Automatic detection device for seizures of severe encephalitis patients based on brain network optimization

Abstract

The application provides a severe encephalitis patient seizure automatic detection device based on brain network optimization, characterized by comprising an acquisition module, a preprocessing module and a detection module, the acquisition module acquires original electroencephalogram signals of the encephalitis patient to be detected; the preprocessing module inputs the detected electroencephalogram signals after preprocessing into a recognition model in the detection module to obtain a seizure detection result; the recognition model in the detection module integrates multiple brain networks and multiple features, more comprehensively reflects the relationship between each node in the brain network from multiple related dimensions; and the network layer and feature layer optimization based on the improved genetic algorithm realizes the optimal determination of multiple network weighting coefficients and multiple network weighting coefficients, takes the optimal network layer and feature layer as machine learning input, thereby greatly improving the accuracy of epilepsy recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, and in particular to an automatic detection device for seizures of patients with severe encephalitis based on brain network optimization. BACKGROUND

[0002] Encephalitis is a serious disease caused by inflammation of the brain parenchyma, which causes neurological dysfunction. The incidence of encephalitis is about 1.5-10.5 / 100,000, of which the incidence of children accounts for 37%-94.2% of the total incidence. Encephalitis is a frequently-occurring and common disease of the central nervous system infection in children. Severe encephalitis is often accompanied by consciousness disorders, seizures, etc., and can even be life-threatening. Electroencephalogram (EEG) records the brain wave pattern by tracking the electrical activity of the brain to diagnose epilepsy. In this stage, in order to detect seizures, human experts need to visually mark long-term EEG records. However, this is a very tedious, time-consuming and costly task. EEG signals contain a lot of information about the fluctuation of brain function behavior, and its application in seizure detection is still a challenging problem. Therefore, it is necessary to develop an automated technique for detecting seizures of patients with severe encephalitis.

[0003] With the wide application of complex network theory, its application in the field of electroencephalogram is also increasing. A commonly used method is to calculate the mutual information (MI) between EEG channels. The complex network constructed by mutual information can further analyze the connectivity of the network, network characteristics and other characteristics of the network, so as to identify and analyze mental diseases such as epilepsy and schizophrenia. In addition to MI, the Pearson correlation coefficient (PCC) is also a commonly used network correlation measure. In addition, the Spearman rank correlation coefficient, the standardized permutation mutual information and the permutation disalignment index (PDI) have also been proposed for constructing brain networks. Since different correlation coefficients reflect different characteristics between network nodes, the network constructed by a single correlation coefficient has a single nature, and different recognition results will be produced when different correlation coefficients are used to construct networks, which has the problems of one-sided recognition, great limitation and low recognition accuracy. SUMMARY

[0004] The present application is provided in order to overcome the deficiencies of the prior art, and provides an automatic detection device for seizures of patients with severe encephalitis based on brain network optimization, which has high recognition accuracy.

[0005] In order to achieve the above object, the application provides a severe encephalitis patient seizure automatic detection device based on brain network optimization, comprising a collection module, a preprocessing module and a detection module, the collection module collects original electroencephalogram signals of the encephalitis patient to be detected; the preprocessing module inputs the preprocessed electroencephalogram signals to be detected into a recognition model in the detection module to obtain a detection result of the seizure; the recognition model in the detection module is obtained by training in the following manner:

[0006] Obtain original electroencephalogram signal samples of a plurality of severe encephalitis patients and label the seizure segments of each channel in each sample;

[0007] Preprocess each electroencephalogram signal sample to obtain a plurality of sub-band signals of different frequency bands;

[0008] Construct a multi-layer weighted brain network and optimize the multi-layer brain network based on an improved genetic algorithm, which comprises: constructing a multi-layer weighted brain network based on a plurality of correlation coefficients respectively, each brain network corresponding to a weighted coefficient, the network weighted coefficients of the plurality of brain networks forming a network layer; dividing the seizure segments in the electroencephalogram signal sample into electroencephalogram segments with a certain length, each electroencephalogram segment corresponding to a plurality of sub-band signals; extracting network features of the multi-layer weighted brain network on different sub-band signals, and weighting and averaging the features corresponding to all sub-band signals to obtain a plurality of segment features, the feature weighting coefficients of each segment feature forming a feature layer corresponding to the network layer; independently optimizing a plurality of network weighted coefficients of the network layer and a plurality of feature weighted coefficients of the feature layer respectively by using an improved genetic algorithm;

[0009] According to the position of each feature in the feature layer and the network layer, multiply the optimized network layer weighted coefficient and the corresponding feature layer weighted coefficient to construct a feature vector of each electroencephalogram segment; multiply a plurality of segment features and a plurality of corresponding feature vectors to obtain a recognition model by training a random forest classification model.

[0010] According to an embodiment of the application, the relationship between the channels in the electroencephalogram signal is quantified by using three correlation coefficients, namely mutual information, Pearson correlation coefficient and standardized permutation mutual information and permutation dislocation index, to construct a three-layer brain network.

[0011] According to an embodiment of the application, the plurality of features of each sub-band signal include weighted degree, clustering coefficient, harmonic centrality, modularity, close centrality and feature vector centrality.

[0012] According to an embodiment of the application, when training the random forest classification model to obtain the epilepsy recognition result, the feature vector of each electroencephalogram segment is constructed based on the three optimized network layer weighted coefficients and six feature layer weighted coefficients to form the input of the random forest model, and the feature vector of each electroencephalogram segment contains eighteen vector elements.

[0013] According to an embodiment of the present application, in each brain network, the weighted degree k i The importance degree of node i in the brain network is represented by the following formula:

[0014]

[0015] wherein M is the set of all nodes in each brain network, and A is the adjacency matrix; A i,j is an element in the adjacency matrix A, representing the relationship between node i and node j; ω i,j is the weight between i and j nodes;

[0016] The clustering coefficient c i reflects the degree to which nodes tend to cluster together, and is represented by the following formula:

[0017]

[0018] wherein k i is the importance degree of node i in the network; d i is the degree of node i; j and h nodes represent the other two vertices that form a triad with node i, d p is the degree of the pth node, used to traverse all nodes;

[0019] The harmonic centrality hc i measures the difficulty of node i to reach other nodes, and is represented by the following formula:

[0020]

[0021] wherein n is the total number of nodes in the network, and dist i,j represents the shortest path between i and j nodes;

[0022] The close centrality clos i is represented by the following formula:

[0023]

[0024] wherein M is the set of all nodes in each brain network, and dist i,j represents the shortest path between i and j nodes;

[0025] The modularity Q represents the strength between groups after dividing the nodes into groups, and is represented by the following formula:

[0026]

[0027] wherein C i is the cluster name of node i, and δ(C i C j1 when i and j nodes belong to the same cluster, otherwise 0; k i is the importance of node i in the network; k j is the importance of node j in the network;

[0028] The eigenvector centrality reflects the importance of the neighbor nodes, and the formula is as follows:

[0029] Ae=λe

[0030] Wherein, e=[e1,…,e i ,…e n ] represents a vector composed of eigenvector centralities of all nodes, A is an adjacency matrix, and λ is a corresponding eigenvalue of the adjacency matrix.

[0031] According to an embodiment of the present application, the preprocessing module in preprocessing the brain electrical signal to be detected and the brain electrical signal sample:

[0032] A band-pass filter is used to filter out frequency components above 48Hz and below 1Hz;

[0033] The filtered brain electrical signal is subjected to five-layer wavelet packet decomposition to obtain six sub-band signals of 1-4Hz, 4-8Hz, 8-12Hz, 12-16Hz, 16-24Hz and 24-32Hz.

[0034] According to an embodiment of the present application, the steps of independently optimizing the network layer and the feature layer based on the improved genetic algorithm include:

[0035] Setting the initial population size, evolution generation number, crossover probability and mutation probability;

[0036] Randomly generating a plurality of network weighting coefficients of the network layer and a plurality of feature weighting coefficients of the feature layer;

[0037] Performing network layer weighting coefficient optimization to select the plurality of network weighting coefficients in the generation with the highest accuracy; setting the constraint condition as the plurality of network weighting coefficients in the network layer adding up to 1, the fitness function as maximizing the accuracy of the random forest classification model in classifying the test set, and performing the plurality of network weighting coefficient optimization through the crossover, recombination and mutation operations of the genetic algorithm;

[0038] Performing feature layer coefficient optimization to select the weighting coefficient in the generation with the highest accuracy; setting the constraint condition as the plurality of feature weighting coefficients in the feature layer adding up to 1, the fitness function as maximizing the accuracy of the random forest classification model in classifying the test set, and performing the weight optimization through the crossover, recombination and mutation operations of the genetic algorithm;

[0039] Repeat the above steps until the test set classification accuracy no longer changes, and the genetic algorithm optimization ends.

[0040] According to one embodiment of the present invention, when the acquisition module acquires the EEG signals of multiple patients with severe encephalitis, the sampling frequency is 256Hz, the electrode distribution adopts the international 10-20 EEG acquisition standard, and a total of 20 channels of EEG data are acquired.

[0041] In summary, the automatic detection device for epileptic seizures in patients with severe encephalitis based on brain network optimization provided by this invention integrates multiple brain networks and multiple features in the detection module, reflecting the relationships between nodes in the brain network more comprehensively from multiple relevant dimensions. Furthermore, the optimization of network layers and feature layers based on the improved genetic algorithm achieves the optimal determination of multiple network weighting coefficients and uses the optimal network layers and feature layers as machine learning input, thereby greatly improving the accuracy of epilepsy recognition.

[0042] To make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0043] Figure 1 The diagram shown is a schematic representation of an automatic detection device for epileptic seizures in patients with severe encephalitis based on brain network optimization, provided in an embodiment of the present invention.

[0044] Figure 2 The training is shown below. Figure 1 A flowchart illustrating the recognition model within the detection module.

[0045] Figure 3 As shown Figure 2 A flowchart of step S30.

[0046] Figure 4 The diagram shows a flowchart of optimizing network layers and feature layers based on an improved genetic algorithm. Detailed Implementation

[0047] like Figure 1As shown, the embodiment provides a severe encephalitis patient seizure automatic detection device based on brain network optimization, which comprises a collection module 10, a preprocessing module 20 and a detection module 30. The collection module 10 collects the original electroencephalogram of the encephalitis patient to be detected; specifically, the collection module 10 samples the electroencephalogram of the severe encephalitis patient based on the electrode distribution using the international 10-20 electroencephalogram collection standard, and the sampling frequency is 256 Hz, and a total of 20 channels of electroencephalogram data are collected. After obtaining the original electroencephalogram of the encephalitis patient, the preprocessing module 20 is preprocessed; specifically, the preprocessing module 20 uses a band-pass filter to filter out frequency components above 48 Hz and below 1 Hz in the original electroencephalogram; and the electroencephalogram obtained after filtering is subjected to five-layer wavelet packet decomposition to obtain six sub-band signals of 1-4 Hz, 4-8 Hz, 8-12 Hz, 12-16 Hz, 16-24 Hz and 24-32 Hz; however, the division of the sub-band signals after preprocessing is not limited by the present application. The six sub-band signals are input into the detection module 30 to obtain the detection result of the seizure.

[0048] In this embodiment, the detection module 30 stores a recognition model for automatically detecting seizures, which is obtained by training. The specific training steps include: obtaining a plurality of original electroencephalogram samples of severe encