A machine learning approach for detection of claustrophobic brain activity in electroencephalography

Abstract

Claustrophobia, a phobia with a specific unreasonable and excessive fear of enclosed spaces, can have a considerable impact on an individual’s life. Electroencephalography (EEG) has been a tool with potential for studying neural processes in anxiety disorders including claustrophobia. In this work, a machine learning algorithm for differentiation between claustrophobic and healthy controls using EEG signals is presented. EEG data were collected from 22 participants under controlled conditions, and preprocessing included filtering, artifact removal, and feature extraction using relative Power Spectral Density (rPSD) across five frequency bands: delta, theta, alpha, beta, and gamma. Classical machine learning models such as Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), Decision Tree (DT), and Random Forest (RF) were applied to assess their suitability in this domain. In addition, deep learning models including Multi-Layer Perceptron (MLP) and a Convolutional Neural Network with Bidirectional Long Short-Term Memory (CNN-BiLSTM) integration were utilized for their capacity to capture complex temporal and spatial patterns in EEG data. Performance testing with five-fold cross-validation revealed that MLP and CNN-BiLSTM performed best in terms of accuracy in classification, with a 95.15% ± 0.77 when all bands of frequencies were combined together. An analysis of brain regions revealed frontal and temporal regions to differentiate between claustrophobic and non-claustrophobic subjects, and beta and theta bands played a significant role in distinguishing between them. These observations unveil high potential for EEG-based machine learning algorithms in objective evaluation of claustrophobia, and propose opportunities for future development in its therapy and diagnosis.

Introduction

A phobia occurs when an individual’s life is highly interfered with an extreme unrealistic fear. For instance, a person with a phobia may avoid certain situations or objects related to the fear and will not be able to function normally¹. The Diagnostic and Statistical Manual of Mental Disorders (DSM-5) identifies various anxiety disorders, including separation anxiety disorder, social anxiety disorder, panic disorder, generalized anxiety disorder, selective mutism, and specific phobias. Specific phobias can be categorized into four groups: animal phobias, natural/environmental phobias, situational phobias, and blood injection injury phobias². Specific phobias are characterized by an overwhelming fear associated with certain activities, people, objects, or situations³.

One of the particular phobias is claustrophobia, a fear of closed spaces (claustro means closed). Common triggers include tunnels, elevators, trains, airplanes, small cars, caves, MRI imaging machines, cellars, and small rooms without windows or windows that can’t be opened⁴. Specific phobias are marked by high and persistent fears of particular objects or situations, which cause the sufferer to exert a great deal of effort to avoid the feared object. The fears sometimes can be expressed in terms of danger or disgust or physical symptoms, like a pounding heart or upset stomach, that may result from exposure to the phobic trigger⁵. Claustrophobic individuals fear not only the enclosed space but also events or occurrences that might happen in the confined space. Almost all enclosed spaces suggest entrapment and a limitation to movement. A person who has this disorder, typically fears the possibility of what could happen in a confined space, such as not having enough oxygen. The suffocation fear is one of the solid and standard features of claustrophobia, felt by many sufferers when entering confined spaces; it is associated with or even forms the sensation of breathlessness⁶.

In more recent years, brain mapping and electroencephalography have been used to increase the understanding of brain function in humans. These methods help to understand the process of behavioral and cognitive functions of the brain^7–11. There are some studies on phobia, from the perspective of brain activity^12–14.

In a study on social phobia (SP), Sachs et al.⁸ examined quantitative EEG in 25 patients with SP off medications, with age- and sex- matched normal controls. They reported significant differences in EEG patterns, particularly in the frontopolar and central regions. Patients with SP were characterized by lower absolute and relative delta and theta power, as well as reduced alpha-adjacent slow beta and fast beta power. They also exhibited greater intermediate beta power in both absolute and relative terms. In a related study, Gemignani et al.⁹ investigated the autonomic as well as EEG reactivity to phobic stimuli among specific phobia patients varying in hypnotizable levels. They compared the responses of these groups with resting conditions and neutral stimuli. The results indicated hypnotizable modulates both autonomic and EEG patterns during phobic stimulation. However, the presence of a specific phobia attenuated the protective effect of hypnotizable against stress-related cardiac issues. In another study of spider phobia, Wiens et al.¹⁰ investigated the effects of in vivo and virtual reality exposure therapy on EEG measures. Spider-phobic individuals received either in vivo or virtual reality exposure treatment. EEG recordings in reaction to spider pictures and other emotional stimuli were evaluated. It was shown that both treatments reduced negative emotional evaluations of spiders; however, no effects on early posterior negativity and late positive potential were found. So, overall effects after in vivo and virtual reality exposure therapy turned out to be similar, but the mechanisms behind initial phases of motivated attention were not altered by the treatment.

In another study, the authors reported that event-related potentials (ERPs) can indicate a gender-dependent recruitment of frontal attention networks in dental phobia¹⁵. Additionally, Bornas and colleagues demonstrated changes in sample entropy and phase synchrony between pairs of EEG channels one week after the phobia condition¹⁶. This study showed that sample entropy increased, while phase synchrony decreased across all sites and pairs of channels.

Whereas these studies examined brain activity in individuals with specific phobias, such as dental phobia¹⁵, snake phobia¹⁶, and acrophobia¹⁷, there are few studies on brain activity related to claustrophobia^18–21. Moradi et al. showed that EEG activity differs between claustrophobic patients and healthy controls. They reported that the relative power spectral density (rPSD) in the theta frequency band increased in the healthy group but decreased in claustrophobic patients. In the alpha band, rPSD increased in both healthy and claustrophobic groups^18,19.

