Deep learning methods

In recent years, substantial efforts have been made to tackle the challenge in general EEG pathology classification through the assistance of deep learning approaches. For instance, Schirrmeister et al. [31] designed a 4-layer deep ConvNet architecture namely BD-Deep4 to identify anomalous events in EEGs, achieving an accuracy of 85.42%. Roy et al. [19] put forward a deep 1D convolutional gated recurrent neural network, i.e., ChronoNet, resulting in 86.57 % classification accuracy. In [14], the authors introduced a novel deep one-dimensional CNN model for the discrimination of two EEG patterns. Moreover, there has been an increasing interest in employing transfer learning and hybrid deep learning techniques for the automatic categorization of EEGs into normal or abnormal. Amin et al. [32] deployed a pre-trained AlexNet model with the last layer replaced by an SVM classifier to conduct EEG pattern recognition, reporting an accuracy of 87.32%. Also, in [22], an AlexNet pre-trained in a non-disclosed database was used to extract features from cropped data, accompanied by a Multilayer Perceptron (MLP) for classification. Beyond that, Khan et al. [3] developed a novel hybrid model that integrates CNN-based feature extraction and LSTM-based classification, resulting in 86.23% accuracy. Likewise, Albaqami et al. [33] combined customized WaveNet and LSTM sub-models to differentiate EEG signals and obtain an accuracy rate of 88.76%.

To comprehensively explore the nature of deep learning approaches, recent studies have undertaken a substantial amount of comparative analysis. In [11], various CNN-based models were systematically compared and analyzed on Temple University Hospital (TUH) dataset. The optimized temporal convolutional network with 456,502 trainable parameters demonstrated superior performance in classifying pathological and non-pathological signals. More recently, Kiessner et al. [17] conducted another holistic examination to evaluate the EEG decoding performance of various deep neural networks on an extended EEG dataset, which is five times larger than the TUH dataset. The outcomes showcased that the most complex model with over four hundred thousand parameters achieved the best performance accuracy of 86.59%.

In summary, even though deep learning approaches can yield marginal improvements in the automatic pathological EEG identification, this is at the cost of a more complex model architecture, a greater need for labeled training samples, and increased computational time and storage requirements [16,17,31]. Moreover, collecting well-labeled datasets is arduous and prone to a relatively low inter-rater agreement, while designing a complex yet high-performing end-to-end structure is difficult [11,34]. Furthermore, the spatial correlation embedded in multi-channel EEG data is another essential yet under-explored factor for accurate EEG detection.

Feature-based methods

In recent years, feature-based approaches have gained prominence in EEG pathology diagnosis, due to lower hardware requirements, simpler model architecture, adaptability in learning meaningful features, and other advantages [2,25]. These approaches primarily follow two stages: feature engineering and classification. Notably, the former including feature extraction and feature selection is critical for boosting EEG decoding performance due to the fact that the classification accuracy highly relies on the learned feature.

Up to now, numerous feature extraction methodologies have been proposed to cope with EEG binary classification, which can be broadly grouped into single-domain, dual-domain, and multi-domain techniques. Representative single-domain analysis techniques are temporal domain analysis methods and frequency domain analysis methods. For instance, various spectral features are captured from channel-, segment-, and EEG-level to detect pathological slowing in EEG signals [15]. The dual-resolution analysis techniques, especially the time-frequency analysis, have garnered increasing attention in discerning EEG patterns owing to their ability to mine complementary information among two distinct domains. For example, Cisotto et al. [13] computed eleven well-established time-frequency domain features in each frame of each EEG channel to distinguish normal and abnormal EEGs. Sharma et al. [12] extracted fuzzy entropy, logarithmic of the squared norm, and fractal dimension feature from wavelet sub-bands, obtaining an accuracy rate of 79.34% with the assistance of the SVM classifier. Similarly, Singh et al. [25] converted brain signals into images via STFT and achieved a classification accuracy of 88.04%. In [2], a hypercube-based feature extractor coupled with DWT was used to decompose signals into a series of physically meaningful narrow-band signals, in conjunction with Neighborhood Component Analysis-based feature selection and the KNN classifier, achieving the accuracy of 87.68%. Along a similar line, Gemein et al. [11] devised EEG pathology diagnosis methods based on DWT. More recently, Zhong et al. [24] implemented WPD and various ensemble learning classifiers to classify EEG data, which achieved a state-of-the-art accuracy of 89.13%.

