48] Y. Yin, X. Zheng, B. Hu, Y. Zhang, X. Cui, EEG emotion recognition using fusion model of graph convolutional neural networks and LSTM, Appl. Soft Comput. 100 (2021) 106954.

[49] P. Zhong, D. Wang, C. Miao, EEG-based emotion recognition using regularized graph neural networks, IEEE Trans. Affect. Comput. 13 (3) (2020) 1290–1301.

[50] X. Lun, S. Jia, Y. Hou, Y. Shi, Y. Li, H. Yang, S. Zhang, J. Lv, GCNs-net: a graph convolutional neural network approach for decoding time-resolved eeg motor imagery signals, 2020, arXiv preprint arXiv:2006.08924.

[51] Z. Wang, Y. Tong, X. Heng, Phase-locking value based graph convolutional neural networks for emotion recognition, IEEE Access 7 (2019) 93711–93722.

[52] M. Jalili, M.G. Knyazeva, Constructing brain functional networks from EEG:partial and unpartial correlations, J. Integr. Neurosci. 10 (02) (2011) 213–232.

[53] M.G. Rosenblum, A.S. Pikovsky, J. Kurths, Phase synchronization of chaotic oscillators, Phys. Rev. Lett. 76 (11) (1996)

1804.

[54] J.-P. Lachaux, E. Rodriguez, J. Martinerie, F.J. Varela, Measuring phase synchrony in brain signals, Hum. Brain Mapp. 8 (4) (1999) 194–208.

[55] C.J. Stam, G. Nolte, A. Daffertshofer, Phase lag index: assessment of functional connectivity from multi channel EEG and

MEG with diminished bias from common sources, Hum. Brain Mapp. 28 (11) (2007) 1178–1193.

[56] Mamcarz, K.. (2020). Analysis of Granger causality between gold and selected financial assets. 64(7). https://doi.org/10.15611/PN.2020.7.06

[57] C.W. Granger, Investigating causal relations by econometric models and cross-spectral methods, Econometrica (1969) 424–

438.

[58] A.K. Seth, A.B. Barrett, L. Barnett, Granger causality analysis in neuroscience and neuroimaging, J. Neurosci. 35 (8) (2015)

3293–3297.

[59] B. He, L. Astolfi, P.A. Valdés-Sosa, D. Marinazzo, S.O. Palva, C.-G. Bénar, C.M. Michel, T. Koenig, Electrophysiological brain connectivity: theory and implementation, IEEE Trans. Biomed. Eng. 66 (7) (2019) 2115–2137.

[60] Pain, S., Roy, S., Sarma, M., & Samanta, D.. (2023). Detection of alcoholism by combining EEG local activations with brain

connectivity features and Graph Neural Network. 85. https://doi.org/10.1016/j.bspc.2023.104851

[61] T.N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, 2016, arXiv preprint arXiv:1609.02907.

[62] F.R. Chung, F.C. Graham, Spectral Graph Theory, Vol. 92, American Mathematical Soc., 1997.

[63] Z. Wang, Y. Tong, X. Heng, Phase-locking value based graph convolutional neural networks for emotion recognition, IEEE Access 7 (2019) 93711–93722.

[64] L.J. Hadjileontiadis, EEG-based tonic cold pain characterization using wavelet higher-order spectral features, IEEE Trans. Biomed. Eng. 62 (8) (2015) 1981–1991.

[65] Demir, A., Koike-Akino, T., Wang, Y., Haruna, M., & Erdogmus, D. (2021, November). EEG-GNN: Graph neural networks for classification of electroencephalogram (EEG) signals. In 2021 43rd Annual International Conference of the IEEE

Engineering in Medicine & Biology Society (EMBC) (pp. 1061-1067). IEEE.

[66] Li, H., Ji, H., Yu, J., Li, J., Jin, L., Liu, L., ... & Ye, C. (2023). A sequential learning model with GNN for EEG-EMG-based stroke rehabilitation BCI. Frontiers in Neuroscience, 17, 1125230.

[67] Vivek, B. S., Adarsh, A., Gubbi, J., Muralidharan, K., Ramakrishnan, R. K., & Pal, A. (2022, September). ST-GNN for EEG

Motor Imagery Classification. In 2022 IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI) (pp.

01-04). IEEE.

[68] Lin, Y. C., Zhang, B., & Prasanna, V. (2022, February). Hp-gnn: Generating high throughput gnn training implementation

on cpu-fpga heterogeneous platform. In Proceedings of the 2022 ACM/SIGDA International Symposium on Field-Programmable

Gate Arrays (pp. 123-133).