CN-122020358-A - Consciousness disturbance classification method and related equipment based on dynamic graph neural network

CN122020358ACN 122020358 ACN122020358 ACN 122020358ACN-122020358-A

Abstract

The embodiment of the application provides a consciousness disturbance classification method and related equipment based on a dynamic graph neural network, belonging to the technical field of medical artificial intelligence and neural engineering intersection. The method comprises the steps of collecting and preprocessing resting state brain electrical signals of a patient, calculating a phase locking value through a sliding window to construct a dynamic brain function network graph sequence, inputting the dynamic graph sequence into a dynamic graph neural network model for processing, wherein the model is composed of a plurality of dynamic graph convolution units connected in time sequence, each unit is combined with a gate control circulation unit through a coupled graph convolution network to extract space-time dynamic characteristics of the brain network, and based on a final hidden state output by the model, a classification result of a micro-consciousness state or a plant state is obtained through a classifier. The application solves the problem that the existing static analysis method can not capture the dynamic characteristics of the brain network and the mismatch between the model and the data structure, realizes the automatic classification of the consciousness disturbance with higher precision, and has good clinical auxiliary diagnosis value.

Inventors

CHEN HE
XIAO JIAQING
WANG YONG
GUAN LONGZHOU
LI XIAOLI

Assignees

华南理工大学

Dates

Publication Date: 20260512
Application Date: 20251224

Claims (10)

1. A method for classifying consciousness disturbance based on a dynamic graph neural network, which is characterized by comprising the following steps: Acquiring a resting state electroencephalogram signal of a patient with consciousness disturbance; preprocessing the electroencephalogram signals to obtain multichannel electroencephalogram signals; based on sliding windows and phase locking values, converting the multichannel brain electrical signals into dynamic brain function network sequences, wherein each time window corresponds to a brain function connection diagram, the brain function connection diagram takes brain electrical acquisition channels as nodes and phase locking values among the channels as side weights; Inputting the dynamic brain function network sequence into a dynamic graph neural network model for processing, wherein the dynamic graph neural network model consists of a plurality of dynamic graph convolution units connected in time sequence, and each dynamic graph convolution unit is used for extracting space-time characteristics through a coupled graph convolution network and a gating circulation unit based on a brain function connection graph of a current time window and the hidden state of a last time window and outputting the hidden state of the current time window; based on the final hidden state output by the dynamic graph neural network model, a classification result of the consciousness disturbance is obtained through a classifier, and the classification result comprises a micro consciousness state and a plant state.
2. The method of claim 1, wherein the converting the multichannel brain electrical signal into a dynamic brain function network sequence comprises: For signals in each time window, calculating phase locking values between every two channels to obtain a phase locking value matrix; And regarding the connection larger than a preset threshold value in the phase locking value matrix as effective connection, and constructing a brain function connection diagram of the time window, wherein the node characteristic is the weight centrality of the node.
3. The method of claim 1, wherein the internal operation of the dynamic graph convolution unit comprises: Calculating an update gating signal based on the current brain function connection graph through a first graph convolution network; calculating a reset gate signal based on the current brain function connection graph through a second graph convolution network; Calculating candidate hidden states based on the current brain function connection diagram through a third graph convolution network and combining the reset gate signal and the previous hidden state; and calculating and outputting a current hidden state based on the updated gate signal, the last hidden state and the candidate hidden state.
4. A method according to claim 3, wherein the first, second and third convolution networks each comprise at least two layers of convolution layers, wherein the propagation rule of at least one layer of convolution layers is: Wherein, the In order to add the adjacency matrix of the self-connection, Is that Is used for the degree matrix of the (c), Is the first The characteristics of the layer node are that, In order for the weight matrix to be trainable, To activate the function.
5. The method of claim 1, wherein the dynamic map convolution unit comprises: At least one spatial feature extraction subunit, composed of successive graph convolution network layers, for extracting spatial features of nodes based on the topology of the graph; The time feature fusion subunit is used for fusing the current space features extracted by the space feature extraction subunit with the hiding state of the previous time step based on the gating circulating unit structure and outputting the hiding state of the current time step; The computation of the update gate, the reset gate and the candidate hidden state in the gating circulation unit is performed by the independent spatial feature extraction subunit based on the currently input graph data.
6. The method of claim 5, wherein the calculation process of the temporal feature fusion subunit is defined by the following formula: Update door : Reset gate : Candidate hidden states : Current hidden state : Wherein, the As the map data of the current time step, The state is hidden for the last time step, 、、 The network subunits are rolled up for three independent figures of identical structure, Is an activation function; Is Hadamard product; to update the trainable weight matrix of the gate, In order to update the gate bias, To reset the trainable weight matrix of the gate, In order to reset the gate bias, A trainable weight matrix for candidate hidden states.
7. The method of claim 1, wherein the classifier comprises a fully connected layer, an activation function layer, a Dropout layer, and an output layer connected in sequence, the output layer outputting normalized class probabilities by a Softmax function.
8. An electronic device comprising a memory storing a computer program and a processor implementing the method of any of claims 1 to 7 when the computer program is executed by the processor.
9. A computer readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the method of any one of claims 1 to 7.
10. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the method of any one of claims 1 to 7.

