QuadTPat: Quadruple Transition Pattern-based explainable feature engineering model for stress detection using EEG signals

Abstract

The most cost-effective data collection method is electroencephalography (EEG), which obtains meaningful information about the brain. Therefore, EEG signal processing is crucial for neuroscience and machine learning (ML). Therefore, a new EEG stress dataset has been collected, and an explainable feature engineering (XFE) model has been proposed using the Directed Lobish (DLob) symbolic language. The first phase of this research is the data collection phase, and an EEG stress dataset was gathered from 310 participants. This collected stress dataset contains two classes: (i) stress and (ii) control. An XFE model has been presented to detect stress automatically. The presented XFE model has four main phases, and these are (i) channel transformer and quadruple transition pattern (QuadTPat)-based feature generation, (ii) feature selection deploying cumulative weighted neighborhood component analysis (CWNCA), (iii) explainable results creation with DLob and (iv) classification with t algorithm-based k-nearest neighbors (tkNN) classifier. The proposed XFE model generates a DLob string, and the explainable results were obtained using this string. Moreover, the presented XFE model attained 92.95% and 73.63% classification accuracy, deploying 10-fold and leave-one subject-out (LOSO) cross-validations (CV). According to the classification performances, the recommended QuadTPat-based XFE is a good model for EEG signal classification. Also, the presented QuadTPat-based XFE model is a good model for explainable artificial intelligence (XAI) since TTPat-based XFE is cooperating with the DLob.

Introduction

Stress is a condition experienced by individuals due to various factors such as work pressure, personal problems, or environmental changes^1,2. Understanding and detecting stress is very important because it significantly affects mental and physical health³. Stress detection involves determining the presence and level of stress in an individual⁴. There are various methods used for stress detection. These methods can be broadly divided into physiological, psychological, and behavioral approaches^5,6. Physiological methods involve monitoring changes in the body due to stress, such as heart rate, blood pressure, and skin conductance^7,8. Psychological methods focus on assessing an individual’s emotional and mental state through surveys and interviews^4,9. Behavioral techniques can be defined as observing changes in an individual’s behavior, such as speech patterns, facial expressions, and physical activity levels¹⁰.

In recent years, technology has played an important role in developing stress detection methods¹¹. Wearable devices such as smartwatches and fitness trackers include sensors that can continuously monitor physiological signals¹². Additionally, machine learning algorithms are used to analyze large amounts of data collected from these devices to detect stress levels¹³ accurately. Mobile applications can also help individuals monitor their stress levels and provide recommendations for stress management^14–16.

Brain-Computer Interfaces (BCIs) have also emerged as a tool in stress detection and management^17,18. BCIs enable direct communication between the brain and external devices¹⁹, allowing for real-time monitoring of brain activity²⁰. By analyzing EEG signals^20–22, BCIs can detect stress-related neural patterns, offering a deeper understanding of an individual’s stress levels. This technology enhances the accuracy of stress detection and opens up new possibilities for personalized stress management interventions²³.

Early detection of stress can help prevent serious health problems such as anxiety, depression, and cardiovascular disease²⁴. It also allows individuals to take proactive measures to manage stress, such as practicing relaxation techniques, participating in physical activities, and seeking professional help^25,26.

Literature review

Some current studies on machine learning-based stress detection in the literature are presented as follows. Al-Alim et al.²⁷ proposed an approach for stress detection using wearable sensors in free-living environments. They utilized the physiological data of 240 subjects. The random forest model achieved the highest accuracy of 98.29% for binary classification, and XGBoost showed the best performance in three-level classification with an accuracy of 98.98%. Nazeer et al.²⁸ presented an enhanced method for stress detection utilizing bio-sensor technology. Their study employed the WESAD dataset. WESAD dataset includes data from 16 subjects collected using wrist-worn and chest-worn sensors. In their study, the XGBoost classifier demonstrated superior performance, achieving an accuracy of 93.65% and an F1 score of 90.1% for three-class classification and 96.12% accuracy for two-class classification using wrist sensors. ​Mathur et al.²⁹ suggested a method for nurse stress detection using body sensors. They collected from 15 nurses data. Their study achieved a weighted F1 score of 99.00 for two- and three-level stress classifications with a decision tree. Migovich et al.³⁰ developed a method for detecting stress in autistic adults during simulated job interviews. Their study utilized data from 15 autistic young adults collected through wearable. They attained the best accuracy of 84.8% for individual models and a group accuracy of 75.4% using leave-one-out cross-validation. Richer et al.³¹ proposed a method for detecting acute psychosocial stress by analyzing body posture and movements. Their study used motion data collected from 59 individuals. They achieved classification accuracies of 75.0% in the pilot study and 73.4% in the main study, demonstrating the potential for using body movements as stress indicators​. Awada et al.³² proposed an analysis of stress appraisal in the workplace. Their study utilized physiological and behavioral data. They obtained the highest prediction accuracy of 82.78% combining physiological and behavioral features with XGBoost. Adarsh and Gangadharan³³ presented an approach to achieve real-time stress assessment. Their study utilized WESAD and SWELL datasets. Their proposed method achieved accuracy rates of 97.75% and 94.48% on the WESAD and SWELL datasets. Vries et al.³⁴ proposed a real-time stress detection system. Their study utilized a multimodal dataset collected from wearable sensors. Their dataset included data from participants with and without ID engaged in various activities to induce stress and relaxation. They achieved a balanced accuracy of 73% across all groups, outperforming the general model, which had an accuracy of 60%. Fontes et al.³⁵ suggested a real-time stress detection system leveraging artificial intelligence. Their study utilized a diverse dataset of physiological signals collected from wearable sensors during daily activities. They achieved an accuracy rate of 91.3% compared to 84.7% for detecting stress levels. Hadhri et al.³⁶ developed a stress detection approach utilizing a voting ensemble classifier. Their research used a dataset containing physiological measurements. They achieved a notable classification accuracy of 78% with the soft voting classifier.

