A Classification Method for Consciousness Disorders Based on Dynamic Graph Convolution and Channel Attention Mechanisms in EEG Signals

Abstract

This invention discloses a method for classifying consciousness disorders in EEG signals based on dynamic graph convolution and channel attention mechanisms, relating to the field of EEG signal recognition technology. According to the method provided in the embodiments of this invention, a complete closed loop is achieved, covering uploading, preprocessing, artifact removal, segmentation, feature extraction, and model prediction. A dynamic graph convolution modeling method with a trainable adjacency matrix is ​​used to adaptively learn functional connections between EEG channels, overcoming the poor generalization problem of static adjacency matrices. Simultaneously, a joint modeling framework of explicit connectivity (PLV) + implicit connectivity (dynamic graph convolution) is used to more robustly capture cross-channel synchronization patterns under low signal-to-noise ratio conditions.

Description

Technical Field

[0001] This invention discloses a method for classifying consciousness disorders based on dynamic graph convolution and channel attention mechanisms of EEG signals, which relates to the field of EEG signal recognition technology. Background Technology

[0002] Currently, the clinical classification of patients with impaired consciousness (including minimally conscious state (MCS) and vegetative state (VS)) is still mainly based on behavioral scales (such as CRS-R), but there are significant shortcomings in bedside, rapid, objective, and repeatable brain function assessment. In conjunction with the web-based EEG processing and prediction system proposed in this application, the existing technology has at least the following technical problems: Highly subjective and reliant on human experience: Behavioral scales require observation of the patient's overt responses, and are greatly affected by the operator's experience, patient fatigue / sedation / hearing, etc., which can easily lead to misdiagnosis or missed diagnosis.

[0003] Low diagnostic efficiency and difficulty in real-time bedside assessment: Scale assessment is time-consuming and difficult to monitor continuously, and cannot be frequently reassessed in intensive care settings.

[0004] EEG signals have a low signal-to-noise ratio and complex artifacts: electromyography, electrooculography, power line interference, etc. can overwhelm weak features related to consciousness; without automated quality control, the model input is unstable.

[0005] Functional connectivity between multiple channels exhibits individual differences and state non-stationarity: changes in conscious state often manifest as changes in cross-channel synchronicity / connection patterns; traditional CNNs treat channels as regular grids, making it difficult to represent the real brain network structure.

[0006] Existing graph convolution methods mostly rely on static, manually constructed adjacency matrices: fixed edges based on geometric distance or experience, which cannot adaptively adjust to patient differences, channel loss after artifact removal, and different sampling rates / segmentation strategies, resulting in limited generalization.

[0007] Attention mechanisms face difficulties in implementation in EEG scenarios: general attention modules have a large number of parameters and high computational overhead, and are prone to amplifying noise channels under low signal-to-noise ratio input, making it difficult to meet the requirements of rapid bedside inference and interpretability.

[0008] Lack of end-to-end engineering pipeline: Most research remains at offline training / inference, and has not formed a reproducible system from file upload, preprocessing, segmentation, feature extraction, parallel acceleration to model prediction and result aggregation, resulting in high deployment barriers. Summary of the Invention

[0009] This invention addresses the problems of existing technologies by providing a method for classifying consciousness disorders based on dynamic graph convolution and channel attention mechanisms in electroencephalogram signals. The technical solution adopted is as follows: Firstly, a method for classifying consciousness disorders based on dynamic graph convolution and channel attention mechanisms of EEG signals includes: S1, retrieve the EEG.set file as raw data for caching, and automatically identify and remove EEG channels; S2, based on the original data, preprocess the EEG signal through a specified time window, and obtain an EEG data sequence with a uniform sampling rate through downsampling processing; S3, based on the EEG data sequence, the EEG signal is divided into multiple time segments according to a preset window length and step size; amplitude threshold detection is performed on each time segment, and abnormal segments exceeding the preset amplitude range are removed to obtain continuous EEG samples and save them as .mat files; S4. Based on each .mat file, a feature matrix is ​​formed by segmenting and extracting features through frequency domain features, complexity features, and connectivity features. S5, input the feature matrix into a brain network containing dynamic graph convolution and ECA (Efficient Channel Attention) mechanism, and output the classification result of each segment of the .mat file; S6, perform voting and average probability summarization on the classification results of each segment, and output the consciousness state judgment and confidence level.

