Dynamic graph neural network-based disturbance of consciousness classification method and related equipment

Abstract

The embodiment of the invention provides a disturbance of consciousness classification method based on a dynamic graph neural network and related equipment, and belongs to the technical field of crossing of medical artificial intelligence and neural engineering. The method comprises the steps that resting-state electroencephalogram signals of a patient are collected and preprocessed; calculating a phase locking value through a sliding window, and constructing a dynamic brain function network diagram sequence; the dynamic graph sequence is input into a dynamic graph neural network model for processing, the model is composed of a plurality of dynamic graph convolution units which are connected in a time sequence mode, and each unit extracts space-time dynamic characteristics of the brain network in a combined mode through a coupled graph convolution network and a gating circulation unit; and based on the final hidden state output by the model, obtaining a classification result of the micro-consciousness state or the plant state through a classifier. According to the method, the problems that an existing static analysis method cannot capture brain network dynamic characteristics and a model and a data structure are mismatched are solved, higher-precision automatic classification of disturbance of consciousness is achieved, and the method has good clinical auxiliary diagnosis value.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of medical artificial intelligence and neural engineering, and in particular to a method and related equipment for classifying consciousness disorders based on dynamic graph neural networks. Background Technology

[0002] Disorders of Consciousness (DoC) are a common clinical state following severe brain injury. Accurate differentiation between minimally conscious state (MCS) and vegetative state (VS) is crucial for treatment decisions and prognostic assessment. Currently, clinical practice mainly relies on the Coma Recovery Scale-Revised (CRS-R) for behavioral assessment; however, this method has drawbacks such as high subjectivity, a high misdiagnosis rate, and dependence on the patient's motor response ability.

[0003] In recent years, objective assessment methods based on electroencephalography (EEG) have gradually emerged. Traditional methods mostly employ machine learning, relying on manual feature extraction, which has limited generalization ability. Deep learning techniques such as convolutional neural networks (CNNs) have been introduced, but CNNs process grid-like data (such as images), while brain functional networks are essentially non-Euclidean graph structures. CNNs need to rearrange the functional connectivity matrix to fit their structure, a process that may introduce bias and disrupt the network's topology.

[0004] Furthermore, existing methods typically analyze the entire EEG signal as a static functional connectivity network, neglecting the dynamic changes in brain functional connectivity over time. Research indicates that the maintenance of consciousness is closely related to the dynamic reorganization of brain networks. Therefore, static analysis methods struggle to capture the temporal dynamics of conscious states, resulting in limited classification performance. Summary of the Invention

[0005] The main objective of this application is to propose a method, electronic device, storage medium, and program product for classifying consciousness disorders based on dynamic graph neural networks, in order to solve the problems of static analysis losing dynamic information and mismatch between traditional models and non-EEG data structures, thereby achieving more accurate automatic classification of consciousness disorders.

[0006] To achieve the above objectives, one aspect of this application proposes a method for classifying consciousness disorders based on a dynamic graph neural network, the method comprising: Obtaining resting-state electroencephalogram (EEG) signals from patients with disorders of consciousness; The electroencephalogram (EEG) signals are preprocessed to obtain multi-channel EEG signals; Based on sliding windows and phase-locking values, the multi-channel EEG signals are converted into a dynamic brain function network sequence, wherein each time window corresponds to a brain function connectivity graph, and the brain function connectivity graph uses the EEG acquisition channels as nodes and the phase-locking values ​​between channels as edge weights. The dynamic brain function network sequence is input into a dynamic graph neural network model for processing. The dynamic graph neural network model consists of multiple temporally connected dynamic graph convolutional units. Each dynamic graph convolutional unit is used to extract spatiotemporal features based on the brain function connectivity graph of the current time window and the hidden state of the previous time window through coupled graph convolutional networks and gated recurrent units, and outputs the hidden state of the current time window. Based on the final hidden state output by the dynamic graph neural network model, a classifier is used to obtain a classification result of consciousness disorder, which includes a minimally conscious state and a vegetative state.

[0007] In some embodiments, the conversion of multichannel EEG signals into dynamic brain functional network sequences includes: For the signal within each time window, calculate the phase lock value between all pairs of channels to obtain the phase lock value matrix; Connections in the phase-locked value matrix that are greater than a preset threshold are considered valid connections, and a brain functional connectivity map for that time window is constructed, wherein the node feature is the weighted degree centrality of the node.