Literature gaps

According to the surveyed literature, the identified gaps are:

• In the literature, well-known feature engineering and deep learning models have generally been used to get high classification performances. This situation causes stagnation for the new generation models proposing.

• Most of the models have focused on classification ability. There are a limited number of explainable artificial intelligence (XAI) models in the EEG signal classification.

Motivation and our model

Our essential motivation is to fill the given literature gaps and proposing a new generation XFE model. Moreover, a new version of the Directed Lobish (DLob)^37,38 has been presented in this research.

The first literature gap is the lack of new EEG signal processing models. In this research, the transformers have inspired us³⁹ since transformers caused a revolution in machine learning (ML). Therefore, we have proposed a channel transformer, the output of which is divided into four layers. Using these layers, we have created an automaton to extract features and demonstrate the relations of these layers. An effective feature extractor should be used after transforming the EEG signals to the channel values. As the literature states, histogram-based feature extractors like local binary pattern (LBP)⁴⁰ are effective feature extractors. Therefore, a new histogram-based feature extraction function has been proposed, and this feature extraction function is called the Quadruple Transition Pattern (QuadTPat).

Moreover, the second innovation of this model is the Cumulative Weight-based Neighborhood Component Analysis (CWNCA)⁴¹, and this feature selector solves the problem of selecting the optimal number of features by computing cumulative weights. In the classification phase, t algorithm-based k-nearest neighbors (tkNN)³⁷ has been used. In this aspect, we have proposed an innovative model has been presented for EEG signal classification.

The second literature gap is the limited XAI models. DLob was used to get explainable results. Moreover, hemisphere transitions, the four-node automaton transitions, and channel information have been used to give explainable results.

According to the above information, our motivations are:

• To bridge the gap in new generation EEG signal processing models by introducing an innovative channel transformer and automaton-based feature extraction method (QuadTPat).

• To advance XAI by improving the DLob symbolic language and using graph-based presentations, contributing to the broader understanding of neural connections and connectome theory.

Innovations

The novelties of this work are:

• A new feature extractor, QuadTPat, has been proposed to extract features with high classification capability.

• A new generation feature selector, CWNCA, has been used to select the most informative features. This research is among the pioneering studies demonstrating the feature selection capability of CWNCA.

• An innovative automaton-based explainable results have been presented in this research.

Contributions

• By proposing a QuadTPat-based XFE model, we have presented a high accurate EEG signal classification model. In this aspect, the recommended model contributes to EEG signal classification.

• By cooperating with DLob, the explainable results have been given. In this aspect, the proposed model presents valuable information about neurosciences.

• The recommended model not only contributes to ML but also contributes to neuroscience.

Methods

The collected dataset

This study was approved by the Non-Invasive Ethics Committee, Firat University, with protocol number December 10, 2023 (2023/14–18). All methods were carried out in accordance with relevant guidelines and regulations, including the Declaration of Helsinki for human subjects. Informed consent was obtained from all participants, and privacy and data protection laws were strictly adhered to, ensuring the confidentiality and anonymity of all data collected. All authors declare no conflicts of interest, and funding sources are fully disclosed, with acknowledgments to the supporting institutions.

Participants were selected from individuals affected by the February 6, 2023, Turkey earthquake series, focusing on those who had experienced earthquake-related trauma. The volunteers were university students residing in the affected earthquake zones during the event. To induce stress, a 3-minute real earthquake video was shown repeatedly, which included footage of the earthquake, immediate reactions of people, and the collapse of buildings. This approach was chosen to capture moments of heightened stress authentically. The participants’ stress levels were determined based on their self-reports after the video exposure, as no additional stress tests were conducted at the end of the experiment.

EEG signals were collected from 310 participants (42 women and 268 men, aged 18–65), resulting in a diverse dataset that reflects a wide range of individual responses to stress. Of these participants, 150 reported experiencing stress, while 160 reported no stress. The EEG data was recorded using the Emotiv Epoch X 14-channel brain cap with a 128 Hz sampling frequency, designed according to the 10/20 system. This system collected EEG data from the Frontal, Temporal, Parietal, and Occipital lobes.

The decision to develop a proprietary dataset instead of using a public database was driven by the need to capture stress responses in a context directly relevant to the study’s objectives. Public stress datasets typically do not encompass stressors of the magnitude and specificity required for this research, particularly those related to large-scale natural disasters like the Turkey earthquake series. As far as we know, there is no existing public stress dataset related to disasters, making our dataset unique and invaluable for future research.

By creating this customized dataset, we were able to ensure that the data collected was directly relevant to the context of earthquake-induced stress, thus increasing the accuracy and appropriateness of the data for the specific context of the study. Although the creation of this dataset required additional time and resources, the resulting data’s relevance and quality significantly enhance the study’s potential to contribute to the field of stress detection and neuropsychological analysis.

This dataset is innovative in several ways:

• The dataset was collected from 310 participants, ensuring a diverse sample that enhances the generalizability of the findings.

• The stress trigger used in this study was a significant real-world event—Turkey’s February 6, 2023, serial earthquakes. This approach not only ensures the authenticity of the stress responses but also highlights the relevance of the dataset to addressing real-world problems through AI technologies.

• To our knowledge, no public stress dataset exists that focuses on disasters. This dataset fills a critical gap, providing a foundation for future research aimed at solving stress-related problems in disaster scenarios using AI technologies.

After removing artifact-laden segments from the collected EEG signals, each signal was divided into 15-second segments, resulting in 3,667 EEG segments. The artifacts removed included significant movement artifacts, such as sudden jerks, standing up, and crying episodes, which severely disrupted the EEG signal quality. These artifacts were carefully identified and excluded to ensure the integrity and accuracy of the data used for analysis. Of the resulting segments, 1,785 were labeled as stress, while 1,882 represented the control class. A visual representation of the EEG collection process is provided in Fig. 1.

Figure 1.

A sample image of the EEG signal collection process.

The proposed method

In this model, we have proposed a new generation XFE model, and the recommended XFE model is named the QuadTPat-based XFE model. The primary objective of the proposed model is to get high and robust classification performance on the collected EEG stress dataset and present interpretable results about post-earthquake stress. The methods used to create the recommended QuadTPat-based XFE model are:

• Channel transformer: The channel transformer transforms the EEG signal to identify the channels by using the amplitude of the EEG signal.

• QuadTPat: After applying the channel transformer to the EEG signal, the features have been extracted by deploying QuadTPat. The proposed QuadTPat extract features deploying four leveled transformed signals. It is a transition table-based feature extractor. Therefore, the presented QuadTPat can detect differences.

• CWNCA: Cumulative weights have cooperated with this feature selector to choose the optimal number of features automatically.

• DLob: It is a symbolic language to generate explainable results. Moreover, it simplifies the interpretability of the EEG signals.

• tkNN: This classifier is a new ensemble version of the kNN classifier and is an iterative classifier. Additionally, iterative majority voting (IMV) has been utilized to increase the classification results.

The overview of the QuadTPat-based XFE model is graphically demonstrated in Fig. 2.

Figure 2.

The schematic overview of the recommended QuadTPat-based XFE model.

Moreover, the QuadTPat-based XFE model’s phases are also explained below.

Phase 1.1: Apply channel transformation to EEG signals to generate signal values for channel indexes.

1

where : transformer signal and : channel transformer function.

Phase 1.2: Extract features from the transformed signal by deploying QuadTPat.

2

where : feature vector and : QuadTPat feature extraction function.

Phase 2: Choose the most meaningful features for deploying CWNCA.

3

where : selected/chosen feature vector and : actual label.

Phase 3: Convert the indexes of the selected features to the DLob string and analyze the string obtained to present explainable results.

Phase 4: Classify the features selected by applying the tkNN classifier to present classification results.

4

More details about the methods used in these phases are given below.

Channel transformer

Channel transformer is a simple method, and the primary objective of this method is to convert EEG signals to channel index. This model highlights channels instead of EEG values/amplitudes because the channels’ importance is more valuable than amplitude in spite of neuroscience. Therefore, we have used this transformation in this research. This transformation has mathematically been defined as below.

5

L2

Herein, : the qualified values, : the transformed signal, : EEG signal, : sorting function, : number of channels and : length of each channel of the EEG signal. In this work, an EEG signal dataset with 14 channels has been used.

Quadruple transition pattern

The theoretical background of the proposed QuadTPat is to get features from the different leveled transformed signals by deploying transition table extraction. This model is also an automaton-based feature extraction function, which is a simple method. To better clarify QuadTPat, the steps are:

• Step 1: Create four levels by using the transformed signal.

6

L2

L3

L4

L5

Herein, four levels of bands have been created, and these bands have been utilized as nodes of the below automaton to extract features. The used automaton is depicted in Fig. 3.

Figure 3.

The used automaton with four nodes to extract features.

The given automaton with four nodes has showcased the connections, with 16 (= 4 × 4) connections in this automaton. Moreover, the given automaton can be shown as a fully connected bipartite graph. Each connection defines each transition table. Thus, 16 transition tables have been computed. This process is also explained in Step 2.

• Step 2: Compute 16 transition tables by using the four-node automaton shown in Fig. 3. The mathematical explanations of this step are:

Firstly, 14 × 14 (14 channels in the collected dataset) sized 16 transition tables have been defined.

7

Herein, : transition table with a size of 19 × 19 since the used dataset has 19 channels.

The transition tables defined above have been completed by deploying the mathematical equations below.

8

L2

L3

L4

L5

L6

L7

L8

L9

L10

L11

L12

L13

L14

L15

L16

L17

L18

The defined transition tables have been filled out by deploying the process above.

• Step 3: Normalize the computed transition tables by deploying z-score normalization.

9

Herein, : mean of the transition table, : standard deviation of the transition table, and : the normalized transition table.

• Step 4: Transform the computed transition tables to vectors.

10

where : g^th vector with a length of 196 (= 14 × 14).

• Step 5: Concatenate the generated vectors to create a feature vector.

11

where : the computed feature vector with a length of 3136 (= 14 × 14 × 16)

Cumulative weight-based neighborhood component analysis

The CWNCA feature selector is the developed version of the NCA⁴² feature selector. CWNCA aimed to solve the problem of selecting the optimal number of features by deploying cumulative weight. The steps of this feature selector are:

• S1: Produce the indices sorted by deploying the NCA feature selector.

12

Herein, : the indices sorted,: the generated weights of the features, : NCA feature selection function, : feature matrix and : actual/real outcome.

• S2: Calculate the optimal number of the features deploying cumulative weight.

13

L2

Here, : cumulative weight, : the length of the features, and : the optimal number of features. For this research, the optimal number of features has been calculated as 180. Therefore, the most informative 180 features have been selected.

• S3: Select the most informative 180 features as a chosen feature vector.

14

L2

where : selected feature vector, : number of the EEG observations.

The three steps (S1-S3) that are given have been defined by the proposed CWNCA feature selector.

Directed Lobish

DLob is a symbolic language used as an XAI method for brain-related data especially for EEG signals. We have used features of this XAI method (DLob). By using the selected indexes, the DLob symbols are generated, and these symbols create a DLob sentence. The DLob sentence generation algorithm is showcased in Algorithm 1.

By using Algorithm 1, the DLob string was computed. To extract meaningful information from this string, the meaning of the DLob letters should be known, as shown in Table 1.

Table 1.

The meaning of the DLob letters used.

Letter	Meaning	Letter	Meaning
FL	Crucial for logical thinking, planning, and decision-making. Associated with speech production (e.g., Broca’s area) and sequential thought processes.	FR	Involved in creative thinking, spatial orientation, and non-verbal communication. Plays a role in attention, impulse control, and emotional regulation.
TL	Represents language comprehension and designing the content of the speech (e.g., Wernicke’s area), verbal memory and understanding spoken and written language.	TR	Represents processing sounds, including language and music. Plays a role in recognizing faces and objects.
PL	Associated with processing language and mathematical operations. Plays a role in understanding and producing speech, as well as in reading and writing.	PR	Important for spatial awareness and navigation. Involved in recognizing patterns, shapes, and the positions of objects in space.
OL	Processing visual information from the right visual field. Involved in recognizing letters, words, and other visual stimuli related to language.	OR	Processes visual information from the left visual field. Involved in recognizing faces, scenes, and spatial orientation of objects.

The meaningful DLob string(s) have been created using the given meaning. Moreover, we used information entropy, a histogram of the channels and letters used to give explainable results. Also, the connections used by the defined automaton (see Fig. 3) have been computed. More explainable results have been presented by computing transition tables of this information.

The t algorithm-based k nearest neighbors

A new generation ensemble classifier, tkNN, was used to get the maximum classification performance from the selected feature vector. The tkNN used an iteration, IMV algorithm, and greedy to get the output with maximum classification accuracy.

• S1: Deploy iterative parameters to obtain parameter-based outcomes.

15

L2

L3

Herein, : the parameters-based outcome of the kNN⁴³, : kNN classifier, : k values, : voting weights of the kNN classifier and : the used distance metric. In this research, 10 k values, three distance metrics and three voting weights have been used in the tkNN classifier. Thus, this classifier generates 90 (= 10 × 3 × 3) parameters-based outcomes.

• S2: Apply IMV and create voted outcomes.

16

L2

L3

where : classification accuracy, : the sorted indexes of the classification accuracies by descending,: classification accuracy computation function, : mode function and : voted outcomes by generating IMV. In this step, 88 (= 90 − 3 + 1) voted outcomes have been produced by IMV since the range of the loop used is from 3 to 90.

• S3: Choose the best 0.

17

L2

L3

Herein, : final outcome and : the index of the maximum accuracy.

The recommended tkNN classifier offers several advantages, as demonstrated by its algorithmic steps (see S1-S3). Unlike traditional kNN with its static parameters, tkNN explores a diverse range of parameter combinations (k-values, distance metrics, and weights). This iterative exploration allows for a more comprehensive optimization process. The tkNN algorithm achieves this by first generating multiple predictions through these parameter variations. Following this, a step referred to as IMV combines these predictions, creating a set of voted outcomes. Finally, tkNN uses a greedy algorithm to select the optimal outcome from these parameter-based and voted outcomes. This classifier has increased the robustness of the kNN and classification results.

In this phase, a 10-fold CV and LOSO CV have been utilized to obtain reliable results.

Results

The presented QuadTPat-based model was developed using MATLAB (version 2024a) through function-based programming. The XFE model is lightweight, allowing it to be implemented on a simply configured laptop in CPU mode, without the need for expensive hardware like a graphical processing unit (GPU). The proposed XFE model is a feature engineering model that employs functional programming, with the functions coded as .m files. This model includes the following functions: (i) main, (ii) channel transformer, (iii) QuadTPat, (iv) CWNCA, (v) tkNN, (vi) DLob sentence generator, and (vii) DLob result generator. By utilizing these functions, the proposed model was applied. Detailed descriptions of the QuadTPat-based XFE model have been provided, enabling researchers to easily develop and extend this model using the shared information. The initial parameters of the proposed model are outlined below:

Feature extraction

• Channel Transformer: A sorting function was used to order the channel identities of the EEG signals in descending order. In this work, a 14-channel EEG signal dataset was used. As a result, the length of the transformed signal is 14 times larger than the original EEG signal.

• QuadTPat: By applying QuadTPat, four layers of the transformed signals were created, with these layers containing 3, 4, 4, and 3 values, respectively, for each of the 14 channels. A fully connected automaton was then employed to extract features. During the feature extraction phase, a transition table was utilized, resulting in the creation of 16 feature vectors, each with a length of 196 (= 14 × 14). In the final phase, these features were merged to form a feature vector with a length of 3,136 (= 14 × 14 × 16).

Feature selection.

• CWNCA: The cumulative weight-based number of the selected features determination has been applied to NCA feature selector.

Classification

• tkNN: By varying the k-value (1–10), distance metrics (City Block, Euclidean, Cosine), and weights (Inverse, Equal, Squared Inverse), 90 (= 10 × 3 × 3) parameter-based outcomes were generated. Through iterative majority voting (IMV) with a range from 3 to 90, an additional 88 (= 90 − 3 + 1) voted outcomes were created. The outcome with the maximum classification accuracy was selected as the final outcome using a greedy algorithm.

Explainable AI (XAI)

• Directed Lobish: This symbolic language consists of 8 symbols, with the meaning of each symbol explained in Table 1. These symbols were generated by utilizing the indexes of the selected features. Each selected feature contains two letters, from which the DLob symbols were extracted. Subsequently, the transition table and histograms of these symbols were generated to obtain explainable results. Additionally, by employing the indexes of these selected features, the transitions of the used automaton were computed.

Firstly, the classification performance of the proposed QuadTPat-based XFE model has been measured. To evaluate the classification performance of this model, we have used classification accuracy, sensitivity, specificity, precision, F1-score, and geometric mean to give comprehensive results. We have also used two validation techniques, a 10-fold CV and an LOSO CV. The computed results by deploying these validation techniques have been shown in Table 2.

Table 2.

The classification results (%) of the recommended QuadTPat-based XFE model deploying 10-fold CV and LOSO CV.

Performance evaluation metric	10-fold CV	LOSO CV
Accuracy	92.94	73.63
Sensitivity	93.00	73.73
Specificity	92.88	73.54
Precision	92.53	72.55
F1-score	92.76	73.14
Geometric mean	92.94	73.63

Table 2 demonstrated that the recommended QuadTPat-based XFE model attained 92.94% and 73.63% classification accuracies by deploying 10-fold CV and LOSO CV techniques, respectively. Moreover, the confusion matrices of these results have been demonstrated in Fig. 4.

Figure 4.

The confusion matrices computed by deploying the QuadTPat-based XFE model. In these confusion matrices, the labels are 1: Stress and 2: Control (No stress). (a) 10-fold CV. (b) LOSO CV.

The confusion matrices showcased in Fig. 4 validated the computed results in Table 1.

Using DLob, we have generated a string to get explainable results, which is demonstrated below.

