CN114219014A - Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method - Google Patents

Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method Download PDF

Info

Publication number
CN114219014A
CN114219014A CN202111425371.0A CN202111425371A CN114219014A CN 114219014 A CN114219014 A CN 114219014A CN 202111425371 A CN202111425371 A CN 202111425371A CN 114219014 A CN114219014 A CN 114219014A
Authority
CN
China
Prior art keywords
node
electroencephalogram
attention
matrix
regression
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111425371.0A
Other languages
Chinese (zh)
Inventor
郭艳蓉
陈涛
郝世杰
洪日昌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei University of Technology
Original Assignee
Hefei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hefei University of Technology filed Critical Hefei University of Technology
Priority to CN202111425371.0A priority Critical patent/CN114219014A/en
Publication of CN114219014A publication Critical patent/CN114219014A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/16—Devices for psychotechnics; Testing reaction times ; Devices for evaluating the psychological state
    • A61B5/165—Evaluating the state of mind, e.g. depression, anxiety
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/369—Electroencephalography [EEG]
    • A61B5/372—Analysis of electroencephalograms
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Psychiatry (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • Veterinary Medicine (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Public Health (AREA)
  • Animal Behavior & Ethology (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Surgery (AREA)
  • Pathology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Probability & Statistics with Applications (AREA)
  • Psychology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Signal Processing (AREA)
  • Physiology (AREA)
  • Fuzzy Systems (AREA)
  • Child & Adolescent Psychology (AREA)
  • Developmental Disabilities (AREA)
  • Educational Technology (AREA)
  • Hospice & Palliative Care (AREA)
  • Social Psychology (AREA)

Abstract

The invention discloses an electroencephalogram-based method for identifying and classifying self-attention-drawing pooled depressive disorder, which comprises the steps of acquiring a data set of electroencephalogram signals and preprocessing the data set; constructing an adjacency matrix based on Euclidean distance and a left-right semi-brain connection relation, and calculating a node importance score based on an attention-seeking pooling operator to obtain a target node to obtain a training-optimized soft label classification model of the depressive disorder; modeling a classification task and a regression task to obtain an output soft label and a regression score, and converting the predicted regression score into label probability distribution; and finally obtaining joint label distribution for classification and clinical scores for regression according to the model. The method constructs an adjacent matrix and introduces a self-attention-seeking pooling model to model the electroencephalogram signals of healthy people and patients with depressive disorder so as to complete identification of the depressive disorder. And (4) mining potential relations between samples by using soft labels, and simultaneously constraining the relation between the regression task and the classification task by adopting JS divergence.