Description

Consciousness disturbance classification method and related equipment based on dynamic graph neural network Technical Field The application relates to the technical field of medical artificial intelligence and neural engineering intersection, in particular to a consciousness disturbance classification method based on a dynamic graph neural network and related equipment. Background Consciousness disturbance (Disorders of Consciousness, doC) is a common clinical state following severe brain injury, where accurate identification of the micro-conscious state (MCS) and plant state (VS) is critical for therapeutic decisions and prognostic assessment. At present, the clinic mainly relies on a revising coma recovery scale (CRS-R) to conduct behavioural assessment, but the method has the defects of strong subjectivity, high misdiagnosis rate, dependence on the motor response capability of patients and the like. In recent years, objective evaluation methods based on electroencephalogram (EEG) are increasingly emerging. The traditional method mostly adopts machine learning, relies on manual extraction of characteristics, and has limited generalization capability. Deep learning techniques such as Convolutional Neural Networks (CNNs) are introduced, but CNNs process grid-like data (e.g., images), whereas brain function networks are essentially non-euclidean structures, CNNs require rearrangement of the function connection matrices to accommodate their structure, a process that can introduce bias and destroy the topology of the network. In addition, the existing method generally averages the whole EEG signal into a static functional connection network for analysis, ignores the dynamic change characteristic of the brain functional connection on a time scale, and researches show that the maintenance of consciousness is closely related to the dynamic recombination of the brain network. Therefore, it is difficult for the static analysis method to capture the time-series dynamic characteristics of the conscious state, resulting in limited classification performance. Disclosure of Invention The embodiment of the application mainly aims to provide a consciousness disturbance classification method, electronic equipment, storage medium and program product based on a dynamic graph neural network, so as to solve the problems of losing dynamic information in static analysis and mismatching of a traditional model and a non-European electroencephalogram data structure, and realize higher-precision consciousness disturbance automatic classification. In order to achieve the above object, an aspect of an embodiment of the present application provides a method for classifying a conscious disturbance based on a dynamic graph neural network, the method including: Acquiring a resting state electroencephalogram signal of a patient with consciousness disturbance; preprocessing the electroencephalogram signals to obtain multichannel electroencephalogram signals; based on sliding windows and phase locking values, converting the multichannel brain electrical signals into dynamic brain function network sequences, wherein each time window corresponds to a brain function connection diagram, the brain function connection diagram takes brain electrical acquisition channels as nodes and phase locking values among the channels as side weights; Inputting the dynamic brain function network sequence into a dynamic graph neural network model for processing, wherein the dynamic graph neural network model consists of a plurality of dynamic graph convolution units connected in time sequence, and each dynamic graph convolution unit is used for extracting space-time characteristics through a coupled graph convolution network and a gating circulation unit based on a brain function connection graph of a current time window and the hidden state of a last time window and outputting the hidden state of the current time window; based on the final hidden state output by the dynamic graph neural network model, a classification result of the consciousness disturbance is obtained through a classifier, and the classification result comprises a micro consciousness state and a plant state. In some embodiments, the converting the multichannel brain electrical signal into a dynamic brain function network sequence comprises: For signals in each time window, calculating phase locking values between every two channels to obtain a phase locking value matrix; And regarding the connection larger than a preset threshold value in the phase locking value matrix as effective connection, and constructing a brain function connection diagram of the time window, wherein the node characteristic is the weight centrality of the node. In some embodiments, the internal operations of the dynamic graph convolution unit include: Calculating an update gating signal based on the current brain function connection graph through a first graph convolution network; calculating a reset gate signal based on