[0010] In some implementations, S4 includes: S41, based on each .mat file, obtain the band power, spectral entropy, spectral boundary frequency and 1 / f slope exponent of multiple frequency bands through Welch power spectrum, and obtain spectral structure difference data as frequency domain features; S42, based on each .mat file, set the window length and step size to form a sliding sub-window to obtain sample entropy, permutation entropy and Lempel-Ziv complexity as nonlinear complexity features; S43. The instantaneous phase is obtained according to the Hilbert transform. The phase difference between each pair of EEG channels is obtained by vectorized broadcasting and the average value is calculated to obtain the PLV matrix, which is used as the connectivity feature.

[0011] In some implementations, S4 further includes: S44, Based on the frequency domain characteristics, nonlinear characteristics and connectivity characteristics of the EEG channel in each segment, a feature matrix is ​​obtained through feature combination; S45, Align the feature dimension of the feature matrix with the preset feature dimension, and perform column-by-column min-max normalization on the feature matrix.

[0012] In some implementations, S5 includes: S51, Load network weights according to the pre-trained model parameters, and initialize the brain network structure according to the preset adjacency matrix; S52, the adjacency matrix is ​​set as a dynamic graph convolution modeling mode with trainable parameters, and functionally connected to the EEG channel through adaptive learning.

[0013] Secondly, embodiments of the present invention provide a classification system for consciousness disorders based on dynamic graph convolution and channel attention mechanisms of electroencephalogram signals, including: The data access caching module is used to obtain the EEG.set file as raw data for caching, and to automatically identify and remove EEG channels; The window preview module is used to preprocess the EEG signal according to the original data through a specified time window, and to obtain an EEG data sequence with a uniform sampling rate through downsampling processing; The artifact removal module is used to divide the EEG signal into multiple time segments according to the EEG data sequence and a preset window length and step size; to perform amplitude threshold detection on each time segment, and to remove abnormal segments that exceed the preset amplitude range, so as to obtain continuous EEG samples and save them as .mat files. The matrix generation module is used to extract feature matrices from each of the .mat files by segmenting them using frequency domain features, complexity features, and connectivity features. The classification processing module is used to input the feature matrix into the classification model and output the classification result for each segment of the .mat file; the classification model includes dynamic graph convolution and ECA channel attention mechanism. The classification prediction module is used to vote on and summarize the average probability of the classification results for each segment, and output the consciousness state judgment and confidence level.

[0014] In some implementations, the matrix generation module includes: The frequency domain feature unit is used to obtain the band power, spectral entropy, spectral boundary frequency and 1 / f slope exponent of multiple frequency bands based on each .mat file through Welch power spectrum, and obtain spectral structure difference data as frequency domain features; The complexity feature unit is used to form a sliding sub-window based on the window length and step size of each .mat file to obtain sample entropy, permutation entropy and Lempel-Ziv complexity as nonlinear complexity features; The connectivity feature unit is used to obtain the instantaneous phase according to the Hilbert transform, obtain the pairwise phase difference of the EEG channels through vectorized broadcasting and calculate the average value to obtain the PLV matrix, which serves as the connectivity feature.

[0015] In some implementations, the matrix generation module further includes: The matrix combination unit is used to obtain a feature matrix by combining features based on the frequency domain characteristics, nonlinear characteristics and connectivity characteristics of the EEG channel in each segment. The input processing unit is used to align the feature dimensions of the feature matrix with preset feature dimensions and perform column-by-column min-max normalization on the feature matrix.

[0016] In some implementations, the classification processing module includes: The model initialization unit is used to load network weights according to the pre-trained model parameters and initialize the brain network structure according to the preset adjacency matrix. The model loading unit is used to set the adjacency matrix as a dynamic graph convolution modeling method with trainable parameters, and to make functional connections with the EEG channel through adaptive learning.

[0017] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein when the one or more computer instructions are executed by the processor, they implement the method described in the first aspect above.

[0018] Fourthly, embodiments of the present invention provide a computer storage medium, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, it implements the method described in the first aspect.

[0019] One or more embodiments of the present invention can bring at least the following beneficial effects: According to the method of this invention, a complete closed loop covering uploading, preprocessing, artifact removal, segmentation, feature extraction, and model prediction can be achieved. The adjacency matrix is ​​set as a dynamic graph convolution modeling method with trainable parameters, which is used to adaptively learn the functional connections between EEG channels and overcome the problem of poor generalization of static adjacency matrices. At the same time, the joint modeling framework of explicit connectivity (PLV) + implicit connectivity (dynamic graph convolution) is used to capture cross-channel synchronization modes more robustly under low signal-to-noise ratio.

[0020] This invention incorporates ECA channel attention to suppress noisy channels and enhance key channel features, meeting the computational power and real-time requirements of clinical deployment. By using feature length alignment in the prediction phase and the same per-sample normalization strategy as in training, the deployment input distribution is consistent with the training distribution, thereby improving reproducibility and stability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for classifying consciousness disorders of EEG signals based on dynamic graph convolution and channel attention mechanism provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a classification model combining dynamic graph convolution and attention modules provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] Example 1: Figure 1 A flowchart of a method for classifying consciousness disorders based on dynamic graph convolution and channel attention mechanisms of EEG signals is shown, such as... Figure 1 As shown, the EEG signal consciousness disorder classification method based on dynamic graph convolution and channel attention mechanism provided in this embodiment includes: S1, retrieve the EEG.set file as raw data for caching, and automatically identify and remove EEG channels; S2, based on the original data, preprocess the EEG signal through a specified time window, and obtain an EEG data sequence with a uniform sampling rate through downsampling processing; S3, based on the EEG data sequence, the EEG signal is divided into multiple time segments according to a preset window length and step size; amplitude threshold detection is performed on each time segment, and abnormal segments exceeding the preset amplitude range are removed to obtain continuous EEG samples and save them as .mat files; S4. Based on each .mat file, a feature matrix is ​​formed by segmenting and extracting features through frequency domain features, complexity features, and connectivity features. S5, input the feature matrix into the classification model, and output the classification result of each segment of the .mat file; the classification model includes dynamic graph convolution and ECA channel attention mechanism; S6, perform voting and average probability summarization on the classification results of each segment, and output the consciousness state judgment and confidence level.

[0025] Furthermore, S4, based on each .mat file, a feature matrix is ​​formed by segmenting and extracting features through frequency domain features, complexity features, and connectivity features, including: S41, based on each .mat file, obtain the band power, spectral entropy, 95% spectral boundary frequency and 1 / f slope exponent of multiple frequency bands through Welch power spectrum, and obtain spectral structure difference data as frequency domain features; S42, based on each .mat file, set the window length and step size to form a sliding sub-window to obtain sample entropy, permutation entropy and Lempel-Ziv complexity as nonlinear complexity features; S43. The instantaneous phase is obtained according to the Hilbert transform. The phase difference between each pair of EEG channels is obtained by vectorized broadcasting and the average value is calculated to obtain the PLV matrix, which is used as the connectivity feature.

[0026] According to S42, Welch power spectrum estimation is used to calculate the band power of multiple frequency bands (default δ / θ / α / β / γ), and further calculates the spectral entropy, 95% spectral boundary frequency (SEF95) and 1 / f slope exponent to characterize the differences in consciousness-related spectral structure.

[0027] Within each segment, a sliding sub-window is formed by pressing winLen / stepSize to calculate the sample entropy, permutation entropy, and Lempel-Ziv complexity, which are used to capture relatively robust complexity changes under low signal-to-noise ratio conditions. According to S43, the instantaneous phase is obtained based on the Hilbert transform. The phase difference between each pair of channels is calculated using vectorized broadcasting, and the average value is obtained to obtain the PLV matrix (20×20) to reflect cross-channel synchronization / functional connection.

[0028] Furthermore, S4 also includes: S44. Based on the frequency domain characteristics, nonlinear characteristics, and connectivity characteristics of the EEG channel in each segment, a 20×F feature matrix is ​​obtained by combination. S45, Align the feature dimension of the feature matrix with the preset feature dimension, and perform column-by-column min-max normalization on the feature matrix.

[0029] Further, in step S5, the classification model is input based on the feature matrix, and the classification result for each segment of the .mat file is output; the classification model includes dynamic graph convolution and ECA channel attention mechanism; including: S51, Load network weights according to the pre-trained model parameters, and initialize the brain network structure according to the preset adjacency matrix; S52, the adjacency matrix is ​​set as a dynamic graph convolution modeling mode with trainable parameters, and functionally connected to the EEG channel through adaptive learning.

[0030] According to S51 and 62, the adjacency matrix A is set as a trainable parameter and updated in reverse by the classification loss in end-to-end training, so that the channel connection weights can be adaptively adjusted according to the differences in patients and the differences in consciousness, rather than relying on fixed geometric distances or empirical edges.

[0031] Non-EEG channels are automatically removed during the upload phase, and a fixed 20-channel structure is adopted at the feature end; the A can be trained to learn a robust connection pattern for key channels, reducing the performance degradation caused by channel quality fluctuations.

[0032] S5 also includes performing convolutional feature extraction on the input EEG signal in the time dimension. Local feature modeling of the time series of the EEG signal is performed through one-dimensional convolution operations to capture the dynamic changes of the EEG signal at different time scales and output a time-domain feature representation. A graph structure is constructed based on the connectivity relationships between EEG channels, and the features of each channel node are propagated and updated through a graph convolutional network to learn the functional connectivity relationships between different EEG channels, thereby extracting cross-channel spatial feature representations.

[0033] Both the time-domain feature extraction unit and the channel-domain feature extraction unit incorporate the ECA channel attention mechanism. By assigning weights to the features of each channel, the feature enhancement of important EEG channels and the suppression of noise channels are achieved, thereby improving feature expression ability and classification performance.

[0034] According to the method provided in the embodiments of the present invention, a complete closed loop covering uploading, preprocessing, artifact removal, segmentation, feature extraction and model prediction is realized; the adjacency matrix is ​​set as a dynamic graph convolution modeling method with trainable parameters, which is used to adaptively learn the functional connections between EEG channels and overcome the problem of poor generalization of static adjacency matrices; at the same time, the joint modeling framework of explicit connectivity (PLV) + implicit connectivity (dynamic graph convolution) is used to capture cross-channel synchronization modes more robustly under low signal-to-noise ratio.

[0035] The method in this invention incorporates ECA channel attention to suppress noisy channels and enhance key channel features, thereby meeting the computational power and real-time requirements of clinical deployment. The feature length alignment during the prediction phase uses the same per-sample normalization strategy as the training phase, ensuring consistency between the deployment input and training distribution, thus improving reproducibility and stability.

[0036] Example 2: Based on Example 1, the adjacency matrix update can introduce sparsification / regularization constraints (e.g., L1, graph Laplacian constraints) or use attention graph learning instead; In S43, the connectivity feature can be replaced or supplemented with coherence, PLI / wPLI, mutual information, etc. In S3, artifact removal can be replaced by automatic artifact removal methods such as ICA / ASR, or a more refined bad segment detection strategy can be adopted. Temporal feature extraction can be replaced by multi-scale one-dimensional convolution, TCN, or Transformer encoder; The deployment model can be expanded from a web-based interface to a bedside all-in-one machine / edge computing device.

[0037] Example 3: Thirdly, embodiments of the present invention provide a classification system for consciousness disorders based on dynamic graph convolution and channel attention mechanisms of electroencephalogram signals, including: The data access caching module is used to obtain the EEG.set file as raw data for caching, and to automatically identify and remove EEG channels; The window preview module is used to preprocess the EEG signal according to the original data through a specified time window, and to obtain an EEG data sequence with a uniform sampling rate through downsampling processing; The artifact removal module is used to divide the EEG signal into multiple time segments according to the EEG data sequence and a preset window length and step size; to perform amplitude threshold detection on each time segment, and to remove abnormal segments that exceed the preset amplitude range, so as to obtain continuous EEG samples and save them as .mat files. The matrix generation module is used to extract feature matrices from each of the .mat files by segmenting them using frequency domain features, complexity features, and connectivity features. The classification processing module is used to input the feature matrix into the classification model and output the classification result for each segment of the .mat file; the classification model includes dynamic graph convolution and ECA channel attention mechanism. The classification prediction module is used to vote on and summarize the average probability of the classification results for each segment, and output the consciousness state judgment and confidence level.

[0038] Furthermore, the matrix generation module includes: The frequency domain feature unit is used to obtain the band power, spectral entropy, 95% spectral boundary frequency and 1 / f slope exponent of multiple frequency bands based on each .mat file through Welch power spectrum, and obtain spectral structure difference data as frequency domain features; The complexity feature unit is used to form a sliding sub-window based on the window length and step size of each .mat file to obtain sample entropy, permutation entropy and Lempel-Ziv complexity as nonlinear complexity features; The connectivity feature unit is used to obtain the instantaneous phase according to the Hilbert transform, obtain the pairwise phase difference of the EEG channels through vectorized broadcasting and calculate the average value to obtain the PLV matrix, which serves as the connectivity feature.

[0039] Furthermore, the matrix generation module also includes: The matrix combination unit is used to obtain a 20×F feature matrix by combining the frequency domain characteristics, nonlinear characteristics and connectivity characteristics of the EEG channel in each segment. The input processing unit is used to align the feature dimensions of the feature matrix with preset feature dimensions and perform column-by-column min-max normalization on the feature matrix.

[0040] Furthermore, the classification processing module includes: The model initialization unit is used to load network weights according to the pre-trained model parameters and initialize the brain network structure according to the preset adjacency matrix. The model loading unit is used to set the adjacency matrix as a dynamic graph convolution modeling mode with trainable parameters, and to make functional connections with the EEG channel through adaptive learning; The Time Domain Feature Extraction (TDFE) module is used to extract convolutional features from the input EEG signal in the time dimension. It performs local feature modeling on the time series of EEG signals through one-dimensional convolution operations to capture the dynamic changes of EEG signals at different time scales and outputs a time domain feature representation.

[0041] The Channel Domain Feature Extraction module (CDFE) is used to construct a graph structure based on the connectivity between EEG channels, and to propagate and update the features of each channel node through a graph convolutional network to learn the functional connectivity between different EEG channels, thereby extracting cross-channel spatial feature representations.

[0042] Both the time-domain feature extraction unit and the channel-domain feature extraction unit incorporate the ECA channel attention mechanism. By assigning weights to the features of each channel, the feature enhancement of important EEG channels and the suppression of noise channels are achieved, thereby improving feature expression ability and classification performance.

[0043] Example 4: This is a schematic diagram of a model structure for a classification system of EEG signals for consciousness disorders based on dynamic graph convolution and channel attention mechanisms, provided in an embodiment of the present invention. The model mainly includes an input module, a feature extraction module, a dynamic graph convolution update module, and a classification output module. Its structure is described below: The input is EEG signal data, represented as a three-dimensional tensor (1, C, T), where C represents the number of EEG channels and T represents the number of time sampling points. The horizontal axis represents the time dimension and the vertical axis represents the channel dimension. (2) Spatiotemporal feature extraction part The input signal sequentially enters the spatiotemporal feature extraction network, which includes multiple feature extraction units. These units include: a TDFE module (Time Domain Feature Extraction Module), used to extract local temporal features of the EEG signal through convolution operations; and a CDFE module (Channel Domain Feature Extraction Module), used to model the spatial relationships between different EEG channels. In this embodiment, the TDFE and CDFE modules are alternately connected to form a composite structure, and this configuration is repeated multiple times to extract deep spatiotemporal features layer by layer. Dynamic graph convolution update module, such as Figure 2 As shown on the right, a multi-layer graph convolutional network (GCN) is introduced during the feature extraction process. In each graph convolutional layer, the adjacency matrix is ​​adaptively updated based on the output features of the current layer, realizing the dynamic adjustment of the functional connectivity relationships of brain regions, thereby constructing a brain network structure that changes over time.

[0044] Example 5: This embodiment also provides an electronic device, including a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method of Embodiment 1; In practical applications, the processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller unit (MCU), microprocessor, or other electronic components to execute the methods described in the above embodiments.

[0045] The method implemented in this embodiment is as described in Embodiment 1.

[0046] Example 6: This embodiment also provides a computer storage medium, in which a computer program is stored, and when the computer program is executed by one or more processors, it implements the method of embodiment one. The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0047] The method implemented in this embodiment is as described in Embodiment 1.

[0048] In the several embodiments provided in this invention, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system and method embodiments described above are merely illustrative.

[0049] It should be noted that, in this document, the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0050] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and variations in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection for this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for classifying consciousness disorders in EEG signals based on dynamic graph convolution and channel attention mechanisms, characterized in that, include: S1, retrieve the EEG.set file as raw data for caching, and automatically identify and remove useless EEG channels; S2, based on the original data, preprocess the EEG signal through a specified time window, and obtain an EEG data sequence with a uniform sampling rate through downsampling processing; S3, based on the EEG data sequence, the EEG signal is divided into multiple time segments according to a preset window length and step size; amplitude threshold detection is performed on each time segment, and abnormal segments exceeding the preset amplitude range are removed to obtain continuous EEG samples and save them as .mat files; S4. Based on each .mat file, a feature matrix is ​​formed by segmenting and extracting features through frequency domain features, complexity features, and connectivity features. S5, input the feature matrix into the classification model, and output the classification result of each segment of the .mat file; the classification model includes dynamic graph convolution and ECA channel attention mechanism; S6, perform voting and average probability summarization on the classification results of each segment, and output the consciousness state judgment and confidence level.

2. The method for classifying consciousness disorders by EEG signals according to claim 1, characterized in that, S4 include: S41, based on each .mat file, obtain the band power, spectral entropy, spectral boundary frequency and 1 / f slope exponent of multiple frequency bands through Welch power spectrum, and obtain spectral structure difference data as frequency domain features; S42, based on each .mat file, set the window length and step size to form a sliding sub-window to obtain sample entropy, permutation entropy and Lempel-Ziv complexity as nonlinear complexity features; S43. The instantaneous phase is obtained according to the Hilbert transform. The phase difference between each pair of EEG channels is obtained by vectorized broadcasting and the average value is calculated to obtain the PLV matrix, which is used as the connectivity feature.

3. The method for classifying consciousness disorders by EEG signals according to claim 2, characterized in that, S4 further includes: S44, Based on the frequency domain characteristics, nonlinear characteristics and connectivity characteristics of the EEG channel in each segment, a feature matrix is ​​obtained through feature combination; S45, Align the feature dimension of the feature matrix with the preset feature dimension, and perform column-by-column min-max normalization on the feature matrix.

4. The method for classifying consciousness disorders by EEG signals according to claim 3, characterized in that, S5 include: S51, Load network weights according to the pre-trained model parameters, and initialize the brain network structure according to the preset adjacency matrix; S52, the adjacency matrix is ​​set as a dynamic graph convolution modeling mode with trainable parameters, and functionally connected to the EEG channel through adaptive learning.

5. A classification system for consciousness disorders based on dynamic graph convolution and channel attention mechanisms of electroencephalogram signals, characterized in that, include: The data access caching module is used to obtain the EEG.set file as raw data for caching, and to automatically identify and remove useless EEG channels; The window preview module is used to preprocess the EEG signal according to the original data through a specified time window, and to obtain an EEG data sequence with a uniform sampling rate through downsampling processing; The artifact removal module is used to divide the EEG signal into multiple time segments according to the EEG data sequence and a preset window length and step size; to perform amplitude threshold detection on each time segment, and to remove abnormal segments that exceed the preset amplitude range, so as to obtain continuous EEG samples and save them as .mat files. The matrix generation module is used to extract feature matrices from each of the .mat files by segmenting them using frequency domain features, complexity features, and connectivity features. The classification processing module is used to input the feature matrix into the classification model and output the classification result for each segment of the .mat file; the classification model includes dynamic graph convolution and ECA channel attention mechanism. The classification prediction module is used to vote on and summarize the average probability of the classification results for each segment, and output the consciousness state judgment and confidence level.

6. The EEG signal consciousness disorder classification system according to claim 5, characterized in that, The matrix generation module includes: The frequency domain feature unit is used to obtain the band power, spectral entropy, spectral boundary frequency and 1 / f slope exponent of multiple frequency bands based on each .mat file through Welch power spectrum, and obtain spectral structure difference data as frequency domain features; The complexity feature unit is used to form a sliding sub-window based on the window length and step size of each .mat file to obtain sample entropy, permutation entropy and Lempel-Ziv complexity as nonlinear complexity features; The connectivity feature unit is used to obtain the instantaneous phase according to the Hilbert transform, obtain the pairwise phase difference of the EEG channels through vectorized broadcasting and calculate the average value to obtain the PLV matrix, which serves as the connectivity feature.

7. The EEG signal consciousness disorder classification system according to claim 6, characterized in that, The matrix generation module also includes: The matrix combination unit is used to obtain a feature matrix by combining features based on the frequency domain characteristics, nonlinear characteristics and connectivity characteristics of the EEG channel in each segment. The input processing unit is used to align the feature dimensions of the feature matrix with preset feature dimensions and perform column-by-column min-max normalization on the feature matrix.

8. The EEG signal consciousness disorder classification system according to claim 7, characterized in that, The classification processing module includes: The model initialization unit is used to load network weights according to the pre-trained model parameters and initialize the brain network structure according to the preset adjacency matrix. The model loading unit is used to set the adjacency matrix as a dynamic graph convolution modeling method with trainable parameters, and to make functional connections with the EEG channel through adaptive learning.

9. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the brainwave signal consciousness disorder classification method as described in any one of claims 1-4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the brainwave signal consciousness disorder classification method as described in any one of claims 1-4.

Images