“FRFLFLFLOLPRTRFRTRFRTLPLTLPLPRPLTLTRPLFLFLOLFRFLFLFRFLFLFLFLPLOLPLOLFLPLTLTLFLTRPRFLFRPRORFLFRPLTLFRTLPLTLPLFROROLFRTRPRORFRFRFLFRFRPLPRFLPRORTLFLFRTRFLPLOROLOLORFRFRPRFRFRTRFRFLFLFLFLFLFLFLTLTROLPLFRFROROLFRFLFLFLFLFRPLTLFRFLTLFLFLORFRFRPRFROROLFRFRFRPRPRORTRTRTLFLFRFROLPLFRFLFRTRFLPLOLPLOLORFLFLPRFLTLFLTLFRFRTRFLFLFRFLFRTRFLFRFLFLFRFRTLFLFRTLPLORPRFLFLFLTLPLFLFROLPRFLFLTROLOROLOROLPLTLORTLOLTRTRPRFRTRFRFLFRFRFLTLFRPLTLFLOLOLFRPRFRFRTRFLPLFLOROLTLTRPRORFRTLFLFLPLFLPRORFLOLOROLORFRTRPRFRFLFRFRFLFRTLFLFRFLFLTLPRORPLTLFLFLPLFLFRFRFLFRFLFRFRFLFLFLFLORFROLFRFRORFRFLFLFRFLFLFLFLFRFRTRFLFRPRORFRTRFLPRFLFLTRORFRFRPROLFLFRORFLTRPRFLPLTRPROLFLFRPRFRFRTROLFLFLORORFLFLPRORTROLFLFRORFLPLTLTLFRFRFRFRTLPLPLPLOLFLFLORORTRPRPR”.

According to the generated string above, the obtained stress detection comments are explained below.

The DLob string obtained for stress detection reveals dominant frontal and temporal lobe activities. The frequent occurrence of FL and FR indicates a high degree of activity in the frontal lobe. The frequent occurrence of FL indicates that the frontal left region’s logical thinking, planning, decision-making, and speech production are frequently activated. Similarly, the frequent use of FR emphasizes the role of the frontal right region in creative thinking, spatial orientation (due to the earthquake), and emotional regulation (possibly called family anxiety).

The frequent activity of TL indicates its involvement in language comprehension, production, and verbal memory, while TR highlights processing sounds, music, and recognizing faces and objects. We describe the reason why these lobes are activated as the need to hear the noises of the earthquake and other individuals.

Although the parietal and occipital lobes (PL, PR, OL, and OR) appear less frequently than the frontal and temporal lobes, they still play an important role. While PL and PR are involved in spatial awareness, navigation (escape plan construction), and mathematical operations, OL and OR are visual processing regions that play a role in the recognition of visual stimuli, faces, and spatial orientation.

Transition models reveal common transitions between FL and FR, demonstrating interactions between logical/creative thinking and planning/emotional regulation. Transitions from FL to TL and TR indicate a transition from logical thinking to language processing and auditory recognition, showing that sounds are perceived and the brain performs a planning function according to the perceived sounds. While PL and PR transitions indicate spatial and pattern recognition activities, OL and OR transitions indicate visual processing and spatial orientation tasks.

High activity and transitions in the frontal and temporal lobes may also be evidence that the individual’s emotional state is impaired due to increased activity in logical and creative thinking, emotional regulation, language processing, and auditory recognition activities often associated with stress responses. The resulting string suggests a complex interplay of brain cognitive, emotional, and sensory processes during stress detection.

Moreover, the transition table of the DLob symbols, histogram of these symbols, transitions of the hemisphere, and transitions of the used automaton have been showcased in Fig. 5 to present explainable results.

Figure 5.

The computed explainable results. (a) Transition table of the DLob symbols. (b) Frequencies of the DLob symbols. (c) Transition table of the DLob symbols. (d) The number of the selected features according to connections.

Figure 6.

The classification performances of the DT: Decision Tree, LD: Linear Discriminant, SVM: Support Vector Machine, kNN: k-nearest neighbors, EkNN: Ensemble kNN, ANN: Artificial Neural Network and KLR: Kernel Logistic Regression classifiers with 10-fold CV.

In Fig. 5, all explainable results are given, and these results are discussed in the discussion section.

Discussion

In this research, we proposed an XFE model that contributes to both machine learning and neuroscience. The proposed model achieved classification accuracies of 92.94% using 10-fold cross-validation (CV) and 73.63% using leave-one-subject-out cross-validation (LOSO CV). Additionally, we used sensitivity, specificity, precision, F1-score, and geometric mean metrics to comprehensively evaluate the classification results, which were further validated by presenting the corresponding confusion matrices.

Through the deployment of DLob and feature analysis, we provided explainable results. The stress parameter used in this study is earthquake stress. According to the generated DLob string, the frontal lobe exhibited high activity. Moreover, there was a significant transition from the frontal lobe (FL) to the temporal lobe (TL) and the right temporal lobe (TR). These transitions indicate a shift from logical thinking processes to language and auditory processing.

In order to show the high classification ability of the proposed QuadTPat-based XFE model, we have presented comparative results in Table 3.

Table 3.

The comparative results.

Study	Method	XAI	Data	Class	Split ratio	Result(s)%
44	Perceived stress scale	No	206 participants	2	10-fold CV	Acc: 85.71
45	2D-CNN, MLP	No	56 participants	2	60:40	Acc: 82.00
46	CNN fusion	No	1. 58 participants (ASCERTAIN) 2. 59 participants (CLAS)	1. 2 2. 2	1. 42 training, 16 testing 2. 43 training, 16 testing	1. Acc: 85.00 F1: 85.00 2. Acc: 80.30 F1: 82.00
47	Statistical feature, kNN	No	46 participants	2	10-fold CV	Acc: 87.56 F1: 87.60 Pre: 87.60
48	Genetic algorithm, kNN	No	110 participants	2	NA	Acc: 89.61 F1: 90.52 Pre: 89.79 Rec: 91.27
Our study	QuadTPat	Yes	310 participants	2	1. 10-fold CV 2. LOSO	1. Acc: 92.94 Sen: 93.00 Spe: 92.88 Pre: 92.53 F1: 92.76 GM: 92.94 2. Acc: 73.63 Sen: 73.73 Spe: 73.54 Pre: 72.55 F1: 73.14 GM: 73.63

CNN: convolutional neural network; kNN: k-nearest neighbor; MLP: multilayer perceptron.

According to Table 3, the proposed model attained high classification performance by deploying a simple model. Also, the presented model is an XAI model and presents explainable results.

The most important points of this work are also discussed below.

• In this research, a quadruple feature extraction function (QuadTPat) has been used. The presented feature extraction function is an automaton-based feature extractor. By analyzing the connections of the used automaton, the findings were obtained. Four nodes were created using the output of the channel transformers, and these are Level 1 (L1), Level 2 (L2), Level 3 (L3), and Level 4 (L4). L1 represents the most dominant channels according to amplitude, while L4 represents the most recessive channels. According to the results of the CWNCA, 42 features of the selected 180 features have been selected from L3 and L4 connections. However, only 1 feature was selected from the L1-L1-based feature extraction model.

• CWNCA is an effective feature selector and chooses 180 out of the generated 3136 features.

• In the classification phase, some classifiers were tested to apply the t algorithm, and the classification accuracies of these classifiers are showcased in Fig. 3. Per Fig. 3, the best classifier is the ensemble k-nearest neighbors (EkNN) since this classifier reached 90.92% classification accuracy, and kNN achieved 90.51% classification accuracy on the collected EEG stress dataset. At the same time, the worst classifier is Decision Tree (DT) since it yielded 77.34% classification accuracy. Therefore, we have applied the t algorithm to kNN, and the tkNN classifier has been used in this research.

• The used tkNN classifier attained 92.94% accuracy. In this aspect, it reached a 2.02% classification accuracy higher than the EkNN classifier.

• By using DLob and selected features, a DLob string was extracted from the extracted DLob string. FL has the highest frequency. FR also shows high frequency (the second frequently used symbol). This finding clearly demonstrated that the earthquake stress mostly affected the frontal lobes.

• Per the selected features (each feature contains two channels of information), the mostly used channel is the O2 channel since we have demonstrated earthquake videos to participants for triggering stress. The second and third channels are the T7 and P7 channels. We need channel information to create a connectome network.

• The presented model contributes to the connectome theorem. The connectome graphs channels have been depicted in Fig. 7. Per the channel transition table, the most frequent transitions are given as follows. T7-P7 (Involvement in understanding spoken language and auditory information): 10 transitions, P8-O2 (Integrating spatial and visual information, which can be crucial in stressful situations requiring environmental awareness and quick visual recognition): 9 transitions, F3-F7 (This transition suggests intensive cognitive processing, likely linked to stress-related tasks requiring complex thought processes and verbal communication): 8 transitions, FC5-F3 (This transition indicate that the individual is planning and executing movements or actions in response to stress.): 8 transitions, O2-O1 (This bilateral activity indicates comprehensive visual analysis and possibly the recognition of visual patterns, which can be heightened under stress): 8 transitions, T8-P8 (This transition suggests that the brain is actively responding to auditory stimuli and translating them into spatial and navigational context, which can be a response mechanism in stressful scenarios): 8 transitions.

• The connectome graphs of the channels are shown in Fig. 8. Transitions between left hemisphere channels (AF3, F7, F3, FC5, T7, P7, O1) and right hemisphere channels (AF4, F8, F4, FC6, T8, P8, O2) indicate bilateral brain activity and coordination between hemispheres during stress detection. This bilateral activity suggests that both hemispheres actively manage stress, integrating cognitive, emotional, and sensory information.

• High activity in the frontal and temporal lobes indicates that cognitive and emotional regulation processes are important during stress. This information can be used to enhance the effectiveness of therapy methods such as CBT (Cognitive Behavioral Therapy) and neurofeedback. For example, during neurofeedback sessions, patients can be taught to control activity in specific brain regions.

• Since EEG analysis shows which brain regions are more associated with stress, pharmacological treatment strategies can be adapted according to these findings. For instance, medications aimed at regulating frontal lobe activity may be more effective in the treatment of stress and anxiety.

Figure 7.

The frequencies of the used channels.

Figure 8.

Connectome graphs of the channels.

Limitations

• To label stress, we used the expressions of the participants. We cannot do medical tests (blood or adrenalin test) to measure stress levels.

• The study uses the Emotiv Epoch X 14-channel brain cap. While this device is suitable for mobile EEG data collection, it has limitations in channel count and sampling frequency (128 Hz), which might not capture all the nuances of brain activity compared to high-density EEG systems.

• Bigger datasets can be collected.

• LOSO CV results are relatively low.

Future works

• The expansion and improvement of the DLob symbolic language are among the near-future plans, with studies planned to make it a more meaningful language through new versions.

• It is intended to simplify disease detection processes by creating a dictionary for the DLob symbolic language, enabling easier detection of diseases from complex EEG signals.

• Although a 14-channel EEG system was used in this study, future research will explore the application of the QuadTPat-based model to high-density EEG systems.

• Future studies will aim to expand the model to detect stress from a broader range of stressors beyond earthquake-related stress. This will include stress resulting from chronic stressors, workplace stress, or psychological conditions such as anxiety and depression.

• The development of real-time stress monitoring systems based on the QuadTPat model is also planned. These systems will be integrated into wearable devices, providing continuous stress assessment and triggering real-time interventions, such as biofeedback or relaxation techniques, to reduce stress.

• The study’s findings contribute to the evolving field of Causal Connectome Theory (CCT) by providing a framework for understanding how different brain regions interact during stress. Future research will build on this by developing models that map causal connections between neural pathways. The reasons behind the formation of complex brain networks will be examined, and the results will be presented.

• Improved versions of the proposed model will be utilized in personalized medicine and personalized treatments.

Conclusions

This research presents a new explainable artificial intelligence (XAI) approach for stress detection using EEG signals through the development of the QuadTPat-based model. Using the DLob symbolic language, the model not only achieves high classification performance but also provides explainable results by offering insights into brain activity during stress. The model demonstrates its reliability in detecting stress by achieving robust classification accuracy of 92.94% using 10-fold cross-validation and 73.63% with leave-one-subject-out cross-validation.

The innovations in this study include the introduction of the QuadTPat feature extraction technique, which captures complex transitions in EEG signals, and the CWNCA feature selector, which effectively identifies the most informative features for classification. The tkNN classifier, developed with these feature engineering techniques, ensures high accuracy in stress classification. The explainable results obtained through the DLob symbolic language provide a clear understanding of brain activity, particularly by highlighting the interactions in the frontal, temporal, parietal, and occipital lobes during stress. The frequent occurrence of FL and FR symbols suggests that the frontal lobe plays an important role in logical thinking, planning, and emotional regulation, while TL and TR symbols indicate the processing of language and auditory information, likely due to the earthquake-related stimuli used in the study. Additionally, the complex connectome diagram obtained shows that stress affects all brain activities.

This research not only contributes to the field of automatic stress detection but also enhances the broader understanding of the brain’s response to stress, particularly the complex interactions between different brain regions. The findings have practical implications for real-time stress monitoring systems, personalized medicine, and stress-related interventions. Future research can explore the extension of the DLob symbolic language, its integration with high-density EEG systems, and the application of the model to a wider range of stress factors. Additionally, the results provide a basis for further development of CCT, with potential applications in understanding complex brain networks and their roles in the stress response.

Author contributions

Conceptualization, VYC, IT, GT, RH, SD, TT; formal analysis, VYC, IT, GT, RH, SD, TT; investigation, VYC, IT, GT, RH, SD, TT; methodology, VYC, IT, GT, RH, SD, TT; project administration, TT; resources, VYC, IT, GT, RH, SD, TT; supervision, TT; validation, VYC, IT, GT, RH, SD, TT; visualization, VYC, IT, GT, RH, SD, TT; writing—original draft, VYC, IT, GT, RH, SD, TT; writing—review and editing, RH, SD, TT. All authors have read and agreed to the published version of the manuscript.

Funding

This work is supported by the 123E357 project fund provided by the Scientific and Technological Research Council of Turkey (TUBITAK) and by the TEKF.24.48 project fund provided by the Scientific Research Projects Coordination Unit of Firat University.

Data availability

The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.

Declarations

Ethical approval

The Non-Invasive Ethics Committee, Firat University, approved this research on ethical grounds on December 10, 2023 (2023/ 14–18). Informed consent to participate and to publish was obtained from participants.

Consent to participate

Written informed consent was obtained from all subjects (patients) in this study.

Competing interests

The authors declare no competing interests.