CN-121987205-A - Electroencephalogram signal depression identification method based on multi-view capsule graph network

CN121987205ACN 121987205 ACN121987205 ACN 121987205ACN-121987205-A

Abstract

An electroencephalogram depression recognition method based on a multi-view capsule graph network belongs to the field of electroencephalogram analysis in the field of biological medicine and comprises the steps of preprocessing electroencephalogram signals, extracting node characteristics, constructing a topological graph, a functional graph and a causal graph according to the node characteristics, carrying out self-adaptive fusion on the topological graph, the functional graph and the causal graph to obtain a fusion graph, constructing a bidirectional regulation mechanism of the capsule graph network, constructing a loss function and a total training target of the capsule graph network, and carrying out depression recognition process based on a dynamic routing mechanism of the capsule graph network. The invention introduces a capsule graph network to carry out structural modeling on the topological structure and the connection mode between the electroencephalogram channels, enhances the interrelation between the electroencephalogram channels through self-adaptive fusion, enables the capsule graph network to more accurately capture the interaction and coupling characteristics of brain areas, introduces a bidirectional adjustment mechanism to flexibly adjust the contribution degree of the characteristics, and enhances the information transmission capacity of the capsule graph network in a global context and the sensitivity and robustness to the depression related characteristics.

Inventors

LI QI
DU YUHANG
ZHANG HANG
WU YAN
GAO NING

Assignees

长春理工大学

Dates

Publication Date: 20260508
Application Date: 20260407

Claims (9)

1. The electroencephalogram signal depression identification method based on the multi-view capsule graph network is characterized by comprising the following steps of: Step one, preprocessing an electroencephalogram signal; step two, extracting node characteristics of the preprocessed electroencephalogram signals; Constructing a topological graph, a functional graph and a causal graph according to the node characteristics to represent different brain connection mechanisms; Step four, a bidirectional regulating mechanism of a capsule graph network; step five, constructing a loss function and a total training target of a capsule graph network; And step six, a depression identification process based on a capsule graph network dynamic routing mechanism.
2. The brain electrical signal depression recognition method based on the multi-view capsule graph network is characterized by comprising the steps of firstly unifying channel names and numbers according to standard electrode layout, unifying sampling rates, removing high-frequency noise and low-frequency noise by adopting band-pass filtering, suppressing 50 Hz power frequency interference by notch filtering, observing waveforms, power spectrums and scalp distribution of independent components in brain electrical signals, eliminating components related to eye movement or myoelectricity, and performing unified reference calibration on all the brain electrical signals by whole brain average heavy reference.
3. The brain electrical signal depression recognition method based on the multi-view capsule graph network according to claim 1, wherein in the second step, brain electrical channels or brain areas are regarded as nodes, inter-channel connection is regarded as edges, preprocessed EEG segments are taken as input, convolution modeling is carried out on time sequences of the nodes, a local mode is captured through sliding of convolution kernels on a time dimension, time structural features are extracted through pooling operation and nonlinear activation, and each node is mapped into a fixed-length time domain embedded vector, namely node features.
4. The brain electrical signal depression recognition method based on the multi-view capsule graph network according to claim 1, wherein in the third step, when a topological graph is constructed, the euclidean distance between any two electrodes is calculated first, the euclidean distance is mapped to a similarity between (0, 1) through a gaussian kernel function to construct a basic topological adjacent matrix, and then a radial basis function is used for initializing the topological relation of the basic topological adjacent matrix to obtain the topological graph.
5. The brain electrical signal depression recognition method based on the multi-view capsule graph network according to claim 1, wherein in the third step, when the functional graph is constructed, mutual information is firstly adopted as a functional connection measure, the statistical dependency intensity is calculated by estimating the joint probability distribution and the edge distribution of the time sequences of two nodes in an EEG segment, and the mutual information of all node pairs is arranged into a matrix and normalized to obtain the functional graph.
6. The brain-electrical signal depression recognition method based on a multi-view capsule graph network according to claim 1, wherein in the third step, when a causal graph is constructed, a causal relationship between nodes is obtained by comparing residual differences when only self history information is used and when two node history information is used for predicting a target node at the same time, GC test is used, prediction errors between the nodes are calculated, and a causal graph of a time-space domain is constructed.
7. The electroencephalogram signal depression recognition method based on the multi-view capsule graph network according to claim 1 is characterized in that in the third step, a dynamic weighting mode is adopted to adaptively fuse a topological graph, a functional graph and a causal graph, three adjacent matrixes of topology, function and causal are combined with time domain embedded vectors of nodes, the time domain embedded vectors of each node pair are spliced, trainable parameter vectors are defined for the topological graph, the functional graph and the causal graph respectively, view attention weights of the node pairs are obtained through softmax, the relative importance of different views is adaptively adjusted on an edge level by using the view attention weights, and a multi-view fused graph is obtained.
8. The brain electrical signal depression recognition method based on the multi-view capsule graph network as claimed in claim 1, wherein in the fourth step, global information of the classified capsules is reversely transmitted to the primary capsules through the advanced capsules to realize feature adjustment from top to bottom Aggregation is carried out to form global class vectors Global class vector As conditional input, the adjustment parameters are obtained by linear transformation And Primary capsule for lower layers Feature scaling and shifting are performed.
9. The method for identifying the depression of the brain electrical signal based on the multi-view capsule graph network according to claim 1, wherein in the fifth step, the loss function comprises a margin loss Reconstruction loss Regular term of attention entropy The total training target is , And Are super-parameters used for balancing the relative importance among classification accuracy, reconstruction capability and attention entropy regularization items.

Description

Electroencephalogram signal depression identification method based on multi-view capsule graph network Technical Field The invention belongs to the technical field of electroencephalogram analysis in the field of biological medicine, and particularly relates to an electroencephalogram depression identification method based on a multi-view capsule graph network. Background Depression is a common mental health problem, and traditional identification methods rely on scale evaluations and clinical interviews, and are highly subjective and prone to misdiagnosis or missed diagnosis. Electroencephalogram (EEG) is used as a noninvasive and objective neurophysiologic index, has millisecond-level time resolution, can capture the electroencephalogram activity in real time, and provides new possibility for auxiliary diagnosis of depression. There are significant differences in EEG signal between depressed patients and healthy people, especially in the frontal lobe area where there is asymmetry in the activity. Therefore, the depression identification technology based on the electroencephalogram has wide application prospect, and can provide a new way for accurate and efficient auxiliary diagnosis of depression. At present, a deep learning-based depression recognition method mainly carries out modeling around physiological signals such as multichannel electroencephalogram and the like, generally represents the electroencephalogram signals as euclidean tensors, and carries out end-to-end feature learning by utilizing a convolution network, a circulation network or an attention mechanism so as to improve the representation capability and stability of a depression-related mode. According to the depression recognition method based on the graph neural network, an electroencephalogram channel or brain region is regarded as a node, inter-channel connection is regarded as an edge, a brain network structure is described based on a certain connection characteristic (such as topological connection or functional connection), and characteristic propagation and aggregation are carried out on the graph to capture association information among channels. The existing depression recognition method closest to the invention is a deep learning method based on EEG, which adopts a TSF-MDD model and combines time-frequency characteristics to recognize major depression (MDD). Specifically, the method first converts the EEG signal into a four-dimensional spatio-temporal frequency domain representation through a data reconstruction scheme, capturing features of the EEG signal in the temporal, spatial and frequency dimensions. Subsequently, the TSF-MDD model uses a 3D convolutional neural network (3D-CNN) and a capsule network (CapsNet) for feature extraction, where the 3D-CNN is used to process four-dimensional data, preserving the spatio-temporal features of the signal, while CapsNet further captures higher-order spatial features. Although the method improves the feature learning capability and robustness of the TSF-MDD model to a certain extent, the following obvious defects still exist: firstly, the multichannel signal is simply regarded as Euclidean tensor, topology interaction between brain regions related to depression is ignored, and channel coupling strength cannot be effectively quantized or nonlinear embedding is adapted, so that modeling of overall brain network constraint is insufficient. Secondly, only a single connection feature (such as topological connection, functional connection or causal connection) is considered, but all connection features are not considered, different connection features cannot be effectively fused, and the associated information between the electroencephalogram channels cannot be fully utilized. Thirdly, the capsule graph network has only a feedback mechanism from bottom to top, and can effectively extract a local brain connection mode, but omits dynamic adjustment of global context, so that aggregation of the bottom capsule graph is not constrained by task semantics, a mode which is irrelevant to depression and is stable in statistics is easy to learn, sensitivity of the model to depression-related features is reduced in a noise environment, and overall classification accuracy and robustness are finally affected. Disclosure of Invention In order to solve the problems of the existing depression identification method, the invention provides an electroencephalogram signal depression identification method based on a multi-view capsule graph network. The technical scheme adopted by the invention for solving the technical problems is as follows: the invention provides an electroencephalogram signal depression identification method based on a multi-view capsule graph network, which comprises the following steps of: Step one, preprocessing an electroencephalogram signal; step two, extracting node characteristics of the preprocessed electroencephalogram signals; Constructing a topological graph, a functional graph