The use of artificial intelligence systems in a range of fields^7,22,23, specially to detect various cognitive disorders by learning from neural activity has increased in recent years^24–26. One field that utilizes machine learning is phobia detection using EEG signals^11,17. Researchers in¹⁷ report that the Convolutional Neural Network (CNN) and Artificial Neural Network (ANN) models can classify degrees of acrophobia using EEG signals with accuracies of 99% and 97%, respectively. Another study is the work of Apicella et al.¹¹, who designed an EEG-based classification system for fear of heights. The use of a VR scenario provided a canyon in which people were subjected to varying degrees of heights and hence varying degrees of fear-inducing stimuli. Both experienced a fear of heights, and the effectiveness of fear induction was measured using psychometric tools. Among other bands, the gamma band was found to be the most important about the fear of heights, followed by the high-beta band, and the most critical area of the scalp was the frontal one. An average accuracy of 68.20% was reached for a three-class fear classification task with EEG features. It was more specific in the area of the frontal region of the scalp, where high-beta and gamma bands had an accuracy rate of 57.90% and 61.30% accuracy, respectively.

Conventional techniques for evaluating claustrophobia, like self-reports and interviews, are arbitrary and might not fully capture the brain underpinnings of the condition. By recording the real-time brain activity linked to fear and anxiety, EEG provides a more objective, non-invasive, and economical option. EEG is more useful for dynamic and portable applications than fMRI. EEG data can help create precise, real-time claustrophobia diagnostic tools when paired with machine learning. It’s critical to assess different traditional machine learning algorithms in order to develop an efficient EEG-based classification system. Decision Trees provide interpretability, Random Forests increase accuracy through ensemble learning, Support Vector Machines excel at handling high-dimensional data, and k-Nearest Neighbors adjust to data structures without making assumptions. Before advancing to more intricate deep learning techniques, investigating these models offers useful baselines and insights^27–30.

In this study, we investigate identification of individuals with claustrophobia from healthy subjects using EEG signals. Our approach involves extracting features such as rPSD and employing both deep learning classifiers (CNN-BiLSTM and MLP) and traditional machine learning classifiers (SVM, KNN, RF, and DT). Additionally, we explore the choice of brain regions and frequency bands and its impact on the classification performance. To the best of our knowledge, no prior research has specifically addressed the classification of claustrophobia patients and healthy controls using EEG signals.

Materials and methods

The proposed method for classifying the claustrophobic and non-claustrophobic participants is shown in Fig. 1. It includes five steps of signal acquisition, pre-processing, feature extraction, and classification. Recording of the EEG signals, from the claustrophobic and non-claustrophobic participants, was performed following the International 10–20 electrode placement system. Before analysis, raw signals were pre-processed with down-sampling, filtering, and Independent Component Analysis (ICA) to boost the signal quality and remove various artifacts. Then decomposition of signals into Delta (0.5–4 Hz), Theta (4–8 Hz), Alpha (8–13 Hz), Beta (13–30 Hz), and Gamma (30–48 Hz) frequency bands was realized alongside rPSD calculations within each band. Machine learning and deep learning approaches were adopted for classification, including SVM, KNN, DT, RF, MLP), and a combined CNN with BiLSTM. Performance metrics, including accuracy, precision, recall, and F1-score, were calculated to evaluate the models, and the results were analyzed with respect to EEG bands and brain areas to determine distinguishing features of claustrophobic versus non-claustrophobic subjects.

Fig. 1.

Block diagram of the designed method.

Participants and behavioral experiments

The study involved recording EEG signals from 22 participants: 9 with self-identified claustrophobia and 13 healthy controls. For details on the experimental setup and recording protocol, refer to our previous study¹⁹. In the following, we provide a brief description. The claustrophobic group’s age ranged from 23 to 48 years (mean ± standard deviation age: 29.11 ± 8.73), with 44% female participants. The healthy control group’s age ranged from 21 to 47 years (27.61 ± 6.39), with 39% female participants. All participants were non-smokers, right-handed, and without a history of neurological illnesses. They received an oral explanation of the experimental procedure and provided informed consent. Participants were informed that they could stop the procedure immediately if they felt uncomfortable. Additionally, participants refrained from taking any medication before the experiment session.

The experiment was conducted with three consecutive conditions, which are marked further as R0, T1, and T2 below. In the first condition, the R0 condition, participants were located in a large and well-illuminated laboratory room. After setting up the data collection headcap, participants sat comfortably while the EEG was recorded for 300 s. In the T1 condition, participants were transferred into a moderately lit, closed-door wooden chamber (measuring180 cm Length × 50 cm Width × 90 cm Height). They took a sitting position on a similar chair for another 300s of EEG data collection. For the third condition (T2), the length of the wooden chamber was reduced to 90 cm, which created a reduced-dimension chamber without doors. This study focused on the T2 condition.

EEG signal and pre-processing

EEG brain activity was recorded using a 31-channel setup at a sampling frequency of 500 Hz. The data was initially of size (22, 31, 150000), implying that there were 22 subjects, 31 EEG channels, and 150,000 time points. Subsequently, the data was down-sampled to the target frequency of 100 Hz, so the dimensions became (22, 31, 30000). Pre-processing included high pass filtering with a cut-off frequency of 0.1 Hz by a 3rd order Butterworth filter, application of a CAR filter, and a short Laplacian filter, introduced to increase spatial resolution and decrease volume conduction effects³¹. ICA for artifact detection and removal preceded signal reconstruction, involving 31 channels. Since the length of data differed after being processed, all data were trimmed to the size of 23,250 samples each. Hence, this made the final size of the input data (22, 31, 23250). All pre-processing was performed using MATLAB 2023b and the EEGLAB toolbox³². We deployed electrodes in distinct regions of the scalp corresponding to different lobes of the brain. For the frontal lobe, we used 13 electrodes: Fp1, Fp2, F3, F4, F7, F8, FC3, FC4, FT7, FT8, FPz, Fz, and FCz. Temporal lobe activity was monitored using 4 electrodes: T7, T8, TP7, and TP8. The parietal lobe was covered by 5 electrodes: P3, P4, P7, P8, and Pz. In the central lobe, 6 electrodes were placed: C3, C4, CP3, CP4, Cz, and CPz. Finally, the occipital lobe activity was recorded using 3 electrodes: O1, O2, and Oz. Figure 2 illustrates these regions, with blue representing the frontal lobe, light pink for the central lobe, yellow for the parietal lobe, green for the temporal lobe, and dark pink for the occipital lobe.

Fig. 2.

Electrode Locations and brain regions.

Feature extraction and machine learning analysis

After preprocessing, we designed a feature extraction pipeline to derive spectral information from EEG signals. As illustrated in Fig. 3, we used a sliding window approach with a window size of 10 s⁷ and an overlap of 90% for effective data augmentation^33,34. Each of these windows was processed to extract five EEG frequency bands: δ (0.5–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz), and γ (30–48 Hz) with 3rd order Butterworth filter (Fig. 3). Subsequently, we calculated the rPSD for each window, channel, and frequency band. The rPSD is defined as¹⁸:

1

Fig. 3.

Extraction of EEG signal in various frequency bands Using a Bandpass Butterworth Filter from 10 s window of signal.

Where Band power is the power of the specific frequency band and Total power is the power of the original, full-spectrum signal. Both band-specific and total power spectral densities were derived using Welch’s method³⁵. This resulted in 223 rPSD values for each frequency band, channel, and subject, providing a detailed spectral representation of the EEG data across spatial and temporal scales.

Machine learning and deep learning models

In the present study, we tested numerous machine learning and deep learning models. For machine learning, we used k-NN, SVM, DT, and RF models. For deep learning, we developed two types of architectures: one is a MLP architecture, and the other is a CNN-BiLSTM. A complete description of the architectures and the parameters used for these models is provided as following sessions.

k-Nearest neighbors (k-NN)

k-Nearest Neighbors is a non-parametric classification algorithm wherein the output of this model is a class membership that can be determined by majority voting of its k-nearest neighbors. A similarity measure based on distance metrics, like Euclidean or Manhattan distance, is generally employed²⁷.

Support vector machine (SVM)

Support Vector Machine classifies the data by finding the hyperplane which best separates classes in a high-dimensional space. In such a way, it can also support linear and nonlinear data using some kernel functions, like the Radial Basis Function (RBF) or polynomial kernels. Here, in this work, we have implemented the RBF kernel with SVM²⁸.

Decision tree (DT)

Decision Tree is a hierarchical model that predicts the value of a target variable by learning simple decision rules that will be inferred from the features of the data. The algorithm will repeatedly divide the data into subsets depending on the more critical attributes and continue until some stop criterion is met²⁹.

Random forest (RF)

Random Forest is a learning method for classification, regression, and other tasks that operate on an ensemble learning method. It does this by training multiple decision trees that, at inference time, perform their prediction of classes, after which the output class is determined by taking the mode of the classes in classification and mean prediction in regression. It reduces overfitting due to its characteristic of tree averaging³⁰.

Multi-layer perceptron (MLP)

The MLP follows a 6-layer Densely Connected architecture with an input layer of 256 neurons, 4 hidden layers of 128, 64, 32, and 8 neurons, and an output layer with one neuron. Input and hidden layers have LeakyReLU activation with α = 0.2 to infuse non-linearity and batch normalization for stabilization and the speedy training of models. To prevent overfitting, a dropout of 0.2 is introduced. This implementation sets a fraction of the input neurons to zero randomly during training. The final layer is a dense layer with a single neuron and a sigmoid activation function for binary classification. The model is compiled using the Adam optimizer to maximize the logarithmic loss with a learning rate set at 0.001. Binary cross-entropy is the loss function that measures classification efficacy to allow suitable model optimization and assessment.

Hybrid CNN-BiLSTM deep classifier

The hybrid CNN-BiLSTM deep classifier was composed of a 1D Convolutional Neural Network (CNN) connected to Bidirectional Long Short-Term Memory (BiLSTM) and Dense layers. As shown in Fig. 4, this architecture has two layers of Conv1D, with 64 filters each, whose kernel size is 3. These are followed by ReLU activation and then MaxPooling1D layers with a pool size of 2 to downsample feature maps. Dropout rate of 0.2 is applied after the pooling and dense layers, respectively, for regularization. This model is set with two bidirectional LSTM layers of 128 units, while the first LSTM layer returns an entire sequence of outputs for further processing. Two dense layers consider ReLU activation for non-linearity, where the first thick layer has 128 neurons and the other one 64 neurons. The last layer is a thick layer with one neuron and a sigmoid function, which classifies binaries. We compiled the model using the Adam optimizer and applied binary cross-entropy as the loss function.

Fig. 4.

Hybrid CNN-BiLSTM Deep Classifier Architecture.

Model training and performance evaluation

To evaluate both performance and generalization, the models were trained using five-fold cross-validation. In five-fold cross-validation, the data set is divided into five equal parts; then, the model is trained once, with one of the parts serving as a validation set, and four parts are used for training the model. The process iterated five times ensures the performance estimate at each fold, so it is robust and reliable, reducing bias and variance. The efficiency of these models was measured with the help of different validation metrics such as accuracy, precision, sensitivity, and F1-Score. Accuracy is the proportion of correct predictions over the total instances, and precision indicates the proportion of true positives out of all optimistic predictions. Sensitivity quantifies the proportion the accurate optimistic prediction makes in all real positive cases, and F1-Score is the harmonic mean between precision and recall, providing a balance between both measures.

2

3

4

5

6

Where TP, TN, FP, and FN are true positives, true negatives, false positives, and false negatives²⁴. These metrics give proper information on how a model would perform with high rates of accuracy, precision, sensitivity, and overall effectiveness, respectively, of the F1 score.

Results

To visualize the power spectral density spatial distribution over different frequency bands, we plotted the brain topography maps for both groups: claustrophobic and non-claustrophobic. These topographies were based on the rPSD values derived during the feature extraction. We calculated the average rPSD of each group’s subjects across all subjects, thus obtaining for each group a 31 × 5 matrix representing the 31 EEG channels and five frequency EEG bands. We derived full topographic maps for both groups that depicted the distinct patterns of neural activity associated with claustrophobic and non-claustrophobic individuals in the restricted experimental setting.

For a more illustrative presentation of differences between groups, we computed a differential topography map: we subtracted the average rPSD values of the non-claustrophobic group from the rPSD values of the claustrophobic group. Group differences are, therefore, visually highlighted for regions and frequency bands in which the differences in neural activity between both groups occur. The statistical significance was tested using the Mann-Whitney U-test, with Bonferroni correction applied to account for multiple comparisons across 31 channels and 5 frequency bands. This adjustment ensured a rigorous control of Type I error while assessing the significance of neural activity differences.

These topographic representations will provide us with necessary information regarding the neural correlates of claustrophobia: pointing out specific brain regions and frequency bands that could play an essential role in the experience of claustrophobic distress.

In the claustrophobic group, Fig. 5 shows distinct neural activity patterns across frequency bands. Delta band activity is higher in the prefrontal region, while theta band activity is concentrated near the central brain, extending to the frontal area. Alpha activity is most prominent in the occipital regions. Beta and gamma bands exhibit similar activity patterns in the parietal regions on both sides of the head, with comparatively lower activation elsewhere. These findings suggest heightened delta activity in the frontal region, focused theta activity in the central area, pronounced alpha activity in posterior regions, and localized beta and gamma activation in temporal-parietal areas. Such patterns may reflect altered neural processing, potentially linked to heightened emotional and cognitive responses in confined spaces among claustrophobic individuals.

Fig. 5.

“EEG Topographical Analysis of rPSD in Claustrophobic and healthy subjects across frequency bands”.

In the control group, Fig. 5 shows that the delta band exhibits less activity in the frontal region compared to the claustrophobic group. The theta band demonstrates higher activity in the center of the head, similar to the claustrophobic group, but with a tendency towards the occipital rather than frontal areas. Alpha activity is stronger in the occipital regions compared to the claustrophobic group, possibly indicating a more relaxed state. The beta and gamma bands display considerably less activity in the parietal regions compared to the claustrophobic group, with beta showing slightly more frontal activation. Overall, the control group exhibited more evenly distributed delta activity and less intense theta activation, with alpha activity notably stronger in the posterior regions. Beta and gamma bands showed subtle distributional differences, with less parietal activation and a slight increase in frontal activation for the beta band. After conducting between-group comparisons between controls and claustrophobic individuals across different brain regions and frequency bands, significant differences emerged following Bonferroni correction (p < .0001). These differences highlighted distinct neural activity patterns across brain regions and frequency bands.

The result shows that in the frontal region, the prefrontal cortex displayed pronounced differences. Fp1 exhibited significant differences across all frequency bands (delta, theta, alpha, beta, and gamma), while Fp2 showed significant differences in the lower frequency bands (delta, theta, and alpha) but not in the higher ones (beta and gamma). The dorsolateral prefrontal cortex (F3, F4) followed a similar trend, with significance across all frequency bands except delta, indicating altered executive function processing in claustrophobic individuals. The central region analysis revealed asymmetrical patterns. C4 showed significant differences across all frequency bands, whereas C3 exhibited selective significance, particularly in the delta and alpha bands. This asymmetry extended to the frontocentral areas, where both FC3 and FC4 demonstrated strong significance across multiple frequency bands, suggesting altered sensorimotor integration in claustrophobic individuals.

The temporal region analysis revealed distinct lateralization effects. T8 (right temporal) exhibited significant differences across all frequency bands, while T7 (left temporal) showed more variable significance patterns. This lateralization points to potential differences in emotional processing and spatial awareness between groups. Similarly, the temporoparietal junction (TP7, TP8) displayed strong significance patterns, suggesting altered multimodal sensory integration.

In the parietal region, electrodes P3, P4, P7, and P8 consistently showed significance across multiple frequency bands, implying differences in spatial processing and sensory integration between groups. The midline parietal electrode (Pz) showed significance in all bands except gamma, whereas the centroparietal (CPz) electrode demonstrated significance in alpha, beta, and gamma bands. The occipital region presented mixed significance patterns. O1 and O2 exhibited variable responses across frequency bands, while Oz showed significance in all bands except gamma. These findings suggest altered visual processing mechanisms in claustrophobic individuals compared to controls.

The comprehensive pattern of significant differences across multiple brain regions and frequency bands indicates widespread alterations in neural processing in claustrophobic individuals, with particularly robust differences in frontal, central, and parietal regions. These findings suggest that claustrophobia is associated with broad changes in brain activity patterns, potentially reflecting altered emotional processing, spatial awareness, and cognitive control mechanisms.

Figure 6 shown the classification performance of six machine learning models—SVM, KNN, DT, RF, MLP, and CNN-BiLSTM—across different EEG frequency bands: All-band, Delta, Theta, Alpha, Beta, and Gamma. The models were assessed based on accuracy, precision, sensitivity, and F1-score, as shown in Table S1.

Fig. 6.

The performance of classifiers (SVM, KNN, RF, DT, MLP, and CNN-BiLSTM) across difference frequency band.

In the All-band scenario, MLP and CNN-BiLSTM stood out with the highest classification accuracy, achieving 95.13% ± 1.15 and 95.15% ± 0.77, respectively. SVM also performed well at 94.68% ± 0.55, followed by KNN with 93.97% ± 0.82. Decision Tree had the lowest accuracy (92.17% ± 1.14). MLP recorded the highest precision (95.17% ± 1.80), while CNN-BiLSTM excelled in sensitivity (94.72% ± 2.40). These trends were consistent for F1-score, with MLP and CNN-BiLSTM achieving 93.95% ± 1.52 and 94.11% ± 0.95, respectively. A one-way ANOVA revealed a significant effect of classifier on accuracy in the All-band scenario, F(5, 24) = 23.43, p < .001, η² = 0.83. Post-hoc pairwise comparisons with Bonferroni correction (α = 0.0033) showed that MLP and CNN-BiLSTM significantly outperformed both Decision Tree and Random Forest. Similarly, SVM and KNN performed significantly better than these conventional models, further confirming the inferior performance of Decision Tree and Random Forest in the all-band analysis.

For the Delta band, SVM and KNN led in accuracy at 93.23% ± 0.63 and 93.80% ± 0.74. MLP closely followed (93.62% ± 0.76), while CNN-BiLSTM scored 92.11% ± 0.62. Decision Tree had the weakest accuracy (82.69% ± 0.95). KNN achieved the highest precision (95.04% ± 0.65), and SVM demonstrated strong precision at 92.73% ± 1.24. CNN-BiLSTM and MLP excelled in sensitivity at 89.29% ± 2.88 and 91.13% ± 2.59, respectively. MLP also led in F1-score (92.10% ± 1.06), slightly ahead of SVM (91.63% ± 0.80). A one-way ANOVA revealed a highly significant effect of classifier type on accuracy in the Delta band, F(5, 24) = 184.79, p < .001, η² = 0.97, indicating a large effect size. Post-hoc pairwise comparisons with Bonferroni correction (α = 0.0033) identified nine significant differences. Specifically, SVM, KNN, CNN-BiLSTM, and MLP significantly outperformed both Decision Tree and Random Forest. Furthermore, Random Forest outperformed Decision Tree, and MLP also outperformed CNN-BiLSTM, highlighting its superior generalization in this frequency band.

In the Theta band, MLP achieved the highest accuracy (94.46% ± 0.84), followed by KNN (93.42% ± 0.90) and SVM (92.66% ± 1.06). Decision Tree recorded the lowest accuracy (85.69% ± 1.32). MLP demonstrated the best precision (97.30% ± 2.55). Sensitivity scores were generally lower, with CNN-BiLSTM and SVM scoring 88.99% ± 2.38 and 88.59% ± 2.46, respectively. MLP led in F1-score at 92.92% ± 1.14, outperforming CNN-BiLSTM (90.76% ± 1.07) and SVM (90.79% ± 1.41). Statistical analysis using one-way ANOVA confirmed a significant effect of classifier on accuracy in the Theta band, F(5, 24) = 49.95, p < .001, η² = 0.91, indicating a very large effect size. Post-hoc pairwise t-tests with Bonferroni correction (α = 0.0033) revealed eight significant differences. Notably, MLP significantly outperformed all other classifiers, including SVM, CNN-BiLSTM, RF, and DT. Additionally, SVM and KNN significantly outperformed Decision Tree, while Random Forest, CNN-BiLSTM, and MLP also achieved significantly better performance than DT. These findings highlight MLP’s dominance in the Theta band, both statistically and performance-wise.

In the Alpha band, MLP again demonstrated the highest accuracy (94.35% ± 1.38), followed by KNN (93.48% ± 0.86) and SVM (91.89% ± 0.90). Decision Tree showed relatively better performance compared to other bands (88.77% ± 0.85). MLP also excelled in precision (97.41% ± 3.35). However, sensitivity was lower, with CNN-BiLSTM and SVM at 85.20% ± 4.36 and 81.97% ± 2.47, respectively. MLP maintained the highest F1-score (92.77% ± 1.80), followed by CNN-BiLSTM (90.49% ± 1.70). One-way ANOVA revealed a significant effect of classifier on accuracy in the Alpha band, F(5, 24) = 23.41, p < .001, η² = 0.83, indicating a large effect size. Post-hoc comparisons using Bonferroni correction (α = 0.0033) identified six statistically significant differences. Specifically, MLP significantly outperformed Decision Tree, SVM, and Random Forest. Additionally, SVM and KNN significantly outperformed Decision Tree, and CNN-BiLSTM also achieved significantly better performance than DT. These results confirm MLP’s statistical superiority in the Alpha band, particularly in terms of accuracy and F1-score.

In the Beta band, MLP achieved the highest accuracy (95.21% ± 0.96), with SVM closely following (94.90% ± 0.51). CNN-BiLSTM also performed well (94.64% ± 0.50). Decision Tree showed lower accuracy (90.64% ± 0.68). KNN had the highest precision (95.37% ± 1.13), while SVM and CNN-BiLSTM performed well in sensitivity at 94.92% ± 0.76 and 95.02% ± 3.86, respectively. F1-score results confirmed this trend, with CNN-BiLSTM achieving 93.54% ± 0.62, followed closely by SVM at 93.84% ± 0.61. A one-way ANOVA revealed a significant effect of classifier on accuracy in the Beta band, F(5, 24) = 20.16, p < .001, η² = 0.81, suggesting a strong influence of classifier choice. Post-hoc pairwise comparisons with Bonferroni correction (α = 0.0033) identified four statistically significant differences: MLP, CNN-BiLSTM, KNN, and SVM each significantly outperformed Decision Tree. While the differences among the top-performing models (MLP, CNN-BiLSTM, and SVM) were not statistically significant, their high accuracy and F1-scores suggest comparable and robust classification capabilities in the Beta band.

In the Gamma band, MLP obtained the highest accuracy (94.94% ± 0.78), followed by SVM (94.64% ± 0.62) and CNN-BiLSTM (94.41% ± 0.75). Decision Tree performed better than in other bands (92.05% ± 0.47). MLP achieved the highest precision (97.41% ± 2.21), with SVM and KNN maintaining consistent precision at 92.94% ± 1.53 and 95.15% ± 1.17, respectively. Sensitivity was highest for CNN-BiLSTM (93.28% ± 4.19) and SVM (94.07% ± 0.62). MLP secured the highest F1-score (93.58% ± 1.05), with SVM following at 93.49% ± 0.71. A one-way ANOVA revealed a significant effect of classifier on accuracy in the Gamma band, F(5, 24) = 13.60, p < .001, η² = 0.74, indicating a moderate to strong effect of classifier type. Post-hoc pairwise t-tests with Bonferroni correction (α = 0.0033) found three significant differences: MLP, SVM, and CNN-BiLSTM all significantly outperformed Decision Tree. However, no significant differences were observed among the top-performing models themselves. These results suggest that although multiple classifiers perform well in the Gamma band, their statistical advantage over simpler models like Decision Tree is most notable.

Overall, MLP and CNN-BiLSTM emerged as top performers, particularly in the All-band, Beta, and Gamma scenarios. MLP excelled in accuracy and precision, while CNN-BiLSTM stood out for sensitivity and F1-score. Traditional models like SVM and KNN showed competitive results in specific bands, especially Delta and Beta. Decision Tree demonstrated the weakest performance, particularly in Delta and Theta, struggling with accuracy and sensitivity. These findings highlight the effectiveness of deep learning models, particularly MLP and CNN-BiLSTM, for EEG signal classification, offering valuable insights for neurocognitive engineering applications.

In our previous analysis, we examined EEG signals through various frequency bands to understand how oscillatory patterns contributed to classification performance. While this approach provided insights into neural activity based on frequency characteristics, it did not capture the spatial distribution of signals across the brain. To address this, we shifted our focus to analyzing EEG activity across different brain regions to explore how localized neural patterns impact classification performance. This region-specific analysis helps determine whether certain brain areas contribute more significantly to classification accuracy, complementing the frequency-based findings. Figure 7 demonstrated that the performance of classifiers (SVM, KNN, RF, DT, Dense Neural network, and CNN) across brain regions. The numerical results are also shown in the table S2.

Fig. 7.

The performance of classifiers (SVM, KNN, RF, DT, Dense Neural network, and CNN) across brain regions.

The MLP model achieved the highest accuracy (95.17% ± 0.61) in Frontal, closely followed by SVM (95.15% ± 0.93) and CNN-BiLSTM (94.84% ± 0.86). These results indicate that deep learning models, particularly MLP and CNN-BiLSTM, are effective in capturing patterns from frontal EEG signals. KNN (94.15% ± 0.84) and RF (93.46% ± 0.91) also performed well. However, the Decision Tree model lagged behind, with an accuracy of 91.50% ± 1.23, highlighting its limited generalization ability.

SVM had the highest precision (95.90%), MLP followed closely (94.82%) and achieved the highest sensitivity (93.58% ± 4.42). The Decision Tree model exhibited the lowest sensitivity (89.64% ± 1.95). A one-way ANOVA revealed a significant effect of classifier on frontal lobe accuracy, F(5, 24) = 29.90, p < .001, η² = 0.86, indicating a strong effect of model type. Post-hoc pairwise t-tests with Bonferroni correction (α = 0.0033) revealed seven significant differences: SVM and KNN significantly outperformed both DT and RF; MLP and CNN-BiLSTM also significantly outperformed DT. Furthermore, MLP significantly outperformed RF. These findings highlight that MLP and CNN-BiLSTM models offer superior generalization and learning capabilities in frontal EEG processing, outperforming traditional classifiers especially Decision Tree and Random Forest.

In Temporal, MLP maintained the highest accuracy (94.33% ± 0.71), followed by CNN-BiLSTM (94.13% ± 0.98) and SVM (93.72% ± 0.70). KNN (93.44% ± 0.82) surpassed RF (92.09% ± 1.05), while Decision Tree again showed the weakest performance (90.38% ± 0.98). MLP and KNN led in precision (94.86% ± 2.56 and 94.57% ± 0.97). Sensitivity scores were lower across all models, with SVM and MLP scoring the highest at 91.23% ± 1.78 and 91.23% ± 3.25, respectively. The Decision Tree model had the lowest sensitivity (88.14% ± 1.14). A one-way ANOVA showed a significant effect of classifier type on temporal lobe accuracy, F(5, 24) = 19.57, p < .001, η² = 0.80, indicating a large effect size. Post-hoc pairwise t-tests with Bonferroni correction (α = 0.0033) identified four significant differences: SVM, KNN, CNN-BiLSTM, and MLP each significantly outperformed Decision Tree. However, no statistically significant differences were observed between the top-performing models (MLP, CNN-BiLSTM, SVM, KNN), suggesting comparable effectiveness across them in temporal region classification.

MLP achieved the highest accuracy (94.17% ± 1.06) in Parietal region, followed by CNN-BiLSTM (93.62% ± 0.52) and KNN (93.48% ± 0.95). SVM (93.15% ± 0.76) and RF (91.93% ± 1.09) also performed well. However, Decision Tree demonstrated the lowest accuracy (87.91% ± 1.16), highlighting a significant gap between tree-based methods and deep learning approaches.

MLP and CNN-BiLSTM continued to perform well in precision (94.44% ± 2.81 and 94.26% ± 1.86). MLP achieved the highest sensitivity (91.28% ± 3.46), while Decision Tree had the lowest (85.70% ± 2.35). MLP also led in F1-score (92.74% ± 1.36). A one-way ANOVA revealed a significant effect of classifier type on parietal lobe classification accuracy, F(5, 24) = 19.38, p < .001, η² = 0.80, indicating a large effect size. Bonferroni-corrected pairwise t-tests (α = 0.0033) showed six significant differences. SVM, KNN, CNN-BiLSTM, and MLP all significantly outperformed Decision Tree. Additionally, both CNN-BiLSTM and MLP also significantly outperformed Random Forest, further emphasizing the advantage of deep learning models in this brain region.

