CN114881089A

CN114881089A - Depression electroencephalogram classification method based on double-branch fusion model

Info

Publication number: CN114881089A
Application number: CN202210622308.4A
Authority: CN
Inventors: 杨淳沨; 苏天; 孔佑勇; 陈阳; 舒华忠
Original assignee: Southeast University
Current assignee: Southeast University
Priority date: 2022-06-02
Filing date: 2022-06-02
Publication date: 2022-08-09

Abstract

The invention discloses a deep learning depression electroencephalogram classification method based on a double-branch fusion model, which comprises the following steps of: (1) acquiring electroencephalograms of electrodes of the prefrontal brain lobes Fp1, Fpz and Fp2 of a plurality of groups of healthy people, (2) acquiring electroencephalograms of electrodes of the prefrontal brain lobes Fp1, Fpz and Fp2 of a plurality of groups of patients with mild depression, (3) acquiring electroencephalograms of electrodes of the prefrontal brain lobes Fp1, Fpz and Fp2 of a plurality of groups of patients with moderate depression, (4) performing training and learning on the double-branch fusion model in the input forms of the healthy control, the patients with mild depression and the patients with moderate depression in the steps (1), (2) and (3), (5) converting the electroencephalograms to be analyzed into corresponding wavelet time-frequency diagrams, and inputting the double-branch fusion model trained in the step (4) to complete analysis of the electroencephalograms. The method has good effect, and can distinguish depression, health and depression degree.

Description

Depression electroencephalogram classification method based on double-branch fusion model

Technical Field

The invention relates to a depression electroencephalogram classification method based on a double-branch fusion model, and belongs to the technical field of computer application.

Background

Depression (MDD) is a common mental Disorder disease, and its clinical symptoms are low interest in all things, poor self-recognition and attention deficit, and even repeated self-mutilation and suicide. According to the World Health Organization (WHO), it is estimated that the number of patients with depression will exceed the sum of all patients with cardiovascular disease by 2030, and depression will become the World's first leading cause of disability. In china, the number of attacks of depression accounts for about 4.2% of the general population, and shows a tendency to rise and become younger year by year. Depression can cause not only serious injury to individuals, but also negative impact on the families and society of patients. If the patient can be correctly diagnosed in the early stage of depression, the condition of the patient can be obviously improved in time by means of psychotherapy, medication, electroshock therapy, lifestyle change and the like. Therefore, the method has great significance for the accurate rate diagnosis of the depression.

On one hand, the electroencephalogram signals can capture the electrical activity of the neurons at the brain millisecond level, and the method has high time resolution and reliability and is relatively low in price. On the other hand, studies have shown that the degree of depression in patients is mainly related to the activity of the prefrontal electroencephalogram and the symmetry of the EEG between the left and right prefrontal lobes. Therefore, the brain prefrontal three-channel electroencephalogram data can be selected as the basis for depression diagnosis.

Disclosure of Invention

In order to accurately distinguish health, depression and depression degree, the invention provides a depression electroencephalogram classification method based on double-branch fusion. The method takes a wavelet time-frequency diagram of frontal lobe three-channel electroencephalogram data and an original electroencephalogram sequence as input to train a two-branch fusion model so as to distinguish healthy electroencephalograms and moderately depressed electroencephalograms from mild electroencephalograms and moderately electroencephalograms.

The invention adopts the following technical scheme for solving the technical problems:

the invention provides a deep learning-based depression electroencephalogram classification method, which comprises the following steps:

(1) acquiring electroencephalogram signals of the prefrontal brain lobes Fp1, Fpz and Fp2 electrodes of a plurality of groups of healthy people, segmenting into window data by using a sliding window, converting into a wavelet time-frequency graph by using wavelet transform, and finally taking the wavelet time-frequency graphs of three channels and corresponding original window sequences as the input of a model.

(2) Electroencephalogram signals of the Fp1, Fpz and Fp2 electrodes of the brain prefrontal lobes of a plurality of groups of patients with mild depression are obtained, the electroencephalogram signals are cut into window data by using a sliding window, then wavelet transformation is used for converting the window data into wavelet time-frequency graphs, and finally the wavelet time-frequency graphs of three channels and corresponding original window sequences are used as input of a model.

(3) Electroencephalogram signals of the Fp1, Fpz and Fp2 electrodes of the brain prefrontal lobes of a plurality of groups of patients with moderate depression are obtained, the electroencephalogram signals are cut into window data by using a sliding window, then wavelet transformation is used for converting the window data into a wavelet time-frequency graph, and finally the wavelet time-frequency graphs of three channels and corresponding original window sequences are used as input of a model.

(4) And (3) training and learning the double-branch fusion model in the input forms of the healthy control, the mild depression patient and the moderate depression patient in the steps (1), (2) and (3).

(5) And (4) converting the electroencephalogram signals of the window to be analyzed into corresponding wavelet time-frequency graphs, inputting the double-branch fusion model trained in the step (4), and completing the analysis of the electroencephalogram signals.

As a further technical scheme of the invention, the brain prefrontal lobe acquisition uses a universal three-lead brain electrical acquisition system.

As a further aspect of the present invention, the sliding window length of (1), (2) and (3) is 2 seconds, and the overlap ratio is 0.

As a further technical solution of the present invention, the wavelet transform used is a wavelet transform based on a complex Morlet wavelet (with a wideband parameter of 3 and a center wavelength of 3).

As a further technical scheme of the invention, the double-branch fusion model is a deep learning classification model built on the basis of a convolutional neural network and a channel attention mechanism.

Has the advantages that: compared with the prior art, the invention adopting the technical scheme has the following technical effects: the dual-branch fusion model provided by the invention can effectively extract time-frequency characteristics of electroencephalogram signals, and effective information of electroencephalogram is extracted from two modes by using a convolutional neural network, so that the accuracy, precision, recall rate and F1 score in distinguishing health and moderate depression, health and mild depression, mild depression and moderate depression and health and mild depression and moderate depression are higher than those of the existing depression electroencephalogram method.

Drawings

FIG. 1 is a flow chart of a disclosed method of the present invention;

FIG. 2 is a two-branch fusion model designed according to the present invention;

fig. 3 is a schematic diagram of a selected prefrontal lobe three-channel electroencephalogram location in the present invention.

Detailed Description

The invention is further elucidated with reference to the drawings and the detailed description.

Example (b): as shown in fig. 1, the flowchart of the method for classifying the depression electroencephalogram based on deep learning disclosed by the present invention specifically includes the following steps:

According to the relation between the prefrontal lobe of the brain and the depression, the electroencephalograms of Fp1, Fpz and Fp2 of the prefrontal lobe are selected as data sources, and a pervasive three-lead electroencephalogram acquisition system is used as an acquisition tool. The specific collection steps are as follows: I. measuring the distance from a nasal root point (an eyebrow center) to the occipital tuberosity of the hindbrain, recording as one tenth of the distance from the eyebrow center to the top of the head, and recording as the position of Fpz; II. Measuring the distance between mastoids passing through the tops of the mastoids on two sides, taking one tenth of the distance from the mastoid on the left ear to the top of the mastoid, and recording as a point; and III, taking one tenth of the occipital bone towards the vertex, and taking the positions Fp1 and Fp2 which pass through the upper left and right tenths respectively.

The method comprises the following steps of preprocessing the acquired electroencephalogram data to obtain cleaner electroencephalogram data: firstly, a 0.5Hz to 100Hz Butterworth band-pass filter is adopted to obtain the effective frequency of the brain electricity, then an ICA algorithm is used for removing artifacts, a 50Hz notch filter is used for removing power frequency noise, and finally min-max data normalization is adopted to prevent the influence caused by dimension and improve the convergence rate of the model.

Secondly, the electroencephalogram data are converted into an input form suitable for a double-branch fusion model by the following steps: firstly, a sliding window is used for intercepting an electroencephalogram signal: the acquired original brain electricity is intercepted by using a sliding window with the length of 2 seconds and the overlapping rate of 0. Then, it becomes a wavelet coefficient matrix by wavelet transform with a complex Morlet wavelet (with a broadband parameter of 3 and a center wavelength of 3) as a base wavelet. Finally, taking the modulus of the complex Morlet wavelet coefficient as the pixel value of the picture and sampling the size of the picture to 224x 224.

And (4-1) the double-branch fusion model consists of a picture feature extraction branch and a time sequence feature extraction branch.

(4-2) time-series feature extraction branch in the two-branch fusion model embeds an attention mechanism based on discrete Fourier transform coefficients in the FCN model. The FCN branch is composed of three basic blocks (Conv1d + BN + ReLU), of which the convolution output channels of the three basic blocks are 128,256,128, respectively. An attention mechanism based on discrete fourier transform coefficients is the spreading of the SE model over the DFT frequency domain.

Example 2:

step (1) (2) (3): collecting the electroencephalogram of a subject, dividing the depression degree according to a depression scale, and then preprocessing the electroencephalogram data. In this example, brain electrical signals of prefrontal lobes Fp1, Fpz and Fp2 of subjects were collected using a generalized three-lead brain electrical acquisition system, the sampling frequency f was 1000Hz, and each subject was grouped by at least two specialized psychiatrists via the Hamilton Depression Scale (HDRS-17) into 20 healthy persons, 34 mildly depressed patients and 29 moderately depressed patients. The three-lead electrical signal of each subject was then sliced using a sliding window of length 2s with an overlap ratio of 0 and subjected to a wavelet transform with a complex Morlet wavelet (with a wideband parameter of 3, a center wavelength of 3) as the basis wavelet into a matrix of wavelet coefficients, followed by taking the modulus of the complex Morlet wavelet coefficients as the pixel values of the picture and sampling the size of the picture to 224x224 size. And finally, forming an input tensor by the three wavelet time-frequency graphs and the corresponding original electroencephalogram sequence.

And (4) constructing a dual-branch fusion model and inputting the input tensor into the dual-branch fusion model for training. In the present embodiment, the parameters of the two-branch fusion model are as follows. Conv below the picture feature extraction branch represents a two-dimensional convolution layer, and Conv of the feature extraction branch represents a one-dimensional convolution layer; BN denotes batch normalization, ReLU denotes the ReLU layer, and MaxPool denotes the maximum pooling layer.

1) The feature extraction module 1: the discarding rate of the one-dimensional convolution layer, the batch normalization layer, the ReLU layer, the Dropout layer and the DFTC-ATT layer is 0.3;

2) the feature extraction module 2: the structure is completely the same as that of the feature extraction module 1; parametrically, the size of the convolution kernel becomes 5, and the number of channels input and output by the convolution layer becomes 128 and 256.

3) The feature extraction module 3: structurally, a comparison feature extraction module 1 replaces DFTC-ATT with GAP; parametrically, the size of the convolution kernel becomes 3, and the number of channels input and output by the convolution layer becomes 256 and 128.

And finally, splicing the feature vectors learned by the two branches into a final feature vector (with the length of 192), and inputting a classification result through a classification module.

And (5) evaluating the learning effect of the model by using the accuracy, the precision, the recall rate and the F1 score.

In the invention, note that Non-De is a healthy subject, Mid-De is a mild depression patient, Con-De is a moderate depression patient, TN (True Negative) is a Negative sample predicted as a Negative by a model, FP (False Positive) is a Negative sample predicted as a Positive by a model, and FN (False Negative) is a Positive sample predicted as a Negative by a model, and then accuracy (accuracycacy) is defined as the probability of correctly classifying all samples:

precision (precision) can be divided into the precision of positive class samples and the precision of negative class samples, the precision of positive class samples is the ratio of the positive class samples to the positive class samples:

the accuracy of the negative class samples is the ratio of the actual negative class in the samples predicted to be the negative class:

the recall ratio (recall) can also be divided into the recall ratio of the positive type samples and the recall ratio of the negative type samples, wherein the recall ratio of the positive type samples is the proportion determined as positive type in the samples actually as positive type:

the recall rate of the negative class sample is the proportion determined as the negative class in the sample actually being the negative class:

the F1 value comprehensively considers the precision rate and the recall rate, is a harmonic mean of the precision rate and the recall rate and is often used as a final evaluation method of a machine learning classification method, and the higher the F1 value of each class is, the better the classification result is. The F1 values under each category are expressed as:

in the present invention, different classification models (ResNet, FCN, MLSTM-FCN, CNN-LSTM) were used for comparison, as shown in tables 1 to 4.

TABLE 1 comparison of results of different classification methods under Non-De and Con-De tasks

TABLE 2 comparison of results of different classification methods under Non-De and Mid-De tasks

TABLE 3 comparison of results of different classification methods under Mid-De and Con-De tasks

TABLE 5 comparison of results of different classification methods under Non-De, Mid-De and Con-De tasks

From the experimental results of tables 1 to 5, it can be seen that the classification effect of the present invention is significantly better than that of other models.

The above description is only an embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can understand that the modifications or substitutions within the technical scope of the present invention are included in the scope of the present invention, and therefore, the scope of the present invention should be subject to the protection scope of the claims.

Claims

1. The deep learning depression electroencephalogram classification method based on the double-branch fusion model is characterized by comprising the following steps of:

(1) acquiring electroencephalogram signals of the prefrontal brain lobes Fp1, Fpz and Fp2 electrodes of a plurality of groups of healthy people, segmenting the electroencephalogram signals into window data by using a sliding window, converting the window data into wavelet time-frequency graphs by using wavelet transformation, and finally taking the wavelet time-frequency graphs of three channels and corresponding original window sequences as input of a model;

(2) acquiring electroencephalogram signals of prefrontal brain lobes Fp1, Fpz and Fp2 electrodes of a plurality of groups of patients with mild depression, segmenting into window data by using a sliding window, converting into a wavelet time-frequency diagram by using wavelet transform, and finally taking the wavelet time-frequency diagrams of three channels and corresponding original window sequences as input of a model;

(3) acquiring electroencephalogram signals of prefrontal brain lobes Fp1, Fpz and Fp2 electrodes of a plurality of groups of patients with moderate depression, segmenting into window data by using a sliding window, converting into a wavelet time-frequency diagram by using wavelet transform, and finally taking the wavelet time-frequency diagrams of three channels and corresponding original window sequences as input of a model;

(4) training and learning the double-branch fusion model in the input forms of the healthy control, the mild depression patient and the moderate depression patient in the steps (1), (2) and (3),

2. The deep learning-based depression electroencephalogram classification method according to claim 1, wherein the sliding window is set to have a length of 2 seconds and an overlap rate of 0.

3. The deep learning-based depression electroencephalogram classification method according to claim 1, characterized in that the algorithm of selecting wavelet transform is wavelet transform using complex Morlet wavelet (broadband parameter is 3, central wavelength is 3) as base wavelet.

4. The deep learning-based depression electroencephalogram classification method according to claim 1, wherein the two-branch fusion model is a deep learning classification model built on the basis of a convolutional neural network and a channel attention mechanism.

5. The deep learning-based depression electroencephalogram classification method according to claim 1, wherein the step (4) is specifically as follows:

(4-1) the two-branch fusion model is composed of a picture feature extraction branch and a time sequence feature extraction branch,

(4-2) the timing feature extraction branch in the two-branch fusion model embeds an attention mechanism based on discrete fourier transform coefficients in an FCN model, wherein the FCN branch is composed of three basic blocks (Conv1d + BN + ReLU), convolution output channels of the three basic blocks are 128,256,128 respectively, and the attention mechanism based on the discrete fourier transform coefficients is an extension of an SE model on a DFT frequency domain.