Description

Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method
Technical Field
The invention relates to the technical field of electroencephalogram signal identification, in particular to an electroencephalogram-based method for identifying and classifying self-attention-drawing pooled depressive disorder.
Background
Depressive disorder is a common mental disorder. Patients with depressive disorders have symptoms of insomnia, anxiety, irritability, etc., while patients with severe depressive disorders may even be accompanied by suicidal behavior. According to the incomplete statistics of the world health organization, there are more than 3.5 million patients with depressive disorders worldwide, while statistical data from china show that more than 9500 million people across the country have depressive disorders. In addition, depressive disorders also place a heavy economic burden on society. Studies have shown that depressive disorders cause economic losses of $ 1 trillion per year. In fact, timely treatment can reduce the burden it carries. With the increasing number of patients with depressive disorders, there is an urgent need to design an objective and automatic algorithm for detecting depressive disorders. In recent years, due to the rapid development of electroencephalogram equipment, an electroencephalogram-based depressive disorder detection algorithm is paid more and more attention by researchers.
The disadvantage of the prior art is that current diagnosis of clinical depressive disorders is based on physician interviews, which largely depend on the degree of expertise and subjectivity of the physician, and questionnaires, which may be fraudulent.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, and in order to realize the purpose, the electroencephalogram-based attention-seeking pool-based depressive disorder identification and classification method is adopted to solve the problems in the background technology.
An electroencephalogram-based method for identifying and classifying self-attention-seeking pooled depressive disorders, comprising:
acquiring a data set of an electroencephalogram signal, and preprocessing the data set;
constructing an adjacency matrix based on Euclidean distance and a left-right semi-brain connection relation, and calculating a node importance score based on an attention-seeking pooling operator to obtain a target node to obtain a training-optimized soft label classification model of the depressive disorder;
modeling a classification task and a regression task to obtain an output soft label and a regression score, and converting the predicted regression score into label probability distribution;
and finally obtaining joint label distribution for classification and clinical scores for regression according to the model.
As a further technical scheme: the specific steps of acquiring the data set of the electroencephalogram signal and preprocessing the data set comprise:
acquiring a data set sample of the electroencephalogram signal, and predefining;
presetting a training set with N samples
Figure BDA0003378153640000021
The ith sample is expressed as: oi＝(Xi，Ai，yi，ti，zi)，(i∈1，…，N)；
Wherein, Xi∈RE×TIs an electroencephalogram signal, E and T respectively represent the number of electrodes and the length of a time sequence; a. thei∈RE×EA adjacency matrix representing the ith sample; y isiAnd ziBinary labels and regression scores respectively representing the ith sample; t is tiIs a soft label for the specimen.
As a further technical scheme: the construction steps of the soft label comprise:
soft label for assisting in constructing sample according to PHQ-9 score
Figure BDA0003378153640000022
Wherein the content of the first and second substances,
Figure BDA0003378153640000023
while normalizing to [0,1 ] according to the fraction]In the meantime, the patient's depressive disorder probability is obtained as
Figure BDA0003378153640000024
And obtaining the probability that the patient is affiliated to the healthy person as
Figure BDA0003378153640000025
As a further technical scheme: the specific steps of constructing the adjacency matrix based on the Euclidean distance and the left and right semi-brain connection relation comprise:
according to the spatial relationship of the similarity feedback electrodes of the electroencephalogram signals, constructing a topological structure of the electroencephalogram signals based on Euclidean distance to obtain an adjacency matrix A belonging to the RE×EWherein E represents the number of electrodes, Ai(p, q) represents the connection relationship between the p channel and the q channel of the ith sample, and if the p channel is connected with the q channel, Ai(p, q) ═ 1, otherwise 0;
the generation formula of the topological structure of the electroencephalogram signal is as follows:
Figure BDA0003378153640000026
and (3) initializing the relevance among the brain electric channels of the samples:
Figure BDA0003378153640000027
wherein δ represents a calibration constant;
and finally, increasing global connection to improve the precision of the adjacent matrix according to the symmetrical connection relation of the left and right half brains.
As a further technical scheme: the specific steps of increasing global connection and improving the precision of the adjacency matrix according to the symmetrical connection relation of the left and right half-brains comprise:
introducing a new self-attention map pooling operator to carry out down-sampling on the electroencephalogram;
calculating the Manhattan distance between the target node and the reconstruction node according to the node with higher importance to obtain the importance score of the node, wherein the calculation formula is as follows:
Figure BDA0003378153640000031
wherein | · | purple sweet1And l represents respectively1The norm and the first layer of the graph neural network,
Figure BDA0003378153640000032
and
Figure BDA0003378153640000033
the feature matrix, the adjacency matrix, the identification matrix and the degree matrix of the ith sample in the neural network of the ith layer diagram.