The performance trends in Central region mirrored previous findings, with MLP achieving the highest accuracy (94.66% ± 0.81), followed by SVM (94.21% ± 0.87) and KNN (93.60% ± 0.90). CNN-BiLSTM also performed well (93.31% ± 0.98), but Decision Tree remained the weakest (88.71% ± 1.23). MLP and SVM exhibited the highest precision scores (96.30% ± 1.43 and 95.92% ± 1.26). Sensitivity followed similar patterns, with MLP (90.43% ± 1.34) and SVM (89.69% ± 2.28) leading. Decision Tree continued to underperform, with a sensitivity score of 86.00% ± 2.50. One-way ANOVA revealed a significant effect of classifier on accuracy in the central lobe, F(5, 24) = 42.40, p < .001, η² = 0.90, indicating a very strong effect of classifier type. Post-hoc pairwise t-tests with Bonferroni correction (α = 0.0033) found seven significant differences. SVM and KNN significantly outperformed Decision Tree and Random Forest, while CNN-BiLSTM and MLP also showed significant superiority over Decision Tree. Additionally, MLP significantly outperformed Random Forest. These results underscore the consistent advantage of deep learning models and SVM/KNN over traditional tree-based classifiers in central EEG signal classification.

Classification accuracy was generally lower in Occipital region. MLP still led with 92.40% ± 1.06, followed by SVM (91.05% ± 1.21) and KNN (91.40% ± 0.79). CNN-BiLSTM remained competitive (91.46% ± 0.47). Decision Tree had the lowest accuracy (86.93% ± 1.74).

MLP and CNN-BiLSTM achieved the highest precision (91.65% ± 1.24 and 91.99% ± 2.72). Sensitivity dropped across all models, with Decision Tree performing the worst (84.55% ± 2.05), while MLP remained the best (89.59% ± 1.98). F1-score results followed a similar pattern, with MLP (90.59 ± 1.36) achieving the highest value and Decision Tree (84.12% ± 2.09) the lowest.

Subject-independent classification

We used leave-one-out cross-validation (LOOCV) for subject-independent classification in order to assess the generalizability of our classifiers across subjects. One subject’s data was used as the test set in this method, and the remaining twenty-one subjects’ data were used to train the model. To make sure the training and testing sets were totally separate, this procedure was repeated for each of the 22 participants (nine claustrophobic and 13 healthy). In accordance with the previously discussed pooled data approach, the EEG data were divided into 10-second windows with 90% overlap. To account for the class imbalance (13 healthy vs. 9 claustrophobic subjects), weighted-average metrics (Accuracy, Precision, Recall, and F1-Score) were used to assess the performance of the classifiers: CNN, SVM, RF, DT, MLP, and CNN-BiLSTM. The findings are summed up in Table 1.

Table 1.

Results of subject-independent classification.

	Accuracy (%)	Precision (%)	Recall (%)	F1-score
KNN	72.73	74.45	72.73	0.70
SVM	72.73	74.45	72.73	0.70
Random forest	72.73	74.45	72.73	0.70
Decision tree	72.73	74.45	72.73	0.70
MLP	72.73	74.45	72.73	0.70
CNN-BiLSTM	54.55	52.98	54.55	0.53

KNN, SVM, RF, DT, and MLP displayed consistent performance with an accuracy of 72.73%, a weighted precision of 0.7445, a weighted recall of 0.7273, and a weighted F1-score of 0.7065, indicating reasonable generalizability for traditional machine learning models. Conversely, the CNN-BiLSTM model underperformed, achieving an accuracy of 54.55%, with a weighted precision of 0.5298, a weighted recall of 0.5455, and a weighted F1-score of 0.5334, potentially due to its complex architecture requiring a larger dataset for effective feature learning. All models performed better on the healthy class (Class 0) compared to the claustrophobic class (Class 1), likely due to a greater number of healthy subjects. The lower performance on claustrophobia indicates a need for additional data and advanced feature engineering. One-way ANOVA in the Occipital lobe revealed significant differences with deep learning models significantly outperforming classical classifiers in decoding occipital EEG signals related to visual processing.

MLP consistently outperformed other models across all brain regions, demonstrating its strong classification capabilities for EEG signal processing. SVM and CNN-BiLSTM also showed strong performance, particularly in the Frontal and Central regions, indicating that both conventional machine learning and deep learning approaches can be effective. KNN and RF provided reliable results, though slightly behind MLP, SVM, and CNN-BiLSTM. The Decision Tree model consistently exhibited the weakest performance, especially in the Parietal and Occipital regions, indicating its limited ability to generalize EEG classification. These findings suggest that deep learning models, particularly MLP and CNN-BiLSTM, are better suited for EEG classification due to their ability to extract meaningful features from complex brain signals.

The performance of various classifiers across different brain areas demonstrates the robustness and effectiveness of SVM and MLP in EEG signal classification for distinguishing between claustrophobic and healthy individuals, while exposed to phobic triggers. These classifiers consistently achieved high accuracy across all brain regions, with particularly strong performance in the frontal, temporal, and central areas. The results suggest that SVM and MLP are well-suited for EEG analysis in neurocognitive engineering applications, providing valuable insights into the neural correlates of claustrophobia.

Computational performance

A critical aspect of evaluating the practical applicability of different machine learning models is their computational cost. This section details the investigation of the running times for the various methods employed in this study. The mean training and testing times for the SVM, KNN, DT, RF, CNN-BiLSTM, and MLP were recorded and are presented below. The computational performance of these models was assessed in two analytical contexts: across different frequency bands and across various brain regions.

Frequency band analysis

The following table summarizes the mean training and testing times for each model when utilizing features from different frequency bands.

Table 2 shows a significant disparity in computational load between traditional machine learning and deep learning models. Deep learning models like CNN-BiLSTM and MLP took 54 to 59 s to train, whereas traditional models were noticeably quicker, with KNN at approximately 0.0013 s. Random Forest had the longest training time among traditional models at around 2.3 s. For testing, Decision Tree was fastest at 0.0005 s, while deep learning models still showed acceptable testing times of just over one second.

Table 2.

Mean training and testing times (in ms) for each model using different frequency bands.

	SVM	KNN	DT	RF	MLP	CNN
Mean training time (ms)	257.8 ± 20.3	1.3 ± 0.4	348.5 ± 161.0	2305.3 ± 275.2	54297.3 ± 2267.4	58763.4 ± 3605.8
Mean testing time (ms)	95.3 ± 20.8	74.7 ± 60.1	0.5 ± 0.1	0.0148 ± 1.8	1163.1 ± 185.4	1112.8 ± 211.6

Brain area analysis

The computational running times were also evaluated based on the analysis of different brain regions. The results are presented in Table 3.

Table 3.

Mean training and testing times (in ms) for each algorithm across different brain regions.

	SVM	KNN	DT	RF	MLP	CNN
Mean Training Time (ms)	367.4 ± 154.7	3.5 ± 0.4	182.2 ± 16.2	1996.0 ± 245.4	64724.5 ± 3238.9	61050.2 ± 2792.6
Mean Testing Time (ms)	161.7 ± 102.8	109.6 ± 9.1	0.5 ± 0.1	20.3 ± 17.5	1128.8 ± 183.3	1195.5 ± 301.0

The analysis shows that trends in brain area and frequency band analysis align. Deep learning models, CNN-BiLSTM and MLP, have the longest training times, over 60 s, while traditional models, particularly KNN, are significantly faster. This indicates a trade-off between model complexity and efficiency; deep models provide high accuracy at the expense of training time. In contrast, KNN and DT deliver quicker training, important for resource-limited applications. The Random Forest model offers moderate training time among traditional methods. These results are pivotal for choosing models based on predictive performance versus computational cost.

Discussion

This study aimed to classify individuals with claustrophobia and healthy participants using EEG signals. The classifiers employed included SVM, KNN, DT, RF, MLP, and CNN. The findings provided several notable insights. Firstly, the frontal lobe demonstrated the highest classification accuracy, suggesting its significant role in claustrophobia. This brain region is associated with functions such as reasoning, planning, speech and movement, emotions, and problem-solving. The results align with previous research linking the frontal lobe to fear responses and anxiety disorders^1–3. Among the classifiers, SVM and MLP demonstrated the highest accuracy. SVM’s effectiveness can be attributed to its ability to handle high-dimensional data and its strength in binary classification tasks. On the other hand, MLP’s superior performance likely stems from its capacity to model non-linear decision boundaries and adapt through real-time learning, making it particularly well-suited for EEG signal classification.

When analyzing frequency bands, the beta band yielded the highest accuracy for SVM, KNN, and MLP classifiers. RF excelled in the gamma band, while DT and CNN performed best using the entire frequency band set. These results suggest that different frequency bands convey distinct information that certain classifiers can better capture. The beta band, associated with active, focused, or anxious thinking and concentration^36,37, may explain its effectiveness for SVM, KNN, and MLP models. The gamma band is associated with functions such as maintaining consciousness, perception, problem-solving, and information processing³⁸. This might explain why it was most effective for the RF classifier. the results suggest that different classifiers exhibit variable performance across frequency bands, likely due to their ability to capture distinct neural information.

EEG analysis revealed significant differences between claustrophobic and non-claustrophobic individuals, particularly in the frontal, temporal, and parietal regions. The beta and theta frequency bands were crucial in distinguishing the two groups. These findings suggest that claustrophobia may involve heightened emotional and cognitive responses in confined spaces. Consistent with previous studies, increased beta activity was associated with heightened anxiety and cognitive processing, while theta activity was linked to emotional regulation³⁹. The observed differences in frontal and temporal regions further suggest that claustrophobic individuals exhibit altered neural activity in areas responsible for fear processing and spatial awareness^18,20.

The high accuracy of MLP and CNN-BiLSTM models highlights machine learning’s potential in medical settings, particularly in diagnosing claustrophobia through EEG. This method is objective and noninvasive, complementing traditional psychological assessments. Identifying key brain regions and frequency bands involved in claustrophobia can enhance neurofeedback and cognitive-behavioral therapies. Despite promising results, clinical applicability must address study limitations, requiring critical steps for effective translation into clinical tools. Future work could explore real-time EEG-based biofeedback to help subjects regulate claustrophobic responses. While the current analysis focused on resting-state EEG, further studies should examine event-related potentials (ERPs) and task-related EEG to capture neural reactions to claustrophobic stimuli. Additionally, fMRI could provide further insights into the neural processes underlying claustrophobia, with the current findings offering a foundation for improved diagnostic tools for claustrophobia and related anxiety disorders.

Limitations

Despite the promising results, it is important to acknowledge a number of restrictions in the current study. The generalizability of the findings could be limited by the relatively small sample size of 22 participants, which included only nine individuals with self-identified claustrophobia and 13 healthy controls. Future work should therefore aim to include larger and more diverse samples to confirm the solidity of the proposed classification models. Another potential disadvantage is the reliance on self-rated claustrophobia; to make diagnoses less sensitive to subjectivity, future studies could incorporate standardized clinical tests, such as heart rate variability. Furthermore, the present study considered resting-state EEG recordings, whereas future investigations could explore event-related potentials (ERPs) and task-related EEG to dynamically record neural activity in reaction to specific claustrophobic stimuli. The performance of the classifiers could also potentially be improved by examining different preprocessing and feature extraction algorithms.

Conclusions

This study demonstrates that deep and machine learning techniques with an EEG can classify claustrophobic subjects with high accuracy. The study reveals significant information regarding neural processes in claustrophobia and confirms the use of EEG in diagnosing anxiety disorder. There is a direction for future studies in terms of having a larger corpus, multimodal neuroimaging, and developing real-time interventions with an EEG basis in an effort to make it even more applicable in a clinic.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.R. performed data analysis and drafted the initial manuscript. N.P. contributed to visualization, conceptualization, and manuscript editing. V.R. reviewed and edited the draft manuscript. F.R. supervised the study, contributed to study design, and oversaw data collection. All authors reviewed and approved the final version of the manuscript.

Data availability

The datasets presented in this study can be found in dataset article: https://doi.org/10.1016/j.dib.2021.107733.

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-19167-4.