[0008] In some embodiments, the internal operations of the dynamic graph convolution unit include: The gate signals are updated based on the current brain functional connectivity graph using the first graph convolutional network; The reset gate signal is calculated based on the current brain functional connectivity graph using the second graph convolutional network; The candidate hidden state is calculated by using the third graph convolutional network based on the current brain functional connectivity map and in combination with the reset gate signal and the previous hidden state. Based on the update gate signal, the previous hidden state, and the candidate hidden state, calculate and output the current hidden state.

[0009] In some embodiments, the first graph convolutional network, the second graph convolutional network, and the third graph convolutional network each include at least two graph convolutional layers, wherein the propagation rule of the at least one graph convolutional layer is:

[0010] in, To add self-connected adjacency matrices, for The degree matrix, For the first Layer node features For trainable weight matrix, This is the activation function.

[0011] In some embodiments, the Dynamic Graph Convolutional Unit (DGCN Cell) includes: At least one spatial feature extraction subunit, consisting of consecutive graph convolutional network layers, is used to extract spatial features of nodes based on graph topology. The temporal feature fusion subunit, based on a gated loop unit structure, is used to fuse the current spatial features extracted by the spatial feature extraction subunit with the hidden state of the previous time step, and output the hidden state of the current time step. The calculation of the update gate, reset gate, and candidate hidden state in the gated loop unit is performed by the independent spatial feature extraction subunit based on the currently input graph data.

[0012] In some embodiments, the calculation process of the time feature fusion subunit is defined by the following formula: Update Gate :

[0013] Reset door :

[0014] Candidate hidden state :

[0015] Current hidden state :

[0016] in, The graph data is for the current time step. The hidden state of the previous time step. , , These are three independent graph convolutional network subunits with identical structures. For activation functions; For Hadamah accumulation; To update the trainable weight matrix of the gate, To update the door offset, To reset the trainable weight matrix of the gate, To reset the door offset, is the trainable weight matrix for the candidate hidden states.

[0017] In some embodiments, the classifier includes a fully connected layer, an activation function layer, a Dropout layer, and an output layer connected in sequence, wherein the output layer outputs normalized class probabilities through a Softmax function.

[0018] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0019] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0020] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0021] Compared with the prior art, this application has the following beneficial effects: 1) Significantly improved classification accuracy: By simultaneously extracting the spatiotemporal dynamic features of the brain network, the classification accuracy on real datasets reached 76.7%, which is 6% higher than the existing static CNN method (70.7%), and has clear clinical significance.

[0022] 2) Truly achieve dynamic modeling: The sliding window is used to construct dynamic graph sequences, and the DGCN Cell is coupled with spatiotemporal feature extraction, which is more in line with the physiological mechanism of brain functional connectivity.

[0023] 3) Structural matching of non-Euclidean data: Graph neural networks are used to directly process brain network graph data without rearranging the brain network matrix, thus preserving the network topology information.

[0024] 4) End-to-end automation: Enables full automation from raw EEG to classification results, reducing manual intervention and improving diagnostic efficiency. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of a static classification model based on CNN in the existing technology.

[0026] Figure 2 This is a flowchart illustrating the steps of a method for classifying consciousness disorders based on a dynamic graph neural network, as described in this application.

[0027] Figure 3 This is a structural diagram of the Dynamic Graph Neural Network (DGCN) processing module in an embodiment of this application.

[0028] Figure 4 This is a diagram of the internal structure of the Dynamic Graph Convolutional Cell (DGCN Cell) in the embodiments of this application.

[0029] Figure 5 This is a schematic diagram of the hardware structure of the electronic device in the embodiments of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0032] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0033] 1) Disorder of consciousness: refers to a state of decreased level of consciousness or altered content of consciousness due to severe brain injury. In this application, it specifically refers to the minimally conscious state and the vegetative state that need to be distinguished.

[0034] 2) Minimal consciousness: refers to a state of impaired consciousness in which the patient has weak but clear evidence of conscious behavior (such as visual tracking, command-driven actions).

[0035] 3) Vegetative state: refers to a state of consciousness disorder in which the patient has completely lost consciousness and cognitive function, but retains only the basic physiological functions of the brainstem and hypothalamus.