It is evident from the above research that feature-based approaches possess tremendous potential in boosting the binary classification performance of multi-channel EEG records. However, few works took the subtle spatial information into account and made use of the mutual complementarity between time, frequency, and spatial domains, causing a decline in classification performance. Moreover, the redundant and irrelevant features embedded in the extracted features, which may cause overfitting and increased computational costs, pose another challenge for EEG pathology detection. Therefore, we put forward a novel feature-based traditional machine learning framework to distinguish normal and abnormal EEG, which integrates a multi-domain feature fusion and a two-step dimension reduction to provide refined and comprehensive information representation.

Data description and preprocessing

To verify the practicality of the proposed method, a large-scale EEG dataset, known as the TUH Abnormal EEG Corpus [18,35], was employed. This dataset was collected by Temple University Hospital under ethical clearance from the Institutional Review Board and is freely available at [36]. All subjects in the dataset provided informed consent before their data were recorded, and the dataset was anonymized in accordance with ethical guidelines for research involving human subjects. Continuously updated and currently at version 2.0.0, this dataset is the most comprehensive open-source EEG benchmark available for research purposes, ensuring compliance with ethical standards for research involving human subjects.

Multi-domain feature extraction

EEG signals encompass intricate and abundant information in time, frequency, and spatial domains [30]. Relying solely on extracting features from a single domain may fall short of capturing the complex nature of this signal. Hence, to overcome this limitation, we attempt to analyze EEG from multiple perspectives simultaneously in this work.

Multi-resolution dual-domain analysis.

Recently, plenty of temporal-frequency domain analysis techniques have been introduced to enhance the accuracy in discerning abnormal brain signals. These dual-domain techniques excel in revealing intricate details concerning the amplitude and phase variations of different frequency components over time. A common technique is STFT, which can capture the signal’s frequency content and time localization information over a defined sliding window [16]. However, its pre-fixed analysis window renders it non-adaptive in multi-resolution dual-domain analysis. Conversely, the wavelet transform techniques, including DWT and WPD, can realize time subdivisions at high frequencies while frequency subdivisions at low frequencies by stretching and translating wavelet functions. However, DWT typically splits high-frequency bands into more subtle subbands, while WPD processes both low- and high-frequency components and generates more subsequences, exacerbating the problem of redundant and irrelevant information as well as higher time complexity [26,28]. This is the rationale behind the utilization of DWT.

DWT can recursively decompose EEG signals into multi-resolution frequency sub-bands at finite layers, as illustrated in Fig 4. At j-layer  ( j = 1 , 2 , … , J ) , the input is broken into high- and low-frequency sub-bands of the same scale respectively through concurrent convolution with a high-pass filter and a low-pass filter . Following this, the intermediate sub-bands undergo 1/2 downsampling to generate an approximate component (Eq (1)) and a detail component (Eq (2)). The former describes the signal’s long-term trend and reflects overall identity, while the latter captures short-term trends and subtle nuances in sub-bands. Notably, the approximate component typically serves as the input for the (j + 1)-layer. This process is iterated until reaching the final decomposition layer J.(1)(2)where L stands for the length of after j–1 decompositions and τ stands for the scale.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Time-frequency feature extraction on raw EEG signals using discrete wavelet transform.

https://doi.org/10.1371/journal.pone.0310348.g004

The decomposition level and the applied mother wavelet significantly impact the quality of EEG decomposition, since they determine the structure of the filter bank and the resulting frequency components. According to [29,37], brainwave rhythms are categorized basically into five distinct frequency bands: Delta, Theta, Alpha, Beta, and Gamma (above 30 Hz); of these, the first four bands have been widely used to study brain function. As an illustration, the features associated with epileptic seizure predominantly appear in the frequency spectrum below 30 Hz [28], while the study [8] reveals that distinct hemisphere asymmetry differences between affected and unaffected subjects can be investigated by assessing the relative signal power in four bands. The Symlets wavelet of order 6 (sym6), a mother wavelet used in DWT, offers several advantages for EEG signal decomposition. It not only reduces dimensionality but also provides optimal band coverage for the Delta, Theta, Alpha, Beta, and Gamma bands, improving classification accuracy [38]. Furthermore, the orthogonality of the sym6 wavelet prevents redundancy during decomposition and reconstruction, guaranteeing precise restoration of the original signal. This property is especially important in processing complex EEG signals and detecting anomalies [39]. However, when using sym6 for EEG signal decomposition, excessive decomposition can lead to an overrepresentation of low-frequency bands, such as the Delta rhythm. This overemphasis may "confuse" the classifier and diminish classification accuracy [38]. Considering the 250 Hz sample rate of the preprocessed EEG data, a decomposition level of 5 was selected to eliminate irrelevant components while preserving key features. The remaining components— (15.625-31.25 Hz), (7.8125-15.625 Hz), (3.90625-7.8125 Hz), and (0-3.90625 Hz)—are retained, as depicted in Fig 4.

Multi-domain feature fusion

Both single-domain and dual-domain analyses face the challenge of achieving effective EEG decoding, as features specific to the time, frequency, and spatial domains can only capture EEG signals from their respective perspectives. Multi-domain fusion strategy is a solution to address this challenge as it is proven to be effective in many other EEG analysis tasks [1,41]. The key motivator behind this strategy is its effectiveness in compensating for the lack of complementarity among heterogeneous characteristics from different domains [1,30]. Nevertheless, a straightforward feature fusion has the hidden danger of high-dimensional feature space. Specifically, according to the preprocessing applied to the TUH Abnormal EEG dataset, it can be inferred that, for each sample, there is a total of 21 channels and 100 segments respectively. Consequently, the time-frequency feature dimension per sample will then be 21 × 100 × 24 = 50400 according to Eq (3). Compared to the spatial feature dimension, which is only 8, it is clear that the former’s dimensionality is extremely high. Such a high dimensionality may give rise to increased model complexity, substantial computational costs, and other issues. As a result, it is necessary to decrease the dimensionality of time-frequency feature space.

Feature aggregation is an excellent countermeasure to condense the extracted features, without compromising feature quality, through the aid of the aggregation function [20]. Therefore, we adopt three Hjorth parameters (i.e., activity, mobility, and complexity) to aggregate each time-frequency statistical feature in every EEG sample from multiple views, as present in Fig 5, owing to their robustness, superior computational efficiency, and the strength of inter-class separation and intra-class aggregation [42]. Take mean for an example:(9)(10)(11)where, , , and denote the three feature aggregation operations using activity, mobility, and complexity functions, respectively. Thus, time-frequency features per sample are transformed into three C × 24 dimensional feature matrices. Then, the aggregated features are fused with spatial features to provide a more comprehensive signal representation in a relatively low-dimensional space. In addition, considering the pronounced correlation between the patient’s age and EEG signals [3,17], we incorporated it into the feature vector.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Process of feature aggregation.

https://doi.org/10.1371/journal.pone.0310348.g005

Statistical significance analysis

Reducing the fused feature space helps attain a set of features with the best class discrimination ability. However, this process poses a significant challenge for EEG pathology detection. Feature selection, a crucial aspect of this endeavor, eliminates redundant and irrelevant features to reduce the computational burden. Recently, various methods (e.g., neighborhood component analysis, Relief, etc.) have been attempted, but most of these methods are vulnerable to noise interference and computationally expensive, especially for EEG signals with low signal-to-noise ratio and high dimensional small samples, potentially reducing the classification accuracy and efficiency [1,2]. For instance, Relief-based feature selection requires multiple searches and calculations of similar samples for each feature, so the calculation cost is high for high-dimensional data sets [43]. Additionally, it is highly sensitive to outliers and noisy data, and its dependence on the distribution of the dataset may lead to important features being overlooked in cases of inter-class imbalance. In contrast, by independently evaluating the statistical relationship between each feature and the target variable, statistical significance analysis typically has faster computation speeds and greater robustness to class imbalances [44]. Therefore, this study applied this method to select highly distinguishable features.

As a rank-based, non-parametric statistical significance analysis method, the Kruskal-Wallis test does not rely on strict distributional assumptions, making it well-suited for feature selection in noisy datasets, like EEG data. Moreover, given the non-normal distribution of features, the Kruskal-Wallis test [45] is utilized to examine features with the most significant statistical impact in classification. The Kruskal-Wallis test formulates the null hypothesis () that no statistically significant difference exists between independent feature groups, while the alternate hypothesis () hypothesizes a difference between them. Highly distinguishable features are selected by testing the null hypothesis. In particular, features are ranked according to their discriminative power, and the p-value of each feature is computed and compared with the level of significance α. Here, α is set to be 0.01 in this work, reflecting a 99% bootstrap Confidence Interval level. If p ≤ α, is accepted, indicating that these features are significant for discrimination, and retained. Conversely, when p > α, we reject and discard these insignificant features. Furthermore, a smaller p-value implies a more important feature for the given task. Finally, all features with the lower p-values (p < 0 . 01) are used to construct a low-dimensional feature matrix and then presented to the classifier.

Classification

After feature engineering, the last phase is to design a classifier for accurately determining EEG classes. As a multi-classifier ensemble algorithm, GBDT has some advantages, including adaptability to various data distributions and varied feature types, as well as the capability to handle complex nonlinear relationships, contributing to its high predictive accuracy across a wide range of applications [4,20,24]. In light of the nonlinear and non-Gaussian nature of EEG data, we employed GBDT’s recent and prominent implementations, namely CatBoost, XGBoost, and LightGBM, for discriminating between normal and abnormal EEG signals and evaluating the proposed feature engineering, as detailed in Section Performance Testing on the Extracted Features with Different Classifiers. The results indicate that CatBoost achieves the best performance among these three classifiers (refer to Fig 6). Consequently, we integrated the proposed feature engineering with CatBoost for comparison against several existing approaches.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Performance of three classifiers.

https://doi.org/10.1371/journal.pone.0310348.g006

Experimental setup

Extensive experiments were conducted on the TUH Abnormal EEG database, which covers various abnormal EEG events such as epilepsy, sleep disorders, and brain injuries, allowing for a comprehensive evaluation of the proposed method’s diagnostic ability for EEG signals related to different brain conditions. Raw EEG data was preprocessed through standard 21-channel selection, downsampling to 250Hz, and 5 s non-overlapping segmentation. In the feature extraction stage, six wavelet-based time-frequency features, i.e., mean, mean deviation, standard deviation, mean absolute value, skewness, and kurtosis were extracted, as elaborated in Section Time-frequency Statistical Feature Extraction. Spatial features were derived from the constructed time-frequency feature space via CSP with eight decomposition components. Simultaneously, activity, mobility, and complexity mapped the time-frequency features into a lower-dimensional feature space, which is subsequently integrated with spatial and age features. To mitigate redundant and irrelevant features, we chose highly discriminative features with p-values below 0.01 and fed them into CatBoost, XGBoost, and LightGBM. The hyper-parameters of these classifiers were fine-tuned as follows: 1080 estimator counts, maximum depths of 4, 6, and 4, and learning rates of 0.03, 0.02, and 0.04, respectively; other hyper-parameters were kept at default values. Consistent with established practices [20,32], the system was separately trained and tested on mutually independent training and testing sets, both sourced from TUH Abnormal EEG. Importantly, this ensures no data leakage at the patient level, thereby simulating cross-patient scenarios in the real world and helping to validate the model’s generalization ability.

To evaluate the effectiveness of pathological EEG detection, three common evaluation metrics, namely accuracy, F1-score, and G-mean, are utilized. Accuracy is an intuitive and standard criterion that represents the proportion of correctly classified instances relative to all instances. Besides, due to the unbalanced class distributions in the training set (refer to Table 1), accuracy is insufficient for comprehensive assessment [16]. Hence, we incorporated both the F1-score and G-mean, each offering a balanced view of the model’s performance, to optimize our evaluation. F1-score serves as a metric that harmoniously combines precision and recall through the computation of their harmonic mean, while the G-mean represents the exact geometric mean of the recalls for both positive and negative classes.(12)(13)(14)where, TP (True Positive) and TN (True Negative) denote the accurate identification counts of positive and negative instances, respectively. FP (False Positive) represents the incorrect assignment of normal instances to the abnormal class, while FN (False Negative) indicates the misclassification of abnormal instances. The confusion matrix, as present in Table 2, encompasses these four metrics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Confusion matrix of EEG detection.

https://doi.org/10.1371/journal.pone.0310348.t002

Performance testing on the extracted features with different classifiers

To evaluate the effectiveness and broad applicability of the proposed feature engineering, this study employed three advanced ensemble learning classifiers: CatBoost, XGBoost, and LightGBM, for categorizing the extracted EEG pathological features. The results obtained using each classifier are presented in Fig 6.

Fig 6 shows that, among these ensemble learning classifiers, the CatBoost-based framework exhibited a superior accuracy rate of 90.22% in distinguishing between non-pathological and pathological EEGs. It is noteworthy that this accuracy exceeds the clinical diagnostic standard, which typically mandates a reliable system to attain an accuracy of 90% for the diagnosis of abnormal EEG [11], highlighting its medical reliability and practical utility. Meanwhile, the high F1-score of 88.89% indicates that this method minimizes both false negatives (which could delay critical interventions) and false positives (which could lead to unnecessary treatments), and the high G-mean of 89.76% reflects the model’s consistent performance across both pathological and non-pathological cases. Together, these metrics demonstrate the method’s reliability and generalizability for clinical abnormal EEG detection. When applying XGBoost and LightGBM, both frameworks demonstrated commendable classification accuracy, achieving 87.32% and 86.59% respectively. These favorable performances are also evident in the F1-score and G-mean, confirming the extensive applicability, stability, and generalizability of the proposed feature engineering across different machine-learning frameworks, as well as its good compatibility with diverse abnormal EEG data. In a word, our feature engineering attains satisfactory classification performance, with accuracy consistently exceeding 86.5% in all cases, robustly supporting the feasibility of multi-domain feature fusion and two-step dimension reduction. Additionally, the outstanding performance of the CatBoost-based approach prompts us to utilize it in subsequent experiments.

The confusion matrices for EEG pathology detection using the above three frameworks on the TUH Abnormal EEG Database are presented in Fig 7. We can observe that, based on the input features, CatBoost misclassified 18 out of 126 abnormal instances and 9 out of 150 normal instances, resulting in a higher false negative rate of 14.29% than the false positive rate of 6%. Similar phenomena, where the model excels in accurately identifying normal EEGs, are evident in the other two classifiers. These findings substantiate that our proposal, when combined with all three classifiers, exhibits high sensitivity to normal class, which is especially pivotal in automated diagnosis [20]. High sensitivity for the normal class helps ensure a low miss rate when diagnosing normal EEG, preventing delays in treatment or intervention [20]. Therefore, the features proposed in this study effectively enhance the reliability and practicality of automated diagnostic systems, particularly in clinical practice, where accurate identification of normal cases is crucial for patient health. Interestingly, this result is in alignment with previous research studies [11,32]. The reason could be the unequal distribution of different classes in the training set (refer to Table 1), specifically the larger quantity of normal EEG, which could induce a bias towards this class in the method.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Confusion matrices of three classifiers.

https://doi.org/10.1371/journal.pone.0310348.g007

Comparative analysis with existing state-of-the-art methods

To demonstrate the superiority of our methodology, a comparison against several representative deep learning and feature-based approaches was carried out on the TUH Abnormal EEG Corpus, as presented in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Comparison of the classification results obtained by different EEG pathology diagnosis approaches on the real-world EEG abnormal dataset.

https://doi.org/10.1371/journal.pone.0310348.t003

Table 3 illustrates that our method exhibits superior performance than other state-of-the-art frameworks. Through comprehensive analysis, we can obtain the following conclusions:

1. (i) In contrast to four deep learning methodologies, the proposed framework consistently achieves the highest accuracy, F1-score, and G-mean. Specifically, compared against two transfer learning methodologies [22,32], both of which were additionally trained on undisclosed EEG data, our approach exhibits notable improvements of 1.09% and 2.90% in accuracy, 1.83% and 3.92% in F1-score, as well as 1.74% and 3.52% in G-mean. This phenomenon can be ascribed to the dependency of deep learning methods on complex end-to-end structures with numerous parameters and a substantial amount of training samples, causing a potential risk of overfitting and increased computational resources. Moreover, the black-box nature of deep learning approaches is another factor impacting their performance. The opaque internal decision logic of these models results in poor interpretability which is crucial in medical applications to ensure diagnostic reliability and safety. Besides, the oversight of subtle spatial information restricts the upper limitation of these methodologies’ performance. In contrast, our study turns to a simple and reliable multi-domain feature-based approach, which achieves higher EEG detection accuracy with lower computational resource consumption.
2. (ii) Also, our framework outperforms five advanced feature-based methods. Specifically, it surpasses two methodologies [2,11] that employ the same time-frequency feature extraction technique as ours, namely DWT. This advancement can be attributed to three key advantages. Firstly, the proposed spatial feature representation compensates for the lack of time-frequency features. The second factor is the complementary information from multi-domain fusion features through comprehensive multi-feature fusion. Thirdly, a two-step dimension reduction strategy efficiently diminishes data dimensions and selects more representative features for EEG pathological detection. On the other hand, our method outperforms the suboptimal feature-based approach [24] in accuracy, F1-score, and G-mean by 1.09%, 1.29%, and 1.16%, respectively. It is noteworthy that this work employs the same classifier as our framework, further signifying the effectiveness of the proposed feature engineering.

In summary, compared to other deep learning and feature engineering methods, the proposed approach significantly outperforms in classification accuracy, F1 score, and G-mean, demonstrating its robustness and broad applicability to real-world EEG data. This is primarily due to the integration of time, frequency, and spatial domain features, fully leveraging their complementary nature to improve signal representation, which results in better generalization across different patients. By extracting spatial features from denoised time-frequency information, the model’s ability to represent multi-dimensional features is further enhanced. Additionally, the two-step dimension reduction strategy reduces feature dimensions while retaining the most discriminative features, minimizing overfitting risk and enhancing the model’s adaptability to high-dimensional data.

Ablation study

The multi-view feature aggregation and spatial information learning are two pivotal elements that affect the performance and efficiency of the EEG binary classification. Hence, in this section, two ablation experiments were conducted on the TUH Abnormal EEG dataset to further investigate and evaluate the individual impact of these two components.

Effect of multi-view feature aggregation.

The multi-view feature aggregation mechanism is a practical implement to reduce the dimension of time-frequency feature space and simplify the model complexity. To fully investigate the effect of feature aggregation, we performed an ablation experiment by gradually increasing the aggregate function. Here, four different cases were compared and analyzed, including:

• Case-1: Without feature aggregation, time-frequency, spatial, and age features are directly concatenated into a feature vector and then input into CatBoost.
• Case-2: Time-frequency features are aggregated using activity and then concatenated with spatial and age features.
• Case-3: Time-frequency features are aggregated using activity and mobility simultaneously and then concatenated with spatial and age features.
• Case-4: Time-frequency features are aggregated using activity and complexity simultaneously and then concatenated with spatial and age features.
• Proposed method: Time-frequency features are aggregated using activity, mobility, and complexity simultaneously and then concatenated with spatial and age features.

To ensure experimental fairness, this study utilized a uniform feature set and a unified CatBoost classifier across all cases. Additionally, the execution time covering feature learning and classification was taken into account for assessing the effect of the mechanism in classification efficiency. Table 4 presents the outcomes of five cases. A comparative analysis reveals several key insights:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Ablation study of multi-view feature aggregation.

https://doi.org/10.1371/journal.pone.0310348.t004

(1) Even without utilizing feature aggregation, Case-1 exhibits a classification accuracy of 81.52%, surpassing the comparative method [13]. This improvement is consistent across the other two indicators, revealing that the proposed spatial information learning and the statistical significance analysis-based feature selection can improve the accuracy of EEG decoding. (2) It is evident that the performance of feature aggregation methods demonstrates a monotonic increase as the aggregation function increments. Specifically, in contrast to Case-1, Case-2 adopting one Hjorth parameter (i.e., activity) yields an accuracy improvement of 4.71% and a substantial time consumption reduction of 140.78 seconds. This phenomenon can be attributed to that feature aggregation effectively compresses high-dimensional features, thereby improving feature recognizability, reducing computational overhead, and enhancing classification performance. Compared to Case-2, Case-3 and Case-4 incorporating a second Hjorth parameter, namely mobility and complexity respectively, lead to improvements in accuracy by 1.81% and 0.73%, F1-scores of 1.98% and 1.05%, along with G-mean of 1.79% and 0.9%. Further analysis revealed that when all Hjorth parameters were employed, the proposed method achieved the peak values across accuracy, F1-score, and G-mean. These results suggest that each Hjorth parameter provides unique insights regarding amplitude, frequency, and other waveform characteristics so that multi-view information integration can comprehensively characterize the time-frequency features, thereby enhancing the power of aggregated features for discriminating EEG signals. Although our method costs a slightly higher time cost compared to Cases 2, 3, and 4, its optimal classification performance justifies the time investment. Meanwhile, in comparison to Case-1, our method reduced execution time by 133.63 seconds and significantly increased the accuracy rate of EEG signal detection by 8.70%.

In order to examine the representational capacity of the features aggregated across the activity, mobility, and complexity dimensions, their p-values are presented in the form of heatmaps, as illustrated in Fig 8. Over 50% of the aggregated features in three figures have p-values below 0.01, especially for the activity-aggregated features exceeding 93%. This demonstrates that the chosen frequency bands contain rich EEG anomaly information, and most selected time-frequency statistical features contribute significantly to subsequent tasks. The multi-view feature aggregation is further proven to not only mitigate redundancy, thus lowering computational complexity and avoiding overfitting, but also combine the advantages of different features to form a more comprehensive and discriminative information representation, which helps better capture complex patterns in EEG signals. Nevertheless, a few features still exhibit lower discriminative power. To address this, we further optimize the feature set using the Kruskal-Wallis test. This mechanism allows the classification model to focus on learning the most informative features so that the model’s efficiency, speed, and generalization capability are enhanced.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Heatmap of p-values for time-frequency features aggregated across activity, mobility, and complexity dimensions.

The color intensity in the heatmap corresponds to the p-value size, representing feature statistical significance for classification.

https://doi.org/10.1371/journal.pone.0310348.g008

In summary, these experimental results substantiate that this multi-view feature aggregation mechanism effectively condenses information by diverse viewpoints to a lower-dimensional and more discriminative information representation. Therefore, this mechanism not only enables the method to efficiently deal with high-dimensional data but also accelerates the execution speed of EEG classification.

Effect of Spatial Information Learning.

In recent years, the majority of feature engineering focuses on exploring time, frequency, and time-frequency domains, with scant attention given to the spatial domain, which is pivotal for improving the precision of EEG pathology detection. In this subsection, to validate the vital effect of spatial information, we compare and analyze two distinct cases: (1) the resultant feature set contains the spatial domain features and (2) the spatial features are left out before classification. To mitigate the inherent impact of the classifier on comparisons, three ensemble learning classifiers, consistent with those in Section Performance Testing on the Extracted Features with Different Classifiers, were used to handle the feature sets. The classification performances of these two scenarios are illustrated respectively in Fig 6 and Fig 9.

Fig 9 depicts the results of Case-1 exclusively learning the time-frequency features. It is evident that CatBoost-based approach achieves accuracy, F1-score, and G-mean of 87.32%, 85.71%, and 86.92% respectively, surpassing most contrast algorithms [11,13,31,32]. This remarkable performance provides supporting evidence for the excellence of the proposed time-frequency feature representation. If these results are further compared against the results in Fig 6, a substantial improvement in classification performance can be observed across all three approaches, attributable to the inclusion of time-frequency spatial information. Taking Catboost as an example, there is a prominent increase in accuracy, F1-score, and G-mean by 2.90%, 3.18%, and 2.84%, respectively. It indicates that the latent spatial information, which carries the important spatial distribution of different classes within the multi-channel EEG signals, effectively compensates for the information gap in the time-frequency feature. Interestingly, this result aligns with [5] affirming that the combination of time, frequency, and spatial features effectively enhances seizure detection. In a word, in the field of abnormal EEG diagnosis, incorporating spatial domain features and fusing multi-domain features enable more accurate detection of subtle EEG changes linked to abnormal brain activity, providing a more reliable technical foundation, which traditional methods may fail to capture.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Performance of three classifiers (without spatial features).

https://doi.org/10.1371/journal.pone.0310348.g009