Brain state visual identification method based on interaction between channel dynamic characteristics and cortical network

Abstract

The invention discloses a visual brain state recognition method for interaction between channel dynamic characteristics and a cortical network, and belongs to the technical field of electroencephalogram characteristic decoding. The method comprises the following steps: acquiring an original EEG signal; performing preprocessing and artificial feature extraction on the original EEG signal to obtain a channel input feature and a channel connectivity input feature; constructing a transparent brain state learning model; and inputting the input features and the channel connectivity input features into a transparent brain state learning model, and outputting a brain state visual recognition result. According to the method, time-space characteristics and connection modes of neural oscillation are captured in a multi-dimensional mode to serve as a priori graph of transparent decoding, convolution of different dimensions is constructed to achieve multi-dimensional space mapping of key characteristics, a Kolmogorov-Arnold representation theorem is introduced to predict a brain state mode, meanwhile, a transparent EEG decoder is adopted for fine-grained revelation, and the method is suitable for being applied to the field of brain state prediction. And a powerful interpretable theoretical support is provided for analyzing brain space activation modes in different states.

Description

Technical Field

[0001] This invention belongs to the technical field of EEG feature decoding, specifically relating to a method for visualizing and recognizing brain states based on channel dynamic characteristics and cortical network interaction. Background Technology

[0002] Brain-computer interfaces (BCIs), as a technology for detecting physiological signals in the brain, can capture relevant neural activity induced by tasks. Among them, electroencephalography (EEG) is widely used to analyze brain state patterns, such as those related to attention and mental fatigue, due to its high temporal resolution and non-invasiveness. Furthermore, decoding dynamic brain activity can provide important data support for clinical medical research.

[0003] Recently, the combination of physiological signals and deep neural networks has demonstrated significant research value. Particularly in pattern recognition, deep neural networks, through multi-layer stacking, can extract highly discriminative high-level features from raw data, overcoming the limitations of traditional methods in feature representation. Simultaneously, existing research has designed various feature engineering techniques to better adapt models to complex EEG signals. For example, mapping EEG signal sequences from multiple electrode channels to a feature plane. To further explore the potential features reflected in EEG signals, many studies have transformed them into other forms of high-level representation. This includes mapping signals to a compact two-dimensional matrix based on the spatial distribution of electrodes in the brain, and treating each EEG signal as an independent time-series feature, capturing its time-varying characteristics. However, the high uncertainty and complexity of deep neural networks make the mapping relationship between input and output difficult to interpret directly using traditional analytical methods.

[0004] In neurophysiology, good interpretability provides new research perspectives and theoretical basis for decoding brain functional networks and their dynamic regulatory mechanisms. Furthermore, combining neural information such as functional connectivity networks, neural oscillation characteristics, and functional localization of specific brain regions can lead to the development of more biologically meaningful deep neural networks. In existing research models, attention mechanisms are often used to dynamically allocate weights, which are then visualized to provide interpretability. For example, the weights in channel attention modules are directly presented as brain topography to analyze spatial distribution patterns of brain activation. Existing techniques use circular connectivity graphs and three-dimensional spatial connectivity visualization to analyze inhibitory relationships between channels. However, while attention weights indicate the model's attention to input features, these weights themselves are not entirely equivalent to the model's decision-making logic. Moreover, the attention mechanism captures highly sensitive key information but lacks spatial consistency, thus its interpretation may be misleading. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a method for visualizing and recognizing brain states based on channel dynamics and cortical network interactions. This method aims to solve the problem that existing attention mechanisms capture key information with high sensitivity but lack spatial consistency, which may lead to misleading interpretations.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for visualizing and recognizing brain states based on channel dynamics and cortical network interactions, comprising the following steps: S1. Obtain the raw EEG signal; S2. Preprocess the raw EEG signal and extract artificial features to obtain channel input features and channel connectivity input features; S3. Construct a transparent brain state learning model; S4. Input the input features and channel connectivity input features into the transparent brain state learning model, and output the brain state visualization recognition results.

[0007] Furthermore, in step S2, the raw EEG signal is preprocessed, specifically as follows: The original EEG signal is divided into different frequency bands using bandpass filtering: In the formula, This represents the signal value after filtering the c-th channel, where... The location of the signal sampling point. T Indicates the number of sampling points; This represents the output signal calculated at time nk after processing by the bandpass filter; Represents the normalization coefficient; M and N represent respectively and The order; k represents the time offset; This represents the original signal acquired at time nk; and These represent the coefficients corresponding to the lower cutoff frequency and the upper cutoff frequency, respectively.

[0008] Furthermore, in S2, the manual feature extraction includes: For the bandpass filtered frequency band signal, the differential entropy is used as its signal feature to obtain the channel input feature. Specifically, the differential entropy is expressed as: In the formula, This represents the value of the calculated differential entropy; This represents a slice of the filtered signal from which the differential entropy is calculated. Represents the Gaussian probability density function; express Variance of the slices; Represents the natural base; Calculate channel connectivity input features based on channel input features: In the formula, This represents the connectivity value between channels i and j; , These represent the signal values ​​of the i-th channel and the j-th channel, respectively. and Let represent the average signal value of the i-th channel and the average signal value of the j-th channel, respectively; This indicates the maximum number of channels in the brainwave.

[0009] Furthermore, in S3, the transparent brain state learning model includes: a channel feature and functional connectivity integrated network and a transparent EEG decoder; The channel feature and functional connectivity integration network includes multiple residual blocks with convolutional kernels of different dimensions and a KAN network; the output of the residual blocks serves as the input of the KAN network.

[0010] Furthermore, step S4 specifically includes the following sub-steps: S41. Input the channel input features and channel connectivity input features into the residual blocks of convolutional kernels of different dimensions for feature sampling, and obtain the channel feature map and connectivity feature map respectively; S42. Perform global average pooling on the channel feature map and the connectivity feature map respectively to obtain the final channel convolutional feature and the final connectivity feature. Then, concatenate the final channel convolutional feature and the final connectivity feature to obtain the fused feature. S43. Input the fused features into the KAN network to predict brain state patterns, and then output the brain state recognition results. S44. Input the channel feature map and connectivity feature map output from the last residual block, as well as the brain state recognition result, into the transparent EEG decoder to calculate the sensitivity of the feature map. S45. Based on the sensitivity and activation pattern of the feature map, the activation map is calculated. The activation map is then fused with the spatial information of the original EEG signal to obtain an interpretive image that visualizes the brain state.

[0011] Furthermore, S41 specifically includes: The channel input features are sampled from the residual block of the one-dimensional convolution kernel, which is represented as: In the formula, Indicates the first lChannel feature maps output by the convolutional layer; Indicates the kernel size of the convolution; The weights of the one-dimensional convolution kernel; Indicates channel input characteristics; Indicates the position index in the convolution weight matrix; The channel connectivity input features are sampled from the residual blocks of the two-dimensional convolution kernel, which is represented as follows: In the formula, For the first l The connectivity feature map output by the convolutional layer; The weights are those of the two-dimensional convolution kernel. Indicates channel connectivity input features; , These represent the position coordinates of the corresponding convolution kernel weights.

[0012] Furthermore, in step S42, global average pooling is performed on the channel feature map and the connectivity feature map respectively to obtain the final channel convolutional features and connectivity features, which are represented as follows: In the formula, and These represent the final channel convolutional features and the final connectivity features, respectively. , This indicates that the maximum value is taken in the corresponding channel feature map; This represents the feature value at position (i,j) in the feature map of channel c. This represents the feature value at position (i,j) in the c-th channel feature map of the connectivity feature.

[0013] Furthermore, in step S43, the fused features are input into the KAN network to predict brain state patterns, and then the brain state recognition result is output, which is represented as follows: In the formula, Indicates the output weights; for function; Indicates fusion characteristics; Indicates the first The function matrix corresponding to the layer, Indicates the first The function matrix corresponding to the layer, This represents the function matrix corresponding to the 0th layer; Use the softmax function to adjust the output weights After normalization, the brain state recognition result is output, which is represented as follows: In the formula, The results of brain state recognition The value is the score for each category; This indicates normalization.

[0014] Furthermore, in S44, the sensitivity of the feature map is calculated, which is expressed as: in, In the formula, Indicates the sensitivity of the feature map; , These represent the height and width of the figure, respectively. Indicates the weighted gradient; Indicates the target class to be predicted; This represents the eigenvalues ​​of the feature map in the last layer; , These represent the horizontal and vertical coordinates of the last layer of the feature map, respectively. This represents the feature value at position (a, b) on the feature map of the specified c-th channel in the last layer.

[0015] Furthermore, S45 specifically includes: calculating the activation map based on the sensitivity and activation mode of the feature map, which is represented as: In the formula, Indicates activation mapping; This represents the activation function.

[0016] The brain state visualization and recognition method based on channel dynamic characteristics and cortical network interaction provided by this invention has the following beneficial effects: This invention first constructs a feature engineering approach, preprocessing and manually extracting features from the raw EEG signal to obtain the node and edge information of each channel. Then, it captures the spatiotemporal characteristics and connection patterns of neural oscillations in multiple dimensions as a priori maps for transparent decoding. To further capture the connectivity of spatial topology, convolutions of different dimensions are introduced to perform multidimensional spatial mapping of key features, maintaining the stability of each channel's characteristics while enhancing the ability to capture the spatial receptive field. Finally, the Kolmogorov-Arnold representation theorem is used to predict brain state patterns. Simultaneously, the carefully designed transparent EEG decoder improves the clarity of the underlying reasoning process of the entire algorithm's decision-making, obtaining fine-grained interpretations of the channels. A lightweight and interpretable network architecture is introduced, ensuring spatial consistency through multidimensional convolutional kernel residual structures, and a classifier is built based on the Kolmogorov-Arnold representation theorem, enhancing the model's generalization ability.

[0017] 2. This invention employs a TransDec (TransDec) to visualize and interpret the model's decision-making process by restoring the activation mapping of the feature map during the process of model recognition of the target brain state.

[0018] 3. This invention has verified the generalization ability and performance of the model on a variety of brain cognitive tasks, so as to comprehensively explore the functional characteristics of the brain under different states. Attached Figure Description

[0019] Figure 1 This is a comparison chart of the effects of different frequency bands on fatigue data in Example 2.

[0020] Figure 2 This is a comparison chart of the effects of different frequency bands on the attention data in Example 2.

[0021] Figure 3 This is a visualization analysis of the channel activation mapping in the conscious state on the fatigue dataset in Example 2.

[0022] Figure 4 This is a visualization analysis of the channel activation mapping in the fatigue state on the fatigue dataset in Example 2.

[0023] Figure 5 This is a visualization analysis of the channel activation mapping in the syncope state on the fatigue dataset in Example 2.

[0024] Figure 6 This is a visualization analysis of channel activation in the confirmed state on the attention dataset in Example 2.

[0025] Figure 7 This is a visualization analysis of channel activation in an uncertain state on the attention dataset in Example 2.

[0026] Figure 8 This is an interpretive visualization analysis of the connectivity activation mapping in the waking state on the fatigue dataset in Example 2.

[0027] Figure 9 This is an interpretive visualization analysis of the connectivity activation mapping in the fatigue state on the fatigue dataset in Example 2.

[0028] Figure 10 This is an interpretive visualization analysis of the connectivity activation mapping in the syncope state on the fatigue dataset in Example 2.

[0029] Figure 11 This is an interpretive visualization analysis of the connectivity activation mapping in the confirmed state on the attention dataset in Example 2.

[0030] Figure 12 This is an interpretive visualization analysis of the connectivity activation mapping in the uncertain state on the attention dataset in Example 2.

[0031] Figure 13 This is a network structure diagram of the brain state visualization and recognition method based on the interaction between channel dynamic characteristics and cortical networks in Example 1.

[0032] Figure 14 This is a flowchart of the brain state visualization and recognition method based on the interaction between channel dynamic characteristics and cortical networks in Example 1. Detailed Implementation

[0033] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0034] Example 1 This embodiment provides a method for visualizing and recognizing brain states based on channel dynamics and cortical network interactions, referencing... Figure 13 and Figure 14 Specifically, it includes the following: S1. Obtain the raw EEG signal ,in, R Represents the set of real numbers; C Indicates the number of channels. T Indicates the number of sampling points.

[0035] S2. Preprocess the raw EEG signal and extract artificial features to obtain channel input features and channel connectivity input features; The original EEG signal Feature engineering is performed to improve the performance and generalization ability of the model. Bandpass filtering is used to transform the raw EEG signal... Signals divided into different frequency bands are widely used to study the characteristics of brain nerve tremors. Among them, five frequency bands [0.5,4], [4,8], [8,12], [12,30], and [30,50] were used, and their filtering process is expressed as follows: In the formula, This represents the signal value after filtering the c-th channel, where... The location of the signal sampling point. T Indicates the number of sampling points; This represents the output signal calculated at time nk after processing by the bandpass filter; Represents the normalization coefficient; M and N represent respectively and The order of; k Indicates the time offset; This represents the original signal acquired at time nk; and These represent the coefficients corresponding to the lower cutoff frequency and the upper cutoff frequency, respectively.

[0036] Manual feature extraction, including: To reduce the computational complexity of the model after filtering, differential entropy is used as the signal feature for the bandpass filtered frequency band signal, thereby obtaining the channel input features. , as input to a branch of the transparent brain state learning model; Specifically, the differential entropy is expressed as: In the formula, This represents the value of the calculated differential entropy; This represents a slice of the filtered signal from which the differential entropy is calculated; assuming the original EEG signal... If it follows a Gaussian distribution, then Represents the Gaussian probability density function; express Variance of the slices; Represents the natural base; Calculate channel connectivity input features based on channel input features. This is used as input to another branch of the transparent brain state learning model, where the channel connectivity input feature is represented as: In the formula, This represents the connectivity value between channels i and j; , These represent the signal values ​​of the i-th channel and the j-th channel, respectively. and Let represent the average signal value of the i-th channel and the average signal value of the j-th channel, respectively; This indicates the maximum number of channels in the brainwave.

[0037] S3. Construct a transparent brain state learning model; The transparent brain state learning model includes: a channel feature and functional connectivity integrated network and a transparent EEG decoder; The channel feature and functional connectivity integration network includes multiple residual blocks with convolutional kernels of different dimensions and a KAN network; the output of the residual blocks serves as the input of the KAN network; and the output of the KAN network serves as the input of the transparent EEG decoder.

[0038] S4. Input the input features and channel connectivity input features into the transparent brain state learning model, and output the brain state visualization recognition results. This specifically includes the following steps: S41. Input the channel input features and channel connectivity input features into the residual blocks of convolutional kernels of different dimensions for feature sampling, and obtain the channel feature map and connectivity feature map respectively; Specifically, the channel input features are sampled from the residual block of the one-dimensional convolution kernel, which is represented as: In the formula, Indicates the first l Channel feature maps output by the convolutional layer; Indicates the kernel size of the convolution; The weights of the one-dimensional convolution kernel; Indicates channel input characteristics; This represents the index of the position in the convolution weight matrix. The channel connectivity input features are sampled from the residual blocks of the two-dimensional convolution kernel, which is represented as follows: In the formula, For the first l The connectivity feature map output by the convolutional layer; The weights are those of the two-dimensional convolution kernel. Indicates channel connectivity input features; , These represent the position coordinates of the corresponding convolution kernel weights.

[0039] S42. Perform global average pooling on the channel feature map and the connectivity feature map respectively to obtain the final channel convolutional feature and the final connectivity feature. Then, concatenate the final channel convolutional feature and the final connectivity feature to obtain the fused feature. The final channel convolution features and connectivity features are represented as follows: In the formula, and These represent the final channel convolutional features and the final connectivity features, respectively. , This indicates that the maximum value is taken in the corresponding channel feature map; This represents the feature value at position (i,j) in the feature map of channel c. This represents the feature value at position (i,j) in the c-th channel feature map of the connectivity feature.

[0040] S43. Input the fused features into the KAN network to classify and predict brain state patterns, and then output the brain state recognition results. Specifically, the fused features are input into the KAN network: In the formula, , This indicates the output weights, and cla represents the number of recognized states; for function; Indicates fusion characteristics; Indicates the first The function matrix corresponding to the layer, Indicates the first The function matrix corresponding to the layer, This represents the function matrix corresponding to the 0th layer; in, In the formula, In the eigenvector exist l The result of the layer vector matrix is ​​specifically represented as follows: In the formula, x Represents the input feature values. and Let these represent the basis functions and their weights, respectively. and These represent the spline function and the weights, respectively.

[0041] Use the softmax function to adjust the output weights After normalization, the brain state recognition result is output, which is represented as follows: In the formula, , The results of brain state recognition The value is the score (logits) for each category; This indicates normalization.

[0042] S44. Output the channel feature map from the last residual block. and connectivity feature map The brain state recognition results are input into the transparent EEG decoder to calculate the sensitivity of the feature map; The gradient with respect to the feature map is calculated using the output score (logits) of the target class. The weighted gradient calculation, which more accurately captures the sensitivity of feature maps from multiple relevant regions, is expressed as follows: in, In the formula, Indicates the sensitivity of the feature map; , These represent the height and width of the figure, respectively. This represents a weighted gradient, calculated relative to the feature map using the output score (logits) of the target category, to more accurately capture multiple relevant regions. Indicates the target class to be predicted; This represents the eigenvalues ​​of the feature map in the last layer; , These represent the horizontal and vertical coordinates of the last layer of the feature map, respectively. This represents the feature value at position (a, b) on the feature map of the specified c-th channel in the last layer.

[0043] S45. Based on the sensitivity and activation mode of the feature map, the activation map is calculated. The activation map is then fused with the spatial information of the original EEG signal to construct various interpretable images. These images will provide an interpretable visual representation for the predictive behavior of the deep neural network. The activation mapping is computed and represented as: In the formula, Indicates activation mapping; This represents the activation function.

[0044] Example 2 This embodiment is used to verify the model (TBSL) or method in Embodiment 1, and specifically includes the following: Dataset: The proposed TBSL was validated using two datasets related to cognitive levels.

[0045] The Fatigue Dataset (SEED-VIG) contains electroencephalograms (EEGs) and electrooculograms (EOGs) collected during simulated driving scenarios in front of a screen. The EEG data was collected from 23 subjects using 17 electrode channels. The EEG characteristics were analyzed at 200 Hz across five frequency bands. Furthermore, the PERCLOSE index was calculated every 8 seconds to assess alertness as a measure of fatigue.

[0046] Attention Dataset: This dataset is a multimodal brain imaging dataset used to measure three cognitive tasks in healthy subjects. A discrimination / selection response task was used for attention assessment. The experiment included 26 participants, each using the first session of the three cognitive tasks. Each phase consisted of several segments of the attention task (40 seconds) and a rest period (20 seconds). 28 EEG channels and 2 EEG channels were recorded at a sampling rate of 1 kHz; this embodiment only uses EEG data.

[0047] Network configuration: The network designed in this embodiment adjusts its output based on the number of classes in different datasets. TBSL employs two sampling branches and a classifier, with parameter configurations shown in Table 1. In the channel feature module, a 3-layer residual block is used, with a 1x3 kernel size and a block stride of 2. Similarly, the functional connectivity module also uses a 3-layer residual block, with a 3x3 kernel size and a block stride of 2. Furthermore, feature engineering utilizes 5 frequency bands in the original data, and the input convolution channels for both sampling modules are set to 5. A single KANLinner layer is used on the KAN classifier, with a function matrix size of 256xN, where N corresponds to the number of states in the dataset.

[0048] Table 1 Configuration Table of the Model The experimental results are as follows: Ablation experiment: To verify the rationale behind the proposed TBSL, its effects on different frequency bands are first presented. For example... Figure 1 As shown, under fatigue conditions, using only Alpha, Beta, and Gamma results demonstrates better performance compared to other frequency bands. And... Figure 2In the attention dataset, the contribution of fatigue to performance is more pronounced on the Delta dataset. However, it's noteworthy that the frequency bands of Beta and Gama show smaller variances on the attention dataset, exhibiting good stability and demonstrating their unique contribution to the attention task. Finally, using five frequency bands as five input feature map channels simultaneously resulted in significant improvements in both fatigue and attention performance. Figure 2 As can be seen, the median accuracy of multi-band on the fatigue dataset was improved to 84%, while in terms of attention, the distribution of the subject prediction effect of multi-band collaboration was more compact and stable.

[0049] Comparative experiment; The TBSL model proposed in this invention maintains a lightweight structure while achieving better prediction results compared to other models. Tables 2 and 3 show the comparison between the model of this invention and other models on two datasets. The specific models in Tables 2 and 3 include: FBCnet: Filter Bank Convolutional Network, designed to overcome the challenges of insufficient training samples and high-dimensional noise in EEG motor imagery decoding. Inspired by neurophysiological mechanisms, this network employs multi-view data representation and spatial filtering to extract discriminative spectral spatial features. Its core innovation lies in introducing a variance layer to effectively aggregate EEG temporal information, thereby achieving efficient learning with limited training data.

[0050] MshallowConvNet: A novel neural network architecture that improves upon the classic ShallowConvNet model. This model significantly improves the decoding performance of EEG signals by addressing the structural problems of the original model, such as the need for repeated parameter optimization on different datasets and unstable training.

[0051] InterpreCNN: A novel interpretable convolutional neural network for EEG-based driver fatigue recognition. This model employs a compact structure and spatially-temporally separable convolutions to process signals, achieving an average accuracy of 78.35% in cross-subject testing, significantly outperforming traditional methods and existing deep learning models. Its core value lies in integrating sample-level feature interpretation technology, which not only identifies key biologically significant features (such as alpha spindle waves characterizing fatigue) but also analyzes the reasons for misclassified samples, providing a new approach to understanding the model's decision-making process and improving recognition accuracy.

[0052] MsGPT: An end-to-end driver fatigue recognition model based on a multi-scale global cue Transformer. This model constructs an intra-scale-inter-scale cascaded framework through multi-scale convolutional block embeddings and introduces a global cue token to guide the interaction between global and local features. It effectively fuses multi-scale features using hybrid tokens and innovatively uses learnable queries to reduce computational complexity to linear.

[0053] CF-FCINet, the model of this invention, is a transparent brain state learning model designed to achieve high-precision decoding of brain states and deep interpretability of decision-making processes. It constructs a priori graph to guide feature extraction by capturing the spatiotemporal characteristics and connectivity patterns of neural oscillations from multiple dimensions. Multidimensional convolution is used to mine spatial topological relationships, and the Kolmogorov-Arnold representation theorem is used to predict brain state patterns, significantly improving feature representation capabilities. Simultaneously, the designed transparent EEG decoder reveals the decision-making reasoning process in fine detail, providing strong theoretical support for analyzing brain spatial activation patterns under different states. Experimental validation on fatigue and attention tasks demonstrates that CF-FCINet not only boasts superior performance but also exhibits exceptional transparency, providing a breakthrough technology for interpretability research in the field of brain state decoding.

[0054] FBCSP (Filter Bank Common Spatial Pattern) is a filter bank-based common spatial pattern algorithm that decomposes EEG signals into multiple frequency bands and extracts CSP features from each band. It then combines this with an automatic feature selection algorithm to filter the most discriminative band-feature combinations, effectively solving the dependency problem of traditional CSP algorithms on frequency band selection. Experiments on public datasets and stroke patient data show that it achieves better cross-validation accuracy than mainstream methods when using specific feature selection and classifier combinations.

[0055] EFMLNet: An EEG-near-infrared fusion network based on end-to-end mutual information learning. It mines the temporal information of EEG and the spatial information of fNIRS through their respective feature extractors, and fuses complementary information using dual parallel mutual learning modules. It achieves a cross-subject classification accuracy of 71.52% on the motor imagery task, outperforming most existing fusion methods.

[0056] DeepCNN: A deep convolutional neural network for end-to-end EEG decoding. By introducing advanced machine learning techniques such as batch normalization, exponential linear units, and pruning training strategies, it achieves an average accuracy of 84.0% in raw EEG signal classification tasks, comparable to the traditional FBCSP algorithm (82.1%). This model eliminates the need for manual feature extraction and can reveal its learned frequency band energy modulation features and spatial topological patterns through novel feature visualization techniques, providing a new paradigm for EEG-based brain function mapping.

[0057] 3D-CNN: An attention state detection model based on a three-dimensional convolutional neural network. It converts EEG signals into a three-dimensional representation to simultaneously preserve spatiotemporal features and employs a multi-scale feature extraction architecture including cascaded structures and parallel convolutional blocks. On a public dataset containing 26 participants, this model outperforms baseline methods in within-subject, between-subject, and adaptive classification scenarios, demonstrating its potential application in neurofeedback and ADHD treatment.

[0058] The comparison shows that the model of this invention maintains a high level of performance, especially on the fatigue dataset, where it is only about 1% worse than the recent MsGPT model, yet the model of this invention has a significantly smaller parameter size. On the attention dataset, this invention shows a greater improvement compared to other models. It outperforms EFMLNet by about 8% and also demonstrates a significant performance advantage compared to other models. Clearly, the model of this invention now achieves a good prediction result. Furthermore, the CF-FCINet of this invention significantly improves the model's lightweight nature. Moreover, the excellent recognition performance provides a practical basis for visualization and interpretability.

[0059] Table 2 Comparison of model performance on fatigue data Table 3 Comparison of model performance on attention data Table 4. Predicted accuracy of fatigue data for each subject Table 5. Predictions from each participant in the attention dataset. Within-subjects experiment; In EEG-related experiments, individual differences among participants can lead to significant deviations in results. Therefore, in-subject training and prediction were performed on each participant using two separate datasets. Tables 4 and 5 show the prediction accuracy for each participant under two cognitive tasks: fatigue and attention. In Table 4, most of the 23 participants achieved accuracy of around 70%–80%, with participants S12 and S04 achieving over 90%, but participants S05 and S17 showed lower accuracy. In Table 5, most of the 26 participants achieved accuracy above 70%, but there were also significant differences in accuracy between participants S07 and S04. In summary, analyzing the prediction results for each participant across the two datasets reveals that individual differences present varying degrees of resistance to EEG decoding. However, the observational results also demonstrate that the method proposed in this invention exhibits certain effectiveness in terms of participant adaptability and stability.

[0060] Explanatory visualization of channel activation; In this invention's model, a one-dimensional convolution kernel is used for the channel signal features. This approach ensures consistency of the feature maps in the original feature space, facilitating visualization and reconstruction. Figures 3 to 7The diagram shows the electrical signal waveforms of each EEG electrode channel, with the yellow sections representing the segments the model focuses on when predicting the current state; the yellower the color, the more the model focuses on it. To improve the visualization of activation, a threshold was set to filter out excessively low activation maps. The brain topography map below each diagram shows the activation maps at a specified time point, facilitating observation of the actual distribution of activated regions in the brain.

[0061] The brain fatigue activation mapping in this experiment is as follows: Figure 3 , Figure 4 and Figure 5 As shown, three different subjects under three different conditions were selected for analysis. Through the activation mapping relationship of the three states of wakefulness, fatigue, and unconsciousness, it was found that the activation mapping values ​​of FT8, O2, O1, and TP8 channels increased with increasing fatigue. Furthermore, observation of the corresponding brain topography maps revealed that the activation values ​​of channels in the occipital and parietal lobe regions were significantly increased in the fatigued state.

[0062] The results on the attention dataset are as follows: Figure 6 and Figure 7 The visualization shows channel activation in two attentional states: confirmation and uncertainty. Participant #25 in the confirmation state and participant #1 in the uncertainty state were selected for visualization. Observation revealed that channel activation in both states was predominantly concentrated in the right hemisphere of the brain. Specifically, there was increased attention paid to channels such as P8, Fp2, and AF4. Further observation... Figure 6 and Figure 7 The topographic map in the lower middle section shows that when attention decreases, the activation value of the right parietal lobe brain region increases significantly, indicating a correlation between the right parietal lobe and spatial attention.

[0063] Interpretive visualization of connectivity activation; To ensure interpretability of channel connectivity, the channel connectivity matrix is ​​also reconstructed from the activation mappings of the feature maps in the model back to the original empty space for visualization. The model activation values ​​are then reconstructed under both fatigue and attention task scenarios. Figures 8 to 12 The results shown demonstrate the connectivity between channels in each state from five perspectives of a 3D model of the brain.

[0064] In this experiment, fatigue was assessed using 17-channel EEG signals, resulting in low connectivity complexity. Figure 8 , Figure 9 and Figure 10 The activation mapping of channel connectivity under three fatigue states is visualized. Observing the connectivity mapping of these three states, it can be found that the model pays more attention to the connectivity between channels within the occipital lobe and between channels and the outside. Specifically, it can be observed that... Figure 8The occipital lobe region shown exhibits significantly lower activation intensity for external connectivity compared to the other two states. Meanwhile... Figure 9 and Figure 10 The study revealed that as fatigue increased, the occipital lobe showed a significant increase in its focus on external connectivity, primarily manifested in the increased activation values ​​of O1, O2, and Oz on connectivity near the left and right temporal lobes. In summary, fatigue is strongly correlated with regions in the occipital lobe. However, this dataset was collected during vehicle driving, and the experimental results are similar to studies on occipital lobe involvement in visual fatigue, echoing the previous interpretative analysis of channel activation.

[0065] Regarding brain attention, 28 EEG channels were selected to visualize channel connectivity activation mapping. The interpretative analysis of the connectivity between the two states of attention is as follows: Figure 11 and Figure 12 As shown. Observation Figure 11 and Figure 12 It was observed that connectivity activation during fatigue was primarily distributed along pathways in the left hemisphere. Furthermore, observation of the maps from both attentional states revealed higher activation intensity in the parietal and frontal lobes, particularly along pathways connecting to PF1, CZ, FC1, and CP1. Notably, significant differences were observed in the brain topography before and after activation mapping; the brain topography in this dataset was concentrated in the right frontal and right parietal lobes, while the connectivity in this experiment was concentrated in the left hemisphere. Further data from more clinical trials is needed to provide further insights.

[0066] In summary, the CF-FCINet and highly interpretable decoder of this invention provide a framework for decoding brain states and analyzing clinical data. This lightweight CF-FCINet demonstrates significant recognition performance across multiple brain state datasets. It employs convolutional kernels of different dimensions tailored to the characteristics of EEG data with varying expressions to extract features while preserving spatial consistency and reducibility of the feature maps. Furthermore, the model is coupled with a KAN classifier, which boasts high mathematical interpretability, further enhancing overall interpretability. Simultaneously, the transparent decoder utilizes activation mapping reconstruction combined with more intuitive visualization techniques in this field, providing highly interpretable evidence for the activation analysis of channel signal fragments and functional connectivity under brain fatigue states.

[0067] To verify the model's effectiveness, this invention selected two different cognitive tasks—fatigue and attention—and validated and visualized their effects on two publicly available datasets. Experimental results show that CF-FCINet not only accurately identifies brain states, but more importantly, through the visualization analysis of channel feature activation mapping and key functional connectivity selection using a transparent decoder, it intuitively reveals the dynamic changes in brain activity under different cognitive states. This multi-layered interpretable analysis provides a new research perspective and technical support for understanding the deeper states of the brain.

[0068] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for visualizing and identifying the state of the brain by interaction of the dynamic properties of the channels and the cortical network, characterized by, The method comprises the following steps: S1, obtaining an original EEG signal; S2, preprocessing and manually extracting features from the original EEG signal to obtain channel input features and channel connectivity input features; S3, constructing a transparent brain state learning model; S4, inputting the input features and the channel connectivity input features into the transparent brain state learning model to output a brain state visual recognition result.

2. The method of claim 1, wherein, In S2, the original EEG signal is preprocessed, specifically: The original EEG signal is divided into different frequency band signals by band-pass filtering: In the formula, represents the signal value after filtering of the cth channel, wherein, is the signal sampling point position, T represents the number of sampling points; represents the output signal after the band-pass filter processing calculated at the n-k moment; represents the normalization coefficient; M and N respectively represent and the order of k represents the time offset; represents the original signal collected at the n-k moment; and respectively represent the coefficients corresponding to the lower limit cutoff frequency and the upper limit cutoff frequency.

3. The method of claim 2, wherein the method further comprises: In S2, the manual feature extraction includes: The differential entropy is used as the signal feature of the frequency band signal after band-pass filtering, and the channel input feature is obtained, wherein the differential entropy is specifically represented as: wherein represents a value of the calculated differential entropy; represents a slice in which the differential entropy is calculated in the filtered signal; represents a Gaussian probability density function; represents a variance of the slice; represents a natural base number; Based on the channel input feature, the channel connectivity input feature is calculated: In the formula, represents the connectivity value between the i and j channels; , respectively represent the signal value of the i channel and the signal value of the j channel; and respectively represent the signal average value of the i channel and the signal average value of the j channel; represents the maximum number of channels of the electroencephalogram.

4. The method of claim 3, wherein, In S3, the transparent brain state learning model includes a channel feature and functional connectivity integrated network and a transparent electroencephalogram decoder. The channel feature and functional connectivity integrated network includes a plurality of residual blocks with different dimension convolution kernels and a KAN network; the output of the residual block is used as the input of the KAN network.

5. The method of claim 4, wherein, S4 specifically includes the following steps: S41, inputting the channel input feature and the channel connectivity input feature into the residual block with different dimension convolution kernels for feature sampling to obtain a channel feature map and a connectivity feature map respectively; S42, performing global average pooling on the channel feature map and the connectivity feature map to obtain final channel convolution features and final connectivity features, and performing splicing processing on the final channel convolution features and the final connectivity features to obtain fusion features; S43, inputting the fusion features into the KAN network for brain state pattern prediction, and then outputting a brain state recognition result; S44, inputting the channel feature map and the connectivity feature map output from the last residual block and the brain state recognition result into the transparent electroencephalogram decoder to calculate the sensitivity of the feature map; S45, calculating an activation mapping based on the sensitivity of the feature map and the activation pattern, fusing the activation mapping with the spatial information of the original EEG signal to obtain an explanatory image of brain state visualization.

6. The method of claim 5, wherein the method further comprises: S41 specifically includes: Inputting the channel input feature into the residual block with one-dimensional convolution kernel for sampling, which is represented as: In the formula, represents the first l channel feature map of the layer convolution output; represents the kernel size of the convolution; is a weight of a one-dimensional convolution kernel; represents the channel input feature; represents a position subscript in the convolution weight matrix; Inputting the channel connectivity input feature into the residual block with two-dimensional convolution kernel for sampling, which is represented as: In the formula, is the first l layer convolution output connectivity feature map; is the weight of the two-dimensional convolution kernel; represents the channel connectivity input feature; , respectively represent the position coordinates of the corresponding convolution kernel weight.

7. The method of claim 6, wherein the method further comprises: In S42, the channel feature map and the connectivity feature map are respectively subjected to global average pooling to obtain the final channel convolution features and the connectivity features, which are represented as: wherein, and denote the final channel convolutional feature and the final connectivity feature, respectively; , denotes taking the maximum value in the corresponding channel feature map; denotes the feature value at the (i, j) position in the specified c-th channel feature map in the channel feature; denotes the feature value at the (i, j) position in the specified c-th channel feature map in the connectivity feature.

8. The method of claim 5, wherein the method further comprises: In S43, the fusion features are input into the KAN network for brain state pattern prediction, and then a brain state recognition result is output, which is represented as: In the formula, represents an output weight value; is a function; represents a fusion feature; represents a function matrix corresponding to the first layer, represents a function matrix corresponding to the first layer, represents a function matrix corresponding to the first layer, represents a function matrix corresponding to the first layer, represents a function matrix corresponding to the first layer, The output weight is normalized by using a softmax function and a brain state recognition result is output, which is expressed as: In the formula, is a brain state recognition result, The value of each class is a score value; denotes normalization.

9. The method of claim 5, wherein the method further comprises: In S44, the sensitivity of the feature map is calculated, which is represented as: Wherein, wherein, denotes the sensitivity of the feature map; , denote the height and width of the map, respectively; denotes the weighted gradient; denotes the predicted target class; denotes the feature value of the feature map of the last layer; , denote the horizontal and vertical position coordinate points of the last layer feature map, respectively; denotes the feature value of the feature map at the (a, b) position on the c-th channel in the last layer.

10. The method of claim 9, wherein the method further comprises: S45 specifically includes: calculating an activation mapping based on the sensitivity of the feature map and the activation pattern, which is represented as: In the formula, denotes an activation mapping; denotes an activation function.

Images