[0036] 4) Electroencephalography (EEG): A physiological monitoring technique that records the electrical activity of neuronal groups in the cerebral cortex using electrodes placed on the scalp. In this application, it specifically refers to resting-state electroencephalography, i.e., the brain electrical signals collected when the patient is quiet, with eyes closed, and without a specific task.

[0037] 5) Phase Lock Value (PLV): A metric used to quantify the phase synchronization between two neural oscillatory signals. Its value ranges from [0, 1], with a larger value indicating a more stable phase relationship and stronger neural synchronization. PLV is a core metric for constructing brain functional networks.

[0038] 6) Functional connectivity: refers to the degree of functional association between different brain regions, usually measured by calculating the statistical dependence (PLV) of neural signals (such as EEG) between two brain regions. Dynamic functional connectivity specifically refers to the phenomenon that the strength of this connection changes over time.

[0039] 7) Graph: A data structure used to represent entities and their relationships, consisting of nodes (or vertices) and edges. In this application, a graph represents a brain functional network within a time window.

[0040] 8) Node: In a graph structure, a node represents an entity. In this invention, each node corresponds to an EEG acquisition electrode channel, representing a specific brain region.

[0041] 9) Edge: In a graph structure, an edge is a line segment connecting two nodes, representing the relationship between the nodes. In this application, an edge represents a significant functional connection (PLV > threshold) between two brain regions, and the edge weight is the PLV value of that connection.

[0042] 10) Node feature: The attribute vector associated with each node. In this application, node feature specifically refers to the degree centrality of each node (brain region), that is, the sum of the weights of all edges connected to that node, used to characterize the local importance of that node in the network.

[0043] 11) Dynamic graph sequence: A collection of graph structures arranged in chronological order. In this application, the dynamic graph sequence is the output of the dynamic brain network construction module, which fully characterizes the evolution of brain functional networks over time.

[0044] 12) Graph Neural Networks: A class of deep learning models specifically designed for processing graph-structured data. They learn the embedded representations of nodes by aggregating information about the nodes' neighbors through a message-passing mechanism.

[0045] 13) Graph Convolutional Network (GCN): A classic implementation of graph neural networks that extracts features from graph data through convolution operations in the spectral or spatial domains.

[0046] 14) Gated Recurrent Unit (GRU): A type of recurrent neural network that effectively captures long-term dependencies in time series and alleviates the vanishing / exploding gradient problem by introducing update gates and reset gates.

[0047] 15) Dynamic Graph Convolutional Network (DGCN): The core model proposed in this application. Specifically, it refers to a neural network architecture that couples graph convolutional networks with gated recurrent units at the unit level, using GCN layers to replace the fully connected layers inside the GRU, for processing dynamic graph sequences. Its basic unit is called a DGCN Cell.

[0048] 16) DGCN Cell: The basic processing unit of the DGCN model in this application. Its input is the graph data of the current time step and the hidden state of the previous time step. Through the internally coupled GCN and GRU operations, it outputs the hidden state of the current time step.

[0049] 17) Degree centrality: One of the metrics for measuring the centrality of nodes in graph theory. In this application, it specifically refers to weighted degree centrality, which is the sum of the weights of all adjacent edges of a node.

[0050] 18) Hidden state: In recurrent neural networks (such as GRU), the internal state vector used to encapsulate and transmit historical sequence information.

[0051] 19) Non-Euclidean Data: Data is broadly divided into two categories: Euclidean data and non-Euclidean data. Euclidean data exhibits excellent translation invariance; each pixel has the same number of neighboring pixels. Common examples of this type include images (2D), text (1D), and speech data. Non-Euclidean data lacks translation invariance; each pixel has a different number of neighboring pixels. Common examples of this type include knowledge graphs, social networks, and chemical molecular structures.

[0052] Accurate diagnosis of disorders of consciousness (DoC), especially the distinction between minimally conscious state (MCS) and vegetative state (VS), remains a core challenge in clinical neuroscience. Currently, the revised Coma Recovery Scale (CRS-R) is the gold standard for behavioral assessment; however, its assessment process suffers from inherent limitations, including high subjectivity, a high rate of misdiagnosis (up to 40%), and reliance on the patient's motor response capabilities.

[0053] To overcome the limitations of behavioral assessment, objective auxiliary diagnostic techniques based on electroencephalography (EEG) have emerged. EEG offers advantages such as high temporal resolution, portable equipment, low cost, and bedside monitoring. Early studies often employed machine learning methods, relying on experts to manually extract EEG features (such as power spectrum, symmetry index, and functional connectivity indices), which resulted in cumbersome feature engineering and weak generalization ability.

[0054] In recent years, deep learning techniques, particularly convolutional neural networks (CNNs) and graph neural networks (GNNs), have been introduced into EEG analysis. CNNs excel at processing image-like data (such as rearranged EEG temporal spectrograms or functional connectivity matrices), but their inherent assumptions of local weight sharing and spatial invariance are incompatible with the complex, non-Euclidean topology of brain functional networks. Graph convolutional networks (GCNs) can naturally handle graph-structured data, providing a superior framework for brain network analysis. However, existing GCN-based methods generally suffer from a fundamental flaw: they typically average data from the entire EEG recording period (e.g., several minutes) to construct a single, static brain functional network. This approach completely ignores the crucial characteristic of consciousness as a dynamic process, where the underlying brain network connections are constantly and dynamically reorganized on millisecond to second timescales. Therefore, static GCN models fail to capture the temporal information that may be crucial for distinguishing states of consciousness, leading to a bottleneck in their classification performance (e.g., accuracy on public datasets is often difficult to exceed 65%).

[0055] For example, one existing technical solution is a CNN-based classification scheme based on the rearranged EEG functional connectivity matrix. This scheme is detailed below: 1) Data preparation and “static” network construction: Resting-state EEG data from patients is collected and preprocessed. The average phase lock value (PLV) between all electrode pairs is calculated over the entire recording period (e.g., 2 minutes), resulting in a unique, globally averaged 62x62 functional connectivity matrix. This matrix represents the "average state" of brain functional connectivity throughout the recording period.

[0056] 2) Unnatural processing performed to adapt to CNN: Because CNNs excel at image processing, and the electrode order of the original PLV matrix (e.g., arranged as Fp1, Fp2, F3, F4...) does not intuitively reflect spatial proximity, the PLV matrix needs to be "rearranged." Based on the 3D coordinates of the electrodes, a clustering algorithm (such as K-means) is used to divide the 68 electrodes into several regions, and the electrodes of these regions are arranged together in the matrix. This step is to "cater" to the architectural characteristics of CNNs, rather than based on the actual functional partitioning of the brain; it is a heuristic data preprocessing method that may introduce bias.

[0057] 3) Model Architecture and Classification: The rearranged static PLV matrix is ​​input into a three-layer CNN model. This CNN contains convolutional layers, pooling layers, normalization layers, and finally outputs the classification result of MCS or VS through a fully connected layer. An example of a CNN model is as follows. Figure 1As shown: the rearranged static PLV matrix is ​​flattened into a vector after 3*(convolution + normalization), and finally outputs the classification probability after a fully connected layer and Softmax operation. Throughout the entire processing flow, this model operates only on this single, averaged static matrix.

[0058] However, the above-mentioned technical solution has the following drawbacks: 1) Dynamic information is completely lost: Calculating the average PLV of the entire EEG segment obliterates the dynamic changes in brain functional connectivity over time. Numerous neuroscience studies have shown that the maintenance of consciousness is closely related to the dynamic connectivity of brain networks, particularly higher cognitive networks such as the prefrontal-parietal lobe. Static methods cannot utilize this crucial information.

[0059] 2) Model and data structure mismatch: Forcibly rearranging the PLV matrix to fit CNNs disrupts the spatial topological relationships of brain regions represented by the functional connectivity matrix itself. This rearrangement is heuristic and may not be reasonable, and the local convolutional kernels of CNNs are unlikely to effectively capture long-range functional connections in brain networks.

[0060] 3) Limited classification accuracy: Due to the two fundamental flaws mentioned above, the best classification accuracy of the closest existing technical solution on the same dataset is only 70.7%, which still has considerable room for improvement and is difficult to meet the needs of high-precision clinical auxiliary diagnosis.

[0061] In view of this, this application provides a method, electronic device, storage medium and program product for classifying consciousness disorders based on dynamic graph neural networks, which aims to solve the problems in the prior art that the static brain network analysis cannot capture the dynamic characteristics of consciousness-related EEG, and the mismatch between the model and the brain network data structure, so as to achieve more accurate automatic classification of MCS and VS patients and provide interpretable decision-making basis for clinical practice.

[0062] This application provides a method for classifying disorders of consciousness based on dynamic graph neural networks, relating to the interdisciplinary field of medical artificial intelligence and neural engineering. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing a method for classifying disorders of consciousness based on dynamic graph neural networks, but is not limited to the above forms.

[0063] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0064] like Figure 2 As shown, this embodiment provides a method for classifying consciousness disorders based on dynamic graph neural networks, including the following steps: S1: Acquire resting-state electroencephalogram (EEG) signals from patients with impaired consciousness; S2: Preprocess the electroencephalogram (EEG) signal to obtain a multi-channel EEG signal; S3: Based on the sliding window and phase lock value, the multi-channel EEG signal is converted into a dynamic brain function network sequence, wherein each time window corresponds to a brain function connection graph, and the brain function connection graph uses the EEG acquisition channel as the node and the phase lock value between the channels as the edge weight. S4: The dynamic brain function network sequence is input into a dynamic graph neural network model for processing. The dynamic graph neural network model consists of multiple time-connected dynamic graph convolutional units. Each dynamic graph convolutional unit is used to extract spatiotemporal features based on the brain function connectivity graph of the current time window and the hidden state of the previous time window through coupled graph convolutional networks and gated recurrent units, and outputs the hidden state of the current time window. S5: Based on the final hidden state output by the dynamic graph neural network model, a classification result of consciousness disorder is obtained through a classifier. The classification result includes minimal consciousness state and vegetative state.

[0065] In one embodiment, the raw EEG signals of patients with altered consciousness undergo preliminary preprocessing, such as denoising, to obtain resting-state multi-channel EEG signals. These signals are then input into a dynamic graph neural network (DGCN) construction module to convert them into dynamic graph sequences that the model can directly process. Each patient's dynamic graph sequence is considered a sample. The resulting sample set is randomly stratified into a training set and a test set (4:1). The training set uses 5-fold cross-validation to train the proposed DGCN dynamic graph neural network model, while the test set evaluates the network performance of the trained DGCN model. The trained model can be directly applied clinically. For patients with altered consciousness requiring diagnostic classification, their resting-state EEG data is collected. After the aforementioned preprocessing and dynamic graph sequence construction, the data is input into the trained DGCN to obtain classification results as minimal disturbance of consciousness or vegetative state.

[0066] The solutions of the embodiments of this application will be described in detail below with reference to the accompanying drawings and specific application examples.

[0067] (1) Dynamic brain network construction module This module is responsible for converting raw, preprocessed resting-state multichannel EEG signals into a dynamic graph sequence that the model can process. Its input is EEG data of dimension [C, T] (C is the number of channels, e.g., 62; T is the number of time steps). The output is a dynamic graph sequence containing T time steps. , , ,..., The main steps are as follows: After extracting the EEG signal of the target frequency band using a filter, the sliding window technique is used to dynamically calculate the PLV value between the two channels, resulting in a PLV sequence with dimensions [T, C, C]. Then, each element of the PLV matrix is ​​compared with a set threshold, and connections greater than the threshold are considered valid connections. Thus, a T-layer brain network effective connection graph is constructed, forming a dynamic brain network.

[0068] (2) Dynamic Graph Neural Network (DGCN) processing module This module is the core of this embodiment, and its structure is as follows: Figure 3As shown, it consists of T time-connected DGCN Cells, with each Cell responsible for processing the graph data at one time step. And pass the hidden state. From left to right, they represent the temporal input, and for each time step, the graph data input... First, the spatial features are extracted using GCN and then passed to GRU. GRU further extracts temporal features based on the hidden state of the previous time step and the spatial features extracted by GCN, and passes these features as the hidden state to the next time step. The basic unit is DGCN Cell.

[0069] DGCN Cell: Its internal structure is as follows Figure 4 As shown, it innovatively couples the spatial feature extractor (GCN) with the temporal feature extractor at the unit level. Specifically, it uses multiple GCN layers to replace the fully connected layers in the traditional GRU for input transformation and state transformation. (From below...) Figure 4 It can be seen that this unit accepts graph data with node features, edge indices, and edge weights. After extracting spatial feature information through three different n-layer GCNs, it is integrated into the GRU. The GRU further combines the hidden state of the previous time step to update the hidden state of the current time step and passes it to the next time step.

[0070] Graph Convolutional Layer (GCN): This embodiment uses two GCN layers consecutively multiple times within the DGCN Cell. The first GCN layer has an input dimension of 1 (node ​​feature dimension) and an output dimension of 32; the second GCN layer has both input and output dimensions of 32. The propagation rule for the GCN layers is as follows:

[0071] in, The graph G has an adjacency matrix with self-connections added. It is the identity matrix. ,Right now yes The degree matrix, This represents the trainable weight matrix of layer l. This represents the sigmoid activation function. After one graph convolution, it can be visualized that for each node in each feature dimension, the feature information of its direct neighbors is first collected and combined with its own features to form a new node feature, which is then updated. Using two consecutive GCN layers means expanding the information propagation range to two levels of neighbor nodes. In this embodiment, it is preferable to use three consecutive GCN layers, where the input and output dimensions of the last two GCN layers are both 32.

[0072] Gated Recurrent Unit (GRU): Let the current input be... ( Includes node features The hidden state at the previous moment is... The formula for calculating its inclusion of GCN is as follows: a) Update Gate :

[0073] b) Reset the door :

[0074] c) Candidate hidden state

[0075] d) Current hidden state

[0076] in, , , It consists of three independent, structurally identical continuous graph convolutional layers (1-dimensional input and 32-dimensional output). The key innovation here is that the gating signal and state update of GRU are calculated by GCN based on the graph structure, rather than by traditional fully connected layers.

[0077] (3) Classification output module 1) Feature flattening: Take the final hidden state output by the last DGCN Cell. (Dimension [C, 32], where C is the number of graph nodes or the number of channels in the EEG), flatten it into a one-dimensional feature vector (dimension [C*32], such as [62*32=1984]).

[0078] 2) Fully connected classifier: This classifier consists of two fully connected layers.

[0079] a) FC1: Input dimension [C*32], output dimension 256, followed by ReLU activation function and Dropout layer (dropout rate p=0.5).

[0080] b) FC2: Input dimension 256, output dimension 2, corresponding to the two categories MCS and VS respectively.

[0081] 3) Output Unit: Apply the Softmax function to the output of FC2 to obtain the normalized class probability, and output the class with the higher probability as the final diagnostic result.

[0082] (4) Advantages and beneficial effects In summary, compared with the prior art, the method of this embodiment has the following advantages and beneficial effects: 1) Significantly improved classification accuracy: By simultaneously and coupledly extracting the spatial topological features and temporal dynamic evolution features of the brain network, this embodiment achieved a classification accuracy of 76.7% on a real dataset containing 220 patients, which is 6 percentage points higher than the closest existing technology (static CNN, 70.7%).

[0083] 2) Truly Dynamic Modeling: The DGCN Cell structure proposed in this embodiment is the first time that GCN and GRU have been integrated at such a close unit level in the field of consciousness disorder diagnosis. This enables the model to learn dynamic reorganization patterns of brain functional connections at millisecond resolution, which is more consistent with the physiological mechanisms of brain operation.

[0084] 3) More suitable for graph data models with non-Euclidean properties of EEG: For brain-related data such as EEG, traditional convolutional neural networks are often difficult to apply due to their inherent non-Euclidean properties. The dynamic graph neural network model proposed in this embodiment is based on graph data driving, which is more advantageous for processing non-Euclidean data and has higher accuracy.

[0085] 4) End-to-end automation: It realizes fully automated analysis from raw EEG signals to final classification results, minimizing human intervention and reliance on expert experience, and improving diagnostic efficiency.

[0086] (5) Industrial applicability For example, the method of this embodiment is applicable to: 1) Auxiliary diagnosis of patients with disorders of consciousness by neurology and rehabilitation departments; 2) Status monitoring and classification in brain-computer interfaces; 3) Brain network analysis of neuropsychiatric diseases (such as Alzheimer's disease and depression); 4) Research institutions conduct research on the dynamic characteristics of brain networks.

[0087] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0088] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0089] Please see Figure 5 , Figure 5 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 501 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 502 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 502 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 502 and is called and executed by the processor 501 using the methods described in the embodiments of this application. The input / output interface 503 is used to implement information input and output; The communication interface 504 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 505 transmits information between various components of the device (e.g., processor 501, memory 502, input / output interface 503, and communication interface 504); The processor 501, memory 502, input / output interface 503, and communication interface 504 are connected to each other within the device via bus 505.

[0090] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0091] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0092] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0093] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0094] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented in the embodiments of this program product are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. The executable computer program code or "code" used to perform the various embodiments can be written in high-level programming languages ​​such as C, C++, Python, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0095] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0096] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0097] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0098] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0099] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0100] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0102] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0104] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for classifying consciousness disorders based on dynamic graph neural networks, characterized in that, The method includes the following steps: Obtaining resting-state electroencephalogram (EEG) signals from patients with disorders of consciousness; The electroencephalogram (EEG) signals are preprocessed to obtain multi-channel EEG signals; Based on sliding windows and phase-locking values, the multi-channel EEG signals are converted into a dynamic brain function network sequence, wherein each time window corresponds to a brain function connectivity graph, and the brain function connectivity graph uses the EEG acquisition channels as nodes and the phase-locking values ​​between channels as edge weights. The dynamic brain function network sequence is input into a dynamic graph neural network model for processing. The dynamic graph neural network model consists of multiple temporally connected dynamic graph convolutional units. Each dynamic graph convolutional unit is used to extract spatiotemporal features based on the brain function connectivity graph of the current time window and the hidden state of the previous time window through coupled graph convolutional networks and gated recurrent units, and outputs the hidden state of the current time window. Based on the final hidden state output by the dynamic graph neural network model, a classifier is used to obtain a classification result of consciousness disorder, which includes a minimally conscious state and a vegetative state.

2. The method according to claim 1, characterized in that, The process of converting multi-channel EEG signals into dynamic brain functional network sequences includes: For the signal within each time window, calculate the phase lock value between all pairs of channels to obtain the phase lock value matrix; Connections in the phase-locked value matrix that are greater than a preset threshold are considered valid connections, and a brain functional connectivity map for that time window is constructed, wherein the node feature is the weighted degree centrality of the node.

3. The method according to claim 1, characterized in that, The internal operations of the dynamic graph convolutional unit include: The gate signals are updated based on the current brain functional connectivity graph using the first graph convolutional network; The reset gate signal is calculated based on the current brain functional connectivity graph using the second graph convolutional network; The candidate hidden state is calculated by using the third graph convolutional network based on the current brain functional connectivity map and in combination with the reset gate signal and the previous hidden state. Based on the update gate signal, the previous hidden state, and the candidate hidden state, calculate and output the current hidden state.

4. The method according to claim 3, characterized in that, The first graph convolutional network, the second graph convolutional network, and the third graph convolutional network each include at least two graph convolutional layers, wherein the propagation rule of at least one graph convolutional layer is as follows: in, To add self-connected adjacency matrices, for The degree matrix, For the first Layer node features For trainable weight matrix, This is the activation function.

5. The method according to claim 1, characterized in that, The dynamic graph convolutional unit includes: At least one spatial feature extraction subunit, consisting of consecutive graph convolutional network layers, is used to extract spatial features of nodes based on graph topology. The temporal feature fusion subunit, based on a gated loop unit structure, is used to fuse the current spatial features extracted by the spatial feature extraction subunit with the hidden state of the previous time step, and output the hidden state of the current time step. The calculation of the update gate, reset gate, and candidate hidden state in the gated loop unit is performed by the independent spatial feature extraction subunit based on the currently input graph data.

6. The method according to claim 5, characterized in that, The calculation process of the time feature fusion subunit is defined by the following formula: Update Gate : Reset door : Candidate hidden state : Current hidden state : in, The graph data is for the current time step. The hidden state of the previous time step. , , These are three independent graph convolutional network subunits with identical structures. For activation functions; For Hadamah accumulation; To update the trainable weight matrix of the gate, To update the door offset, To reset the trainable weight matrix of the gate, To reset the door offset, is the trainable weight matrix for the candidate hidden states.

7. The method according to claim 1, characterized in that, The classifier comprises a fully connected layer, an activation function layer, a Dropout layer, and an output layer connected in sequence. The output layer outputs normalized class probabilities through the Softmax function.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images