As a further technical scheme: the specific steps of performing node pooling operation according to the obtained node importance scores include:
firstly, selecting an importance node to form a smaller graph, wherein the specific formula is as follows:
Figure BDA0003378153640000034
Figure BDA0003378153640000035
Figure BDA0003378153640000036
wherein the content of the first and second substances,
Figure BDA0003378153640000037
an operator is an index used to identify a more important node and return that node,
Figure BDA0003378153640000038
and
Figure BDA0003378153640000039
is to obtain a feature matrix and an adjacency matrix according to the indexAnd are sent into the neural network of the next layer of diagram together;
computing nodes from parameterized projection vectors c
Figure BDA00033781536400000310
And
Figure BDA00033781536400000311
the relationship between the nodes is calculated by the following formula:
Figure BDA00033781536400000312
Figure BDA00033781536400000313
Figure BDA00033781536400000314
where | represents tandem operation, c is a trainable parameter,
Figure BDA00033781536400000315
representing a directed connection relationship between node p and node q,
Figure BDA00033781536400000316
representing a non-directional node connection relationship between node p and node q,
Figure BDA00033781536400000317
is an undirected graph learned based on the previous layer adjacency matrix and the current layer;
and simultaneously comparing the similarity between different node pairs, and generating a sparse graph structure by using a sparsemax function:
Figure BDA00033781536400000318
wherein τ is a set threshold;
integrating source node information into a target node for reservation based on a self-attention mechanism, wherein the specific formula is as follows:
Figure BDA0003378153640000041
wherein the content of the first and second substances,
Figure BDA0003378153640000042
and
Figure BDA0003378153640000043
respectively representing a query matrix and a key matrix,
Figure BDA0003378153640000044
represents a self-attention matrix;
meanwhile, a predefined brain topological structure is introduced into an attention mechanism and is constrained by a non-negative proportion parameter epsilon, and the specific formula is as follows:
Figure BDA0003378153640000045
a vector representation of fixed size is generated using the Readout function and a concatenation of average pooling and maximum pooling is performed for each sub-graph, defined as:
Figure BDA0003378153640000046
wherein the content of the first and second substances,
Figure BDA0003378153640000047
is the output of the l +1 th layer, and the ith sample can be expressed as:
Figure BDA0003378153640000048
as a further technical scheme: the loss function comprises classification loss, regression loss and inconsistency loss;
the predicted label probability distribution is measured using the Kullback-Leibler divergence to minimize the difference
Figure BDA0003378153640000049
And a true soft label tiThe difference between the two is specifically represented by the following formula:
Figure BDA00033781536400000410
wherein D isKL(. I. represents Kullback-Leibler divergence);
based on the difference between the predicted regression score and the true regression score and using l1Loss was used as an evaluation index:
Figure BDA00033781536400000411
wherein, | · | represents an absolute operator;
scoring according to predicted PHQ-9
Figure BDA00033781536400000412
Conversion to soft labels
Figure BDA00033781536400000413
And using JS divergence in predicted tag distribution
Figure BDA00033781536400000414
And prediction-based regression scores
Figure BDA00033781536400000415
Converted label distribution
Figure BDA00033781536400000416
Applying inconsistent constraint between the label and the label, and obtaining the probability distribution of two labels measured by using JS divergence, wherein the specific formula is as follows:
Figure BDA0003378153640000051
the overall objective function is obtained as: l isloss＝Lc+λLr+γLd；
Wherein, λ and γ are weight parameters.
Compared with the prior art, the invention has the following technical effects:
by adopting the technical scheme, the depression disorder soft label classification model based on the electroencephalogram signals is constructed. Firstly, constructing an adjacency matrix based on Euclidean distance and left and right semi-brain connection relation. And selecting an importance node according to a scoring module, and then introducing a self-attention mechanism to integrate the information of the deleted node into a reserved node. The learned embeddings are further fed back into the network with the JS divergence module to explicitly model the classification and regression tasks to output soft tags and regression scores. By converting the predicted regression scores into a label probability distribution, the model can ultimately obtain a joint label distribution for classification and clinical scores for regression. Thereby solving the problem of difficult diagnosis of clinical depression disorder at present. A more professional, accurate and quick method for identifying depressive disorders is provided.
Drawings
The following detailed description of embodiments of the invention refers to the accompanying drawings in which:
FIG. 1 is a schematic diagram of the steps of a self-attention-seeking pooled depressive disorder identification and classification method disclosed herein;
fig. 2 is a schematic diagram of a left-right half brain electrical brain channel connection mode disclosed in the present application.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, in an embodiment of the present invention, a electroencephalogram-based method for identifying and classifying a depressive disorder in a self-attention-seeking pool includes:
s1, acquiring a data set of the electroencephalogram signals, and preprocessing, wherein the method specifically comprises the following steps:
s2, constructing an adjacency matrix based on Euclidean distance and left and right semi-brain connection relation, and meanwhile calculating node importance scores based on an attention-seeking pooling operator to obtain target nodes to obtain a training-optimized soft label classification model of the depressive disorder;
s3, modeling a classification task and a regression task to obtain an output soft label and a regression score, and converting the predicted regression score into label probability distribution;
and S4, finally obtaining the joint label distribution for classification and the clinical score for regression according to the model.
In a specific embodiment, symbol predefining is performed: given a training set
Figure BDA0003378153640000061
The data set has a total of N training samples. Each sample can be represented as: o isi＝(Xi，Ai，yi，ti，zi)，(i∈1，…，N)；
Wherein, Xi∈RE×TIs an electroencephalogram signal, E and T respectively represent the number of electrodes and the length of a time sequence; a. thei∈RE×EAn adjacency matrix representing the ith sample, which is used for describing the connection relation between the electroencephalogram electrodes; y isiAnd ziBinary labels and regression scores respectively representing the ith sample; t is tiIs a soft label of the sample, and the sum of the descriptive degrees of different classes for the same sample is 1, namely
Figure BDA0003378153640000062
As the GNN layers propagate, the node pooling operation causes a change in the number of nodes between layers. For this purpose,
Figure BDA0003378153640000063
represents the ith sample instance of the ith layer. The corresponding feature matrix and adjacency matrix are respectively expressed as
Figure BDA0003378153640000064
And
Figure BDA0003378153640000065
wherein ElAnd TlRespectively representing the number of nodes and the characteristic dimension of the ith layer. Xi(p, q) is the element of p rows and q columns of the ith sample.
Because the data set does not provide soft labels per se, the method utilizes the PHQ-9 score to assist in constructing the sample soft labels
Figure BDA0003378153640000066
In a specific embodiment, the range interval of PHQ-9 is set to [0, 27 ]]The threshold was set to 5 (sample scores greater than 5 were considered depressive disorder patients, otherwise healthy people). Normalize the score to [0,1]Thereby obtaining a patient's depressive disorder probability of
Figure BDA0003378153640000067
On the basis, the probability that the patient is affiliated to the healthy person is further obtained as
Figure BDA0003378153640000068
Since the above data set does not provide a predefined adjacency matrix, a euclidean distance-based adjacency matrix construction method is proposed.
The specific steps for constructing the adjacency matrix are as follows:
there is evidence that the three-dimensional physical distance between electrodes may reflect the connection relationship between the electrodes, which may be used to guide the generation of the adjacency matrix. However, existing depressive disorder electroencephalographic datasets do not provide electrode three-dimensional spatial coordinates. The similarity of the brain electrical signals can also reflect the spatial relationship of the electrodes. Thus, the euclidean distance of the brain electrical signal is calculated to indicate the generation of the topology of the brain. The specific formula is as follows:
Figure BDA0003378153640000069
wherein the adjacency matrix A ∈ RE×ERepresenting the topology of the brain electrical signal, E representing the number of electrodes, Ai(p, q) represents the connection relationship of the p-channel and the q-channel of the ith sample. If channel p and channel q are connected, then Ai(p, q) ═ 1, and vice versa 0.
Meanwhile, the relevance among the sample brain electrical channels is further initialized:
Figure BDA0003378153640000071
where δ represents a calibration constant.
Directly constructing the adjacency matrix by using euclidean distances will lead to the problem of introducing isolated nodes, i.e. euclidean distances between nodes and other nodes cannot meet the threshold requirement. If the threshold is lowered, many invalid connections will be introduced and the computational cost will be increased. To overcome this problem, when an orphaned node is encountered, the present invention selects the first three nodes closest to the orphaned node as its neighbors. However, adjacency matrices determined in this manner often connect adjacent brain electrical channels, ignoring global connections.
In order to better extract the topological structure of the electroencephalogram channel, the method increases global connection to improve the precision of the adjacency matrix. Studies have shown that neuronal activity between the hemispheres of the brain can provide additional information. As shown in fig. 2, in order to fully consider the brain symmetry information, some pairs of symmetrical left-right half brain connections are further selected to make the most of the brain symmetry information.
In electroencephalogram signal modeling, the invention introduces a new self-attention-diagram pooling operator to carry out down-sampling on electroencephalogram. The module is key to how to calculate the importance scores of the nodes, and in the specific implementation, if one node can participate in the reconstruction of many other nodes, the importance of the node is relatively high. For this purpose, the manhattan distance between the target node and the reconstruction node is further calculated. The formula for calculating the importance score is:
Figure BDA0003378153640000072
wherein | · | purple sweet1And l represents respectively1Norm and l-th layer of the graph neural network.
Figure BDA0003378153640000073
And
Figure BDA0003378153640000074
the feature matrix, the adjacency matrix, the identification matrix and the degree matrix of the ith sample in the neural network of the ith layer diagram. In this way, the importance scores of the nodes are obtained, and top-m nodes are further selected according to the importance scores. On the basis, node pooling operation is executed, and importance nodes are selected to form a smaller graph, wherein the specific formula is as follows:
Figure BDA0003378153640000075
Figure BDA0003378153640000076
Figure BDA0003378153640000077
wherein the content of the first and second substances,
Figure BDA0003378153640000078
an operator is an index used to identify more important nodes and return those nodes.
Figure BDA0003378153640000079
And
Figure BDA0003378153640000081
the feature matrix and the adjacency matrix are obtained according to the index, and are sent to the neural network of the next layer graph together.
The above operation brings about the following problems:
1) an isolated node: assume that there is a node of higher importance, however, the neighbor nodes of the node are of lower importance. In this case, the node selection mechanism will cause the target node to become an isolated node in subsequent figures, thereby making it inaccessible to other nodes. This will affect the integrity of the graph structure and hinder the propagation of information.
2) Information loss: for high-dimensional brain electrical signals, each electrode can provide node-specific information. That is, even for nodes of lower importance, it provides some unique information that other electrodes do not possess. A node delete operation means discarding this information. Therefore, how to ensure the connectivity of the graph and the integrity of the information is an urgent problem to be solved.
For the problem of the first isolated node, the invention encodes the subgraph
Figure BDA0003378153640000082
Potential relationships between each node pair in (a). In a specific embodiment, the nodes are first calculated using a parameterized projection vector c
Figure BDA0003378153640000083
And
Figure BDA0003378153640000084
attention weight of (1). Because the brain graph structure is undirected, the relationship between the nodes is calculated by the formula:
Figure BDA0003378153640000085
Figure BDA0003378153640000086
Figure BDA0003378153640000087
where | represents tandem operation, c is a trainable parameter,
Figure BDA0003378153640000088
representing a directed connection relationship between node p and node q,
Figure BDA0003378153640000089
representing a non-directional node connection relationship between node p and node q,
Figure BDA00033781536400000810
is an undirected graph learned based on the previous layer adjacency matrix and the current layer.
The method can ensure the integrity of the subgraph. Meanwhile, in order to compare similarities between different node pairs, the work of the prior art is usually normalized by using a softmax function, so that a dense full-connectivity graph is generated. But most connections are noisy and redundant, increasing computational cost and reducing graphics accuracy. Thus, here the sparse graph structure is generated using the sparsemax function:
Figure BDA00033781536400000811
where τ is the set threshold.
To address the second information loss problem, the present invention introduces a self-attention mechanism. All nodes of the current layer are regarded as source nodes, and the reserved importance nodes are regarded as target nodes. Thus, the self-attention mechanism will integrate source node information into the target node.
Simplifying the notation, provided that
Figure BDA0003378153640000091
Is a feature matrix of
Figure BDA0003378153640000092
The importance node of the layer is
Figure BDA0003378153640000093
The goal of the self-attention mechanism is to assign ElIntegration of individual node information into El+1The specific formula of each node is as follows:
Figure BDA0003378153640000094
wherein the content of the first and second substances,
Figure BDA0003378153640000095
and
Figure BDA0003378153640000096
respectively representing a query matrix and a key matrix,
Figure BDA0003378153640000097
representing a self-attention matrix. In order to preserve as much original graphical information as possible, a predefined brain topology is introduced into the self-attention mechanism, thereby enhancing the connectivity. To prevent the adjacency matrix values from being too large or too small for the self-attention matrix, we constrain the model using a non-negative scaling parameter ε, defined as:
Figure BDA0003378153640000098
while a Readout function is further employed to generate a fixed-size vector representation. The Readout function performs a concatenation of average pooling and maximum pooling for each sub-graph, defined as follows:
Figure BDA0003378153640000099
wherein
Figure BDA00033781536400000910
Is the output of layer l + 1. Finally, the ith sample can be expressed as:
Figure BDA00033781536400000911
the loss function comprises classification loss, regression loss and inconsistency loss;
first, the first loss is the supervised classification loss. To measure predicted label probability distribution
Figure BDA00033781536400000912
And a true soft label tiThe difference between them, we use the Kullback-Leibler divergence to minimize the difference, which is defined as follows:
Figure BDA00033781536400000913
wherein D isKL(. DELTA.. DELTA.) represents the Kullback-Leibler divergence.
The regression task calculates the difference between the predicted regression score and the true regression score and uses l1Loss was used as an evaluation index:
Figure BDA0003378153640000101
wherein, | · | represents an absolute operator;
in order to strengthen the link between the classification task and the regression task so that they can guide each other in an explicit way. In this example, the predicted PHQ-9 was further scored
Figure BDA0003378153640000102
Conversion to soft labels
Figure BDA0003378153640000103
And using JS divergence in predicted tag distribution
Figure BDA0003378153640000104
And prediction-based regression scores
Figure BDA0003378153640000105
Converted label distribution
Figure BDA0003378153640000106
Applying an inconsistency constraint therebetween. Thus, we can measure two tag probability distributions with JS divergence as follows:
Figure BDA0003378153640000107
the overall objective function is obtained as:
Lloss＝Lc+λLr+γLd；
wherein, λ and γ are weight parameters.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents, which should be construed as being within the scope of the invention.

Claims (7)

1. An electroencephalogram-based method for identifying and classifying self-attention-seeking pooled depressive disorders, comprising the following steps:
acquiring a data set of an electroencephalogram signal, and preprocessing the data set;
constructing an adjacency matrix based on Euclidean distance and a left-right semi-brain connection relation, and calculating a node importance score based on an attention-seeking pooling operator to obtain a target node to obtain a training-optimized soft label classification model of the depressive disorder;
modeling a classification task and a regression task to obtain an output soft label and a regression score, and converting the predicted regression score into label probability distribution;
and finally obtaining joint label distribution for classification and clinical scores for regression according to the model.
2. The electroencephalogram-based method for identifying and classifying depressive disorder in self-attention-map pooling according to claim 1, wherein the specific steps of acquiring a data set of electroencephalogram signals and preprocessing the data set comprise:
acquiring a data set sample of the electroencephalogram signal, and predefining;
presetting a training set with N samples
Figure FDA0003378153630000011
The ith sample is expressed as: o isi＝(Xi，Ai，yi，ti，zi)，(i∈1,…，N)；
Wherein, Xi∈RE×TIs an electroencephalogram signal, E and T respectively represent the number of electrodes and the length of a time sequence; a. thei∈RE×EA adjacency matrix representing the ith sample; y isiAnd ziBinary labels and regression scores respectively representing the ith sample; t is tiIs a soft label for the specimen.
3. The electroencephalogram-based self-attention-seeking pooling depressive disorder identification and classification method according to claim 2, wherein the soft label construction step comprises:
soft label for assisting in constructing sample according to PHQ-9 score
Figure FDA0003378153630000012
Wherein the content of the first and second substances,
Figure FDA0003378153630000013
while normalizing to [0,1 ] according to the fraction]Get ill betweenThe probability of depressive disorder in the subject is
Figure FDA0003378153630000014
And obtaining the probability that the patient is affiliated to the healthy person as
Figure FDA0003378153630000015
4. The electroencephalogram-based method for identifying and classifying depressive disorders based on self-attention-map pooling according to claim 1, wherein the specific steps of constructing the adjacency matrix based on Euclidean distance and left and right semi-brain connection relations comprise:
according to the spatial relationship of the similarity feedback electrodes of the electroencephalogram signals, constructing a topological structure of the electroencephalogram signals based on Euclidean distance to obtain an adjacency matrix A belonging to the RE×EWherein E represents the number of electrodes, Ai(p, q) represents the connection relationship between the p channel and the q channel of the ith sample, and if the p channel is connected with the q channel, Ai(p, q) ═ 1, otherwise 0;
the generation formula of the topological structure of the electroencephalogram signal is as follows:
Figure FDA0003378153630000021
and (3) initializing the relevance among the brain electric channels of the samples:
Figure FDA0003378153630000022
wherein δ represents a calibration constant;
and finally, increasing global connection to improve the precision of the adjacent matrix according to the symmetrical connection relation of the left and right half brains.
5. The electroencephalogram-based method for identifying and classifying depressive disorders based on self-attention-seeking pooling, according to the left and right semi-brain symmetric connection relation, wherein the specific steps of increasing global connection and improving accuracy of adjacency matrix include:
introducing a new self-attention map pooling operator to carry out down-sampling on the electroencephalogram;
calculating the Manhattan distance between the target node and the reconstruction node according to the node with higher importance to obtain the importance score of the node, wherein the calculation formula is as follows:
Figure FDA0003378153630000023
wherein |1And l represents respectively1The norm and the first layer of the graph neural network,
Figure FDA0003378153630000024
and
Figure FDA0003378153630000025
the feature matrix, the adjacency matrix, the identification matrix and the degree matrix of the ith sample in the neural network of the ith layer diagram.
6. The electroencephalogram-based self-attention-chart pooling depressive disorder identification and classification method according to claim 1 or 5, wherein the specific steps of performing node pooling operation according to the obtained node importance scores comprise:
firstly, selecting an importance node to form a smaller graph, wherein the specific formula is as follows:
Figure FDA0003378153630000026
Figure FDA0003378153630000027
Figure FDA0003378153630000028
wherein the content of the first and second substances,
Figure FDA0003378153630000029
an operator is an index used to identify a more important node and return that node,
Figure FDA00033781536300000210
and
Figure FDA00033781536300000211
obtaining a characteristic matrix and an adjacent matrix according to the index, and sending the characteristic matrix and the adjacent matrix into a neural network of a next layer of graph together;
computing nodes from parameterized projection vectors c
Figure FDA00033781536300000212
And
Figure FDA00033781536300000213
the relationship between the nodes is calculated by the following formula:
Figure FDA0003378153630000031
Figure FDA0003378153630000032
Figure FDA0003378153630000033
where | represents tandem operation, c is a trainable parameter,
Figure FDA0003378153630000034
representing a directed connection between node p and node qThe relationship is such that,
Figure FDA0003378153630000035
representing a non-directional node connection relationship between node p and node q,
Figure FDA0003378153630000036
is an undirected graph learned based on the previous layer adjacency matrix and the current layer;
and simultaneously comparing the similarity between different node pairs, and generating a sparse graph structure by using a sparsemax function:
Figure FDA0003378153630000037
wherein τ is a set threshold;
integrating source node information into a target node for reservation based on a self-attention mechanism, wherein the specific formula is as follows:
Figure FDA0003378153630000038
wherein the content of the first and second substances,
Figure FDA0003378153630000039
and
Figure FDA00033781536300000310
respectively representing a query matrix and a key matrix,
Figure FDA00033781536300000311
represents a self-attention matrix;
meanwhile, a predefined brain topological structure is introduced into an attention mechanism and is constrained by a non-negative proportion parameter epsilon, and the specific formula is as follows:
Figure FDA00033781536300000312
a vector representation of fixed size is generated using the Readout function and a concatenation of average pooling and maximum pooling is performed for each sub-graph, defined as:
Figure FDA00033781536300000313
wherein the content of the first and second substances,
Figure FDA00033781536300000314
is the output of the l +1 th layer, and the ith sample can be expressed as:
Figure FDA00033781536300000315
7. the electroencephalogram-based method for identifying and classifying depressive disorders based on self-attention-map pooling according to claim 6, wherein the loss functions include classification loss, regression loss and inconsistent loss;
the predicted label probability distribution is measured using the Kullback-Leibler divergence to minimize the difference
Figure FDA0003378153630000049
And a true soft label tiThe difference between the two is specifically represented by the following formula:
Figure FDA0003378153630000041
wherein D isKL(. I. represents Kullback-Leibler divergence);
based on the difference between the predicted regression score and the true regression score and using l1Loss was used as an evaluation index:
Figure FDA0003378153630000042
wherein, | · | represents an absolute operator;
scoring according to predicted PHQ-9
Figure FDA0003378153630000043
Conversion to soft labels
Figure FDA0003378153630000044
And using JS divergence in predicted tag distribution
Figure FDA0003378153630000045
And prediction-based regression scores
Figure FDA0003378153630000046
Converted label distribution
Figure FDA0003378153630000047
Applying inconsistent constraint between the label and the label, and obtaining the probability distribution of two labels measured by using JS divergence, wherein the specific formula is as follows:
Figure FDA0003378153630000048
the overall objective function is obtained as: l isloss＝Lc+λLr+γLd；
Wherein, λ and γ are weight parameters.
CN202111425371.0A 2021-11-26 2021-11-26 Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method Pending CN114219014A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111425371.0A CN114219014A (en) 2021-11-26 2021-11-26 Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111425371.0A CN114219014A (en) 2021-11-26 2021-11-26 Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method

Publications (1)

Publication Number Publication Date
CN114219014A true CN114219014A (en) 2022-03-22

Family

ID=80698656

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111425371.0A Pending CN114219014A (en) 2021-11-26 2021-11-26 Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method

Country Status (1)

Country Link
CN (1) CN114219014A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114764858A (en) * 2022-06-15 2022-07-19 深圳大学 Copy-paste image recognition method, device, computer device and storage medium
CN117497140A (en) * 2023-10-09 2024-02-02 合肥工业大学 Multi-level depression state detection method based on fine granularity prompt learning
CN117497140B (en) * 2023-10-09 2024-05-31 合肥工业大学 Multi-level depression state detection method based on fine granularity prompt learning

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114764858A (en) * 2022-06-15 2022-07-19 深圳大学 Copy-paste image recognition method, device, computer device and storage medium
CN114764858B (en) * 2022-06-15 2022-11-01 深圳大学 Copy-paste image identification method and device, computer equipment and storage medium
CN117497140A (en) * 2023-10-09 2024-02-02 合肥工业大学 Multi-level depression state detection method based on fine granularity prompt learning
CN117497140B (en) * 2023-10-09 2024-05-31 合肥工业大学 Multi-level depression state detection method based on fine granularity prompt learning

Similar Documents

Publication Publication Date Title
CN115249539B (en) Multi-mode small sample depression prediction model construction method
CN111611367B (en) Visual question-answering method introducing external knowledge
CN113673244B (en) Medical text processing method, medical text processing device, computer equipment and storage medium
CN112256866A (en) Text fine-grained emotion analysis method based on deep learning
Zhu et al. Deep multi-modal discriminative and interpretability network for alzheimer’s disease diagnosis
Ma et al. Echo memory-augmented network for time series classification
Jeyaraj et al. Classifier feature fusion using deep learning model for non-invasive detection of oral cancer from hyperspectral image
CN114781382A (en) Medical named entity recognition system and method based on RWLSTM model fusion
CN116541472A (en) Knowledge graph construction method in medical field
CN116072265A (en) Sleep stage analysis system and method based on convolution of time self-attention and dynamic diagram
CN116110565A (en) Method for auxiliary detection of crowd depression state based on multi-modal deep neural network
CN109935321B (en) Risk prediction system for converting depression patient into bipolar affective disorder based on functional nuclear magnetic resonance image data
Nasrullah et al. [Retracted] Detection of Types of Mental Illness through the Social Network Using Ensembled Deep Learning Model
CN114219014A (en) Electroencephalogram-based attention-seeking pooling depressive disorder identification and classification method
Lyu et al. Multiscale echo self-attention memory network for multivariate time series classification
Lin et al. Learning patient similarity via heterogeneous medical knowledge graph embedding
CN117036760A (en) Multi-view clustering model implementation method based on graph comparison learning
Liu et al. Multi-type decision fusion network for visual Q&A
CN113836398A (en) Patent transaction recommendation method based on attribute heterogeneous network representation learning
CN115700826A (en) Receipt processing method, receipt display method, receipt processing device, receipt display device, computer equipment and storage medium
Feng et al. Spatial-attention and demographic-augmented generative adversarial imputation network for population health data reconstruction
Ma et al. Modeling Multivariate Time Series via Prototype Learning: a Multi-Level Attention-based Perspective
Zhang et al. Prediction of behavior and activity of patients through wearable devices based on deep learning
CN116936108B (en) Unbalanced data-oriented disease prediction system
CN116665906B (en) Resting state functional magnetic resonance brain age prediction method based on similarity twin network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination