CN113768474B - Anesthesia depth monitoring method and system based on graph convolution neural network - Google Patents

Abstract

The invention discloses a method and system for monitoring depth of anesthesia based on a graph convolutional neural network. The system comprises: a data preprocessing module: used for preprocessing cerebral cortex electrogram signals; a functional network building module: used for calculating The phase lag index PLI of the sample data, each sample calculates an adjacency matrix to obtain the network topology map samples of different anesthesia stages; the graph conversion module: used to convert the dual graph, and convert the graph samples into a weighted graph constructed based on the phase lag index , and construct a new graph for node features at the same time; two-stream graph convolutional neural network module: used to store two models of two-stream graph convolutional neural network, model 1 is used to extract edge weight information, model 2 is used to extract node feature information, The prediction results are obtained by summing the predicted probability classifications of the output of the two models. The invention finds a new feature for distinguishing different anesthesia states, and the classification accuracy of awake state, moderate anesthesia state and deep anesthesia state reaches 95.4%, and can monitor different states of anesthesia well.

Description

A method and system for anesthesia depth monitoring based on graph convolutional neural network

technical field

The invention relates to the fields of biomedical signal processing technology and deep learning technology, in particular to a method and system for monitoring depth of anesthesia based on a graph convolutional neural network.

Background technique

In general anesthesia surgery, the anesthesiologist needs to monitor the patient's anesthesia status in real time. Anesthesia monitors can assist anesthesiologists in grasping the depth of anesthesia in patients and avoid accidental intraoperative awareness. If the anesthesia is too deep, it may make it difficult for the patient to wake up after surgery, and even have adverse sequelae to the nervous system; if the anesthesia is too shallow, it may cause the patient to wake up during the operation, leaving a psychological shadow on the patient. Therefore, it is very important to monitor the depth of anesthesia in real time for patients during surgery.

Currently, clinical techniques commonly used to monitor the depth of anesthesia include EEG bispectral index analysis, auditory evoked potential (AEP), and anesthesia entropy. These techniques are all used to monitor the depth of anesthesia by processing EEG signals, which record the surface signals of the sulcus gyrus of the brain, and have the advantages of non-invasive, non-invasive, and easy-to-obtain. The current popular anesthesia depth monitoring technology still has some defects. For example, BIS is not effective for isoflurane-induced anesthesia, and there are great differences in different people, and its algorithm is not public. Therefore, it is necessary to explore a more stable anesthesia depth monitoring algorithm.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, construct a brain function network based on EEG signals, and provide a method and system for monitoring depth of anesthesia based on a graph convolutional neural network in combination with a dual-stream graph convolutional neural network.

In order to achieve the above purpose, a method for monitoring the depth of anesthesia based on a graph convolutional neural network designed by the present invention is special in that the method comprises the following steps:

1) Collect several channels of EEG signals, and preprocess the original signals;

2) Intercept several time segments of different anesthesia stages as data samples, calculate the phase lag index PLI, calculate an adjacency matrix for each sample, and obtain the network topology map samples of different anesthesia stages, the anesthesia stages include awake stage, moderate anesthesia stage, deep anesthesia stage;

3) Convert the adjacency matrix of the network topology graph into a dual graph, the edge connections of the dual graphs obtained by the conversion are all the same, and the edge weight information is retained on the node features of the dual graph, and the graph samples before conversion are based on the phase lag index. The constructed weighted graph, while retaining the node features, and constructing a fully connected matrix to represent the topology of these node features;

4) Constructing a two-stream graph convolutional neural network, dividing the graph samples into two-stream graph data, and finding two common adjacency matrices of the two-stream graph data; in the two-stream graph data, the first-stream graph data is based on phase lag The weighted graph constructed by the index, and the other flow graph data is a fully connected graph that retains the characteristics of the original node;

5) Input the two-stream graph data into the two models of the two-stream graph convolutional neural network respectively, perform graph coarsening and fast pooling, reduce the data dimension, aggregate similar nodes, and output the predicted value of each anesthesia stage through the fully connected layer;

6) The predicted values of different anesthesia stages output by the two models of the dual-stream graph convolutional neural network are added respectively, and the category with the largest predicted value is output as the prediction result of the anesthesia stage.

Preferably, the step 1) EMG signal of the midbrain cortex is a 16-channel ECoG signal of the frontal lobe-parietal lobe of the subject's brain; the preprocessing includes 0.1-100 Hz filtering, 50 Hz sag filtering and 200 Hz resampling.

Preferably, the calculation method of the phase lag index PLI is:

和

，使用希尔伯特变换计算瞬时相位

： Let the signal sequence of the two channels be

and

, using the Hilbert transform to calculate the instantaneous phase

:

表示

的希尔伯特(Hilbert)变换，i=1或2，j为虚数符号： in,

express

The Hilbert transform of , i = 1 or 2, and j is the imaginary sign:

In the formula, PV represents the principal value of Cauchy, t is the time, and τ is the integral variable;

The relative lock between the two channels is calculated as:

In the formula, z ₂ *( t ) is the complex conjugate of z ₂ ( t );

Then the PLI value is calculated as follows:

The range of PLI is between 0 and 1, with 0 indicating no phase locking between the two channels and 1 indicating perfect phase coupling between the two channels.

Preferably, in the step 4), the dual-stream graph convolutional neural network is constructed using the spectral domain graph convolution method GCN, the graph convolution is extended to the frequency domain of the graph through Fourier transformation, and the signal is filtered by using a filter. .

Preferably, in the step 5), both models of the dual-stream graph convolutional neural network adopt the spectral graph convolution method in which the convolution kernel is approximated by Chebyshev polynomials, and the graph coarsening and fast graph based on the Graclus multi-level clustering algorithm pooling method.

Preferably, the Graclus multi-level clustering algorithm adopts a greedy algorithm to measure the continuous roughness of the graph, so as to minimize the spectral clustering objective.

Preferably, in the step 2), the average absolute value of the signal amplitude of each channel of each time segment is calculated as the node feature.

Preferably, the data samples in the step 2) are randomly divided into a training set, a validation set, and a test set according to the ratio of 8:1:1, and the training set is used to train the neural network model of the graph and the validation set to adjust the hyperparameters of the model and adjust the model's hyperparameters. The ability of the model is initially evaluated, and the test set evaluates the generalization ability of the final model.

The present invention also provides an anesthesia depth monitoring system based on a graph convolutional neural network, the system includes a data preprocessing module, a functional network building module, a graph conversion module and a dual-stream graph convolutional neural network module;

The data preprocessing module: used to preprocess the cerebral cortex signal;

The functional network building module is used to cut the sample data into several time segments of different anesthesia stages, calculate the phase lag index PLI, calculate an adjacency matrix for each sample, and obtain the network topology map samples of different anesthesia stages. Stages include awake stage, moderate anesthesia stage, and deep anesthesia stage;

The graph conversion module is used to convert the adjacency matrix of the network topology graph sample into a dual graph. The edge connections of the converted dual graphs are the same, and the edge weight information is retained on the node features of the dual graph. The sample is a weighted graph constructed based on the phase lag index. In addition, the original node features are retained and used to construct a new fully connected graph;

The two-stream graph convolutional neural network module: used to store two models of the two-stream graph convolutional neural network, model 1 is used to extract edge weight information, model 2 is used to extract node feature information, and the two models are output prediction probability. The classification is added to obtain the prediction result.

The present invention further provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, the above-mentioned method for monitoring anesthesia depth based on a graph convolutional neural network is implemented.

The beneficial effects of the present invention include:

1) The traditional spectral graph convolution method can achieve excellent graph classification benefits on the basis of an adjacency matrix. It is sensitive to node feature information, but it cannot handle graphs with different network topologies, while the present invention adopts the dual graph conversion method. , the dual graph transformation cleverly converts the weighted graph obtained by the phase lag index calculation in this application into a graph of the same adjacency matrix, and more importantly, converts the edge weight features that are not sensitive to the graph convolution into node features. In this way, the traditional spectral convolution method, which is very efficient, can be directly used.

2) The graph itself includes its edge weights and its node features. The edge weights of the graph are converted into a dual graph and become the node feature input of the first model. Then, for the node characteristics of the graph, the present invention proposes a second graph convolution model. After the edge weights of the graph are extracted, the original node information is retained, and the relationship between the remaining nodes is considered equal. , so for the second graph convolution model, the present invention uses a fully connected matrix without self-connection as its adjacency matrix.

3) The present invention designs a dual-stream graph convolutional neural network structure, both of which use the graph classification method of spectral graph convolution. One stream is used to extract edge weight information, and the other stream is used to extract node feature information. After the two models are trained, , sum the predicted probabilities to predict the test set;

4) The present invention finds a new feature to distinguish different anesthesia states, and combines the brain network and the graph convolutional neural network to monitor the depth of anesthesia. The different states of anesthesia are well monitored, which provides a new method for clinical anesthesia monitoring.

The idea proposed by the present invention is not only applicable to this anesthesia data, but also applicable to EEG signal classification in other scenarios.

Description of drawings

FIG. 1 is a structural block diagram of the system of the present invention.

FIG. 2 is a 16-channel map covering the prefrontal-parietal lobe of the macaque selected by the present invention.

Figure 3 is an example of calculating the adjacency matrix of the phase lag index.

Figure 4 is a network topology diagram of the prefrontal-parietal lobe of the macaque monkey in different states (from left to right: awake state, moderate anesthesia, deep anesthesia).

Figure 5 is the adjacency matrix of the prefrontal-parietal lobe of the macaque in different states (from left to right: awake state, moderate anesthesia, deep anesthesia).

Figure 6 is an example of dual graph conversion.

Figure 7 shows the adjacency matrices of the two models.

FIG. 8 is a confusion matrix for predicting the test set by model 1, model 2 and dual-stream model in the present invention.

FIG. 9 is a ROC curve diagram of the prediction of the test set by the model 1, the model 2 and the dual-stream model in the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

A method for monitoring depth of anesthesia based on a graph convolutional neural network proposed by the present invention includes the following steps:

1) Collect several channels of EEG signals, and preprocess the original signals;

2) Intercept several time segments of different anesthesia stages as data samples, calculate the phase lag index PLI, calculate an adjacency matrix for each sample, obtain the network topology map samples of different anesthesia stages, and calculate the absolute value of the amplitude of each time segment. The average value of , as the node feature of the graph sample, the anesthesia stage includes the awake stage, the moderate anesthesia stage, and the deep anesthesia stage;

3) Convert the adjacency matrix of the network topology graph to a dual graph. The edge connections of the dual graphs obtained by the conversion are all the same, and the edge weight information is retained on the node features of the dual graph. The graph samples before conversion are constructed based on the phase lag index. Weighted graph, while retaining node features, and constructing a fully connected matrix to represent the topology of these node features;

4) Constructing a two-stream graph convolutional neural network, dividing the graph samples into two-stream graph data, and finding two common adjacency matrices of the two-stream graph data; in the two-stream graph data, the first-stream graph data is based on phase lag The weighted graph constructed by the index, and the other flow graph data is a fully connected graph that retains the characteristics of the original node;

5) Input the two-stream graph data into the two models of the two-stream graph convolutional neural network respectively, perform graph coarsening and fast pooling, reduce the data dimension, aggregate similar nodes, and output the predicted value of each anesthesia stage through the fully connected layer;

6) The predicted values of different anesthesia stages output by the two models of the dual-stream graph convolutional neural network are added respectively, and the category with the largest predicted value is output as the prediction result of the anesthesia stage.

The implementation process of each step is described in detail below:

In order to accomplish the purpose of the proposed invention, the present invention uses the experimental data of anesthesia of rhesus monkeys in the public database Neurotycho (http://neurotycho.org/) to explore the depth of anesthesia of rhesus monkeys induced by ketamine-medetomidine, and the experimental data used are 2 16-channel ECoG signals covering the prefrontal-parietal lobe of the brain from 5 anesthesia experiments in a single macaque. ECoG signals are electrocortical signals, and like EEG signals, they are all recorded electrical activity in the area between paired electrodes on the scalp. ECoG is an invasive brain-computer interface, which has higher spatial resolution and higher signal quality than EEG signals. The experiment included the awake stage before anesthesia, the anesthesia induction stage, the anesthesia maintenance stage, the anesthesia recovery stage, and the post-anesthesia awake stage. After intercepting time segments of the same length (1s) at different stages, a brain topology network was constructed based on the functional connectivity method (phase lag index), and the average absolute value of the signal amplitude of each channel in each time segment was calculated as the node feature to obtain three stages. (Awake, moderate anesthesia, deep anesthesia) graph samples, the data is randomly divided into training set, validation set, and test set according to the ratio of 8:1:1, and the training set is used to train the graph neural network model, and the validation set is used to adjust the model. Hyperparameters and initial evaluation of the ability of the model, test set to evaluate the generalization ability of the final model.

Firstly, the anesthesia experiment of rhesus monkeys is introduced in detail. The anesthesia experiment includes three stages.

11) Awake stage:

(a) AwakeEyeOpened-START/END: The macaque rests with its eyes open.

(b) AwakeEyeClose-START/END: By covering the eyes, the macaques rest with their eyes closed.

12) Anesthesia stage:

(a) AnesticDrugInjection: Intramuscular injection of ketamine-medetomidine.

(b) Anesthetized-START/END: Macaques are in a state of loss of consciousness (LOC). When macaques no longer respond to monkey hand manipulation or nostril touch with cotton swabs, or in humans, they enter the LOC state. In addition, LOC can also be confirmed by observing slow-wave oscillations in neural signals.

13) Recovery period:

(a) AntagonistInjection: Atipamezole was injected to recover monkeys from anesthesia.

(b) RecoveryEyeClosed-START/END: The point at which the slow-wave oscillations in the neural signal disappeared was regarded as the beginning of eye-closed recovery, in which the macaques rested peacefully with their eyes closed.

(c) RecoveryEyeOpened-START/END: By removing the blindfold, the monkey sat calmly with its eyes open.

The above is the task design of the anesthesia experiment.

The experimental data are 5 experiments of 2 rhesus monkeys, the signal includes all stages before and after anesthesia of rhesus monkeys, and the sampling rate is 1kHZ. A 16-channel ECoG signal covering the prefrontal-parietal lobe of the macaque is selected, as shown in Figure 2, the black dots mark the signal channels selected by the present invention.

In step 2), including intercepting data samples, calculating PLI, and calculating node characteristics of the network topology map.

21) Intercept data samples:

After preprocessing the data, multiple 1s segments were evenly intercepted at each stage of each experiment, and 1,000 1s segments were intercepted in each experiment in the awake stage, moderate anesthesia stage, and deep anesthesia stage, and the amount of moderate anesthesia data was small. , so the step of the sliding window is 0.1s.

(a) The awake phase data were intercepted during the pre-anesthesia awake phase and the post-anesthesia awake phase, where the post-anesthesia awake phase was the late recovery phase to determine the macaques' awake state.

(b) Moderate anesthesia data is the data of the intermediate period of anesthesia induction period (anesthetic injection - reaching LOC).

(c) The deep anesthesia data is the data of the anesthesia maintenance period.

Therefore, 5×3×1000 is obtained, that is, 15,000 data samples, 5 is 5 experiments, 3 is 3 stages, and 1000 is 1,000 samples in each stage of each experiment. Each sample includes 16 channels and 200 sampling points after resampling, and one sample corresponds to a network topology map.

22) Calculate the PLI to obtain the adjacency matrix;

The correlation between channels is calculated using the phase lag index (PLI), which is calculated as:

和

，使用希尔伯特变换计算瞬时相位

： Let the signal sequence of the two channels be

and

, using the Hilbert transform to calculate the instantaneous phase

:

表示

的希尔伯特(Hilbert)变换，i=1或2，j为虚数符号，计算方 法如下：

express

The Hilbert transform of , i = 1 or 2, j is an imaginary number symbol, the calculation method is as follows:

where PV represents the principal value of Cauchy, t is the time, and τ is the integral variable; after calculating the phase of each channel signal, the relative locking between the two channels can be calculated as:

In the formula, z ₂ *( t ) is the complex conjugate of z ₂ ( t );

The range of PLI is between 0 and 1, with 0 indicating no phase locking between the two channels and 1 indicating perfect phase coupling between the two channels. The PLI value is calculated as follows:

，对应邻接矩阵(1,2)的位置，而节点2于节点1的相关值

，即 得到的邻接矩阵是一个实对称矩阵。同样地，节点3于节点15的相关值对应邻接矩阵(3,15) 的位置，节点15于节点3的相关值对应邻接矩阵(3,15)的位置，同时

，以 此类推，便计算出了整个16×16邻接矩阵。 After the data samples are obtained, the phase correlation between channels is calculated using the phase lag index calculation formula, and a 16×16 adjacency matrix is obtained for each sample, and the adjacency values are distributed between 0-1. As shown in Figure 3, the phase lag index calculation adjacency matrix example: the 1s electrical signal (200 points) of node 1 and node 2 calculates a correlation value according to the phase lag index formula in the summary of the invention

, corresponding to the position of the adjacency matrix (1,2), and the correlation value between node 2 and node 1

, that is, the obtained adjacency matrix is a real symmetric matrix. Similarly, the correlation value of node 3 to node 15 corresponds to the position of the adjacency matrix (3, 15), the correlation value of node 15 to node 3 corresponds to the position of the adjacency matrix (3, 15), and at the same time

, and so on, the entire 16×16 adjacency matrix is calculated.

Figure 4 shows the network structure of the prefrontal-parietal lobe of the rhesus monkey Chibi brain at three different stages; Figure 5 shows the adjacency matrix corresponding to the topological network map of the rhesus monkey Chibi at three different stages (for the distinguishing effect of different stages with more significant performance) , the adjacency matrix drawn here adds self-connections).

23) Calculate the absolute value of the signal amplitude of each channel of each segment, and then take the average value as the node feature of each network topology map, and each node corresponds to a feature value.

An example of dual graph conversion in step 3) is shown in Figure 6: The idea of dual graph conversion is to convert edges into nodes and nodes into edges. If there are nodes connected between edges and edges, add edges to the obtained dual graph .

As shown in Figure 6, the original graph contains edges 01, 02, 03, 12, 13, and 23, which are turned into nodes in the right graph, and their edge weights become node features. In addition, 01, 02 have a common node 0; 01, 03 have a common node 0; 02, 03 have a common node 0; and so on, add a connection to the edge with a common node on the new graph, and the connection value is regarded as 1.

Based on the idea of the above dual graph conversion, the present invention converts multiple adjacency matrices (16×16) obtained by calculating the phase lag index into a dual graph (120×120), where 120=(16×16-16)/2. The edge connections of the converted dual graph are the same, that is, the adjacency matrix shown in Figure 7 (left), and the original edge weights are represented in the node features of the new graph.

On the other hand, the node features calculated by the amplitude information previously are also retained as the input of another graph convolutional neural network model, namely GCN model 2, GCN model 2 will be based on a 16 × 16 adjacency matrix to these The reserved node features are filtered. Since the edge information between the nodes and the nodes has been taken out and entered into the GCN model, it can be considered that the relationship between the nodes and the nodes of the reserved node information is equal, and one will be removed from the node. Connected fully connected matrix as its adjacency matrix.

Step 4) Build a two-stream graph convolutional neural network:

The dual-stream graph convolutional neural network includes two models. GCN model 1 (hereinafter referred to as model 1) designs a 6-layer graph convolution-pooling structure to extract edge weight information, that is, graph data with a size of 120×120 , GCN model 2 (hereinafter referred to as model 2) designs a 4-layer graph convolution-pooling structure for extracting node feature information, that is, 16×16 graph data.

In computer vision, CNN can effectively extract image features with neatly arranged pixels, that is, Euclidean structure data, and there are many non-Euclidean structure data in scientific research, such as social network, protein structure, etc. To apply machine learning methods to these non-European structured data, GCN has become the focus of research.

The difficulty in applying convolution to graph-structured data at the beginning is the problem of parameter sharing. Pictures with neatly arranged elements can satisfy translation invariance, and convolution kernels of the same size can be defined on the entire graph to perform convolution operations. The number of adjacent nodes of each node is different, so a convolution kernel of the same size cannot be used for convolution operations.

Graph convolutional neural networks include spectral methods and spatial methods. Faced with the problem of parameter sharing, spectral methods try to define convolutions in the spectral domain, not in the node domain. The node domain cannot satisfy translation invariance and cannot define convolutions of the same size. Therefore, the definition of convolution is implemented in the spectral domain, and then changed back to the spatial domain, which solves the problem of parameter sharing. The spatial method directly defines convolution in the node domain. In order to solve the problem of parameter sharing, the spatial method first determines the neighbor nodes of the target node, then arranges these neighbor nodes in sequence, and selects a fixed number of neighbor nodes for each node. Parameter sharing can be achieved, which is a different idea from the spectral method. Both methods perform better on graph-related tasks. The invention adopts the spectral graph convolution method to realize graph classification of two models.

The invention adopts the method GCN to define convolution directly in the spectral domain, which realizes the classification of graphs based on an adjacency matrix.

Spectral graph convolution: Extends graph convolution to the frequency domain of the graph through Fourier transform.

，在傅里叶域取一个

为参数的滤波器

： for input signal

, take one in the Fourier domain

filter for parameters

:

where U is the eigenvector matrix of the Laplacian matrix L of the graph. Laplacian matrix:

是拉普拉斯矩阵L的特征值组成的对角矩阵，

就是图上的傅里叶变换。 Among them, A is the adjacency matrix, D is the degree matrix,

is the diagonal matrix composed of the eigenvalues of the Laplace matrix L,

It is the Fourier transform of the graph.

用切比雪夫多项式进行K阶逼近，得到改进的卷积核： In order to reduce the amount of calculation, the

K-order approximation with Chebyshev polynomials to get an improved convolution kernel:

，

为拉普拉斯矩阵L中最大的特征值，

是切比 雪夫多项式的系数。 in

,

is the largest eigenvalue in the Laplace matrix L,

are the coefficients of the Chebyshev polynomial.

信号x进行滤波： Next, use the filter

Filter the signal x:

是在缩放后的拉普拉斯为

的k阶的切 比雪夫多项式，意味着

，我们可以用递归关系和

，

来计算

。in

is the scaled Laplacian as

A Chebyshev polynomial of order k, which means

, we can use the recurrence relation and

,

to calculate

.

After completing data preprocessing, data interception at different stages, and functional network construction, 30,000 graph samples at different stages are obtained. The obtained graph is a weighted graph constructed based on the phase lag index, which is different from the conventional binary graph. The weighted graph expresses the topological structure of the brain network in more detail; in addition, in order to achieve higher classification accuracy, the present invention also selects the absolute value and average of the channel signal amplitude as the node feature of the graph.

After the graph samples are constructed based on the topology and node information, the graph samples are divided into two-stream graph data, and two common adjacency matrices of the two types of graph data are found, and the adjacency matrices will be used as the topology of these graphs. A network that computes the Laplacian matrix of a graph. The first-stream graph data input into the graph is obtained by the dual graph method, and the edge weights are converted into graph convolution-sensitive node features, and an adjacency matrix of size 120×120 is obtained; the other stream data is the retained original node features, the original edge After the weights are taken out, the relationship between nodes is regarded as equal, so a 16×16 fully connected adjacency matrix (self-connection has been removed) is constructed.

In step 5), the obtained two types of graph data are respectively input into the two models of the two-stream graph convolutional neural network, such as the two-stream graph convolutional neural network method described in the content of the invention, the features of the first stream graph convolutional neural network input x is one-dimensional, that is, 120×1, and the input adjacency matrix is shown in Figure 7 (left), with a size of 120×120; the feature x input by the second flow graph convolutional neural network is also one-dimensional, that is, 16× 1. The input adjacency matrix is shown in Figure 7 (right), with a size of 16×16. The two-stream graph neural network model uses the same convolution method and pooling method, that is, the spectral graph convolution method that approximates the convolution kernel with Chebyshev polynomials, and the graph coarsening and graph coarsening based on the Graclus multi-level clustering algorithm. Fast pooling method. It's just that different convolution-pooling structures are designed due to the different sizes of the input data features.

，对中度麻醉状态的 预测概率为

，对深度麻醉状态的预测概率为

。模型2对清醒状态的预测概率为

，对 中度麻醉的预测概率为

，对深度麻醉的预测概率为

。将两个模型的对每个类别的预 测概率相加分别得到双流图卷积神经网络对清醒状态、中度麻醉状态、深度麻醉状态的预 测概率：

，

，

，取其中的最大值，便得到该值对应的 预测类别。 In step 6), the two models of the dual-stream graph convolutional neural network output the predicted probabilities respectively. The two models in the dual-stream graph convolutional neural network go through the softmax classification layer respectively and output the predicted probability of each category. The trained model 1 and model 2 respectively predict the test set, and obtain the predicted probability of model 1 for the awake state. for

, the predicted probability of moderate anesthesia is

, the predicted probability of deep anesthesia is

. Model 2 predicts the probability of being awake as

, the predicted probability for moderate anesthesia is

, the predicted probability of deep anesthesia is

. The predicted probabilities of the two models for each category are added to obtain the predicted probabilities of the two-stream graph convolutional neural network for the awake state, the moderate anesthesia state, and the deep anesthesia state:

,

,

, and take the maximum value to get the prediction category corresponding to this value.

Based on the above method, the anesthesia depth monitoring system based on the graph convolutional neural network proposed by the present invention is shown in FIG.

Data preprocessing module: It is used to preprocess the cerebral cortex signal, filter and downsample the original data, filtering can filter out clutter and alternating current signal; downsampling can reduce the amount of data to be processed. Preprocessing includes: 0.5-100HZ filtering; 50HZ notch filtering; 200HZ resampling. The above preprocessing operations are all done in MATLAB2016b.

Functional network building block: It is used to cut the sample data into several time segments of different anesthesia stages, calculate the correlation between channels through the phase lag index method, construct the adjacency matrix, obtain the network topology map of key brain areas, and calculate the signal amplitude The mean absolute value is used as a graph node feature.

Graph conversion module: It is used to convert the adjacency matrix of the network topology graph sample into a dual graph, and convert the edge weights into graph convolution-sensitive node features, so that different graphs can obtain the same adjacency matrix. The direct application of the product graph classification method; at the same time, the original node features are preserved, and a new fully connected adjacency matrix is constructed as another flow graph data.

Two-stream graph convolutional neural network module: used to store two models of the two-stream graph convolutional neural network, model 1 is used to extract edge weight information, model 2 is used to extract node feature information, and the output prediction probability classification of the two models is related to each other. Add to get the prediction result. The dual-stream graph convolutional neural network module designs a dual-stream structure, which is used to learn the two types of graph data obtained by the graph conversion module, and uses spectral graph convolution to extract graph features, graph coarsening and fast pooling operations to aggregate similar nodes and reduce computation. Finally, the softmax layer predicts the predicted probabilities of different states of anesthesia, adds the predicted probabilities of the two models to achieve model fusion, and uses the added probabilities to predict the test set.

，

表示每一图卷积层的滤波器数量。 Table 1 shows the internal parameters of the two-stream graph convolutional neural network module model 1, and Table 2 shows the internal parameters of the two-stream graph convolutional neural network module model 2, where O represents the number of anesthesia depth classification tasks,

,

Represents the number of filters per graph convolutional layer.

The structure of the double-layer graph convolutional neural network built in the present invention is described as follows:

In the GCN model of the present invention, after the graph convolution layer, the dimension of the graph is unchanged. For the maximum pooling layer, the dimension of the graph will be reduced by half, which means that the N×N Laplacian matrix, after the maximum After the pooling layer, the dimension of the graph will become N/2×N/2. For an adjacency matrix of size 120×120, a 6-layer pooling structure of 64-32-16-8-4-2-1 can be used or The 3-layer pooling structure of 120-60-30-15, the present invention chooses the former. For an adjacency matrix of size 16×16, the present invention selects a 4-layer pooling structure of 16-8-4-2-1.

The input adjacency matrix gives a description of the graph structure. Coarse the adjacency matrix to obtain a multi-level coarsening matrix. At the same time, according to the multi-level coarsening matrix obtained by the coarsening, the original data is rearranged and quickly pooled. The raw data is reorganized into 3D data by the rearranged relations returned by the coarsening, which are fed into the network for convolution.

The graph convolutional layer learns the features of the graph data, and the pooling layer reduces the data dimension and aggregates similar nodes. After the two models are trained, use the trained model to test the test set to get the predicted category.

Table 3 shows the prediction accuracy of model 1 and model 2 and the combined dual-stream model on the test set. It can be seen that model 1 and model 2 can achieve quite good accuracy, and the combination of the two models can make the prediction reach better performance, which shows that the two models learned the discriminativeness of different features.

FIG. 8 is the confusion matrix of the test set, and FIG. 9 is the ROC curve diagram of the test set, including the evaluation of Model 1, Model 2 and the two-stream model, the test set evaluates the generalization ability of the final model, and the present invention obtains three classifications of the test set The accuracy reaches 95.4%.

The training and testing process of the dual-stream graph convolutional neural network model were completed in the Python3.6 TensorFlow 1.13.1 environment.

Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solution of the patent and not to limit it. Although the patent has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solution of the patent can be Modifications or equivalent substitutions are made without departing from the spirit and scope of the technical solutions of this patent, and they should all be covered by the scope of the claims of this patent.

Claims (9)

1. an anesthesia depth monitoring system based on graph convolutional neural network, is characterized in that: described system comprises data preprocessing module, functional network building module, graph conversion module and dual-flow graph convolutional neural network module; The data preprocessing module: used to preprocess the cerebral cortex signal; The functional network building module is used to cut the sample data into several time segments of different anesthesia stages, calculate the phase lag index PLI, calculate an adjacency matrix for each sample, and obtain the network topology map samples of different anesthesia stages. Stages include awake stage, moderate anesthesia stage, and deep anesthesia stage; The graph conversion module is used to convert the adjacency matrix of the network topology graph sample into a dual graph. The edge connections of the converted dual graphs are the same, and the edge weight information is retained on the node features of the dual graph. The sample is a weighted graph constructed based on the phase lag index; The two-stream graph convolutional neural network module: used to store two models of the two-stream graph convolutional neural network, model 1 is used to extract edge weight information, model 2 is used to extract node feature information, and the two models are output prediction probability. The classification is added to obtain the prediction result. 2. a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, a method for monitoring the depth of anesthesia based on a graph convolutional neural network is realized, characterized in that: the method Include steps: 1) Collect several channels of EEG signals, and preprocess the original signals; 2) Intercept several time segments of different anesthesia stages as data samples, calculate the phase lag index PLI, calculate an adjacency matrix for each sample, and obtain the network topology map samples of different anesthesia stages, the anesthesia stages include awake stage, moderate anesthesia stage, deep anesthesia stage; 3) Convert the adjacency matrix of the network topology graph into a dual graph, the edge connections of the converted dual graphs are the same, the edge weight information is retained on the node features of the dual graph, and the converted graph samples are based on the phase lag index. The constructed weighted graph, while retaining the node features, and constructing a fully connected matrix to represent the topology of these node features; 4) Constructing a two-stream graph convolutional neural network, dividing the graph samples into two-stream graph data, and finding two common adjacency matrices of the two-stream graph data; in the two-stream graph data, the first-stream graph data is based on phase lag The weighted graph constructed exponentially, and the data of the other flow graph is a fully connected adjacency matrix that retains the characteristics of the original node; 5) Input the two-stream graph data into the two models of the two-stream graph convolutional neural network respectively, perform graph coarsening and fast pooling, reduce the data dimension, aggregate similar nodes, and output the predicted value of each anesthesia stage through the fully connected layer; 6) The predicted values of different anesthesia stages output by the two models of the dual-stream graph convolutional neural network are added respectively, and the category with the largest predicted value is output as the prediction result of the anesthesia stage. 3 . The computer-readable storage medium according to claim 2 , wherein the ECoG signal of the mesencephalic cortex in the step 1) is a 16-channel ECoG signal of the frontal lobe-parietal lobe of the brain of the research subject; the preprocessing Including 0.1~100Hz filtering, 50Hz notch filtering and 200Hz resampling. 4. A computer-readable storage medium according to claim 2, wherein the calculation method of the phase lag index PLI is: Let the signal sequence of the two channels be

and

, using the Hilbert transform to calculate the instantaneous phase

:

in,

express

The Hilbert transform of , i = 1 or 2, and j is the imaginary sign:

In the formula, P.V. represents the principal value of Cauchy, t is the time, and τ is the integral variable; The relative lock between the two channels is calculated as:

In the formula, z ₂ *( t ) is the complex conjugate of z ₂ ( t ); Then the PLI value is calculated as follows:

The range of PLI is between 0 and 1, with 0 indicating no phase locking between the two channels and 1 indicating perfect phase coupling between the two channels. 5. A computer-readable storage medium according to claim 2, characterized in that: in the step 4), a dual-stream graph convolutional neural network is constructed using the spectral domain graph convolution method GCN, and the graph is transformed by Fourier transformation. Convolution extends into the frequency domain of the graph and uses filters to filter the signal. 6 . The computer-readable storage medium according to claim 2 , wherein in the step 5), the two models of the dual-stream graph convolutional neural network both adopt the spectrogram that approximates the convolution kernel with Chebyshev polynomials. 7 . Convolution method, and graph coarsening and fast pooling method based on Graclus multi-level clustering algorithm. 7 . The computer-readable storage medium according to claim 6 , wherein the Graclus multi-level clustering algorithm adopts a greedy algorithm to measure the continuous thickness of the graph, so as to minimize the spectral clustering target. 8 . 8 . The computer-readable storage medium according to claim 2 , wherein in the step 2), the average absolute value of the signal amplitude of each channel of each time segment is calculated as the node feature. 9 . 9 . The computer-readable storage medium according to claim 2 , wherein the data samples in the step 2) are randomly divided into a training set, a verification set, and a test set according to a ratio of 8:1:1, 10 . The training set is used to train the graph neural network model, the validation set is used to adjust the hyperparameters of the model and the ability of the model is initially evaluated, and the test set is used to evaluate the generalization ability of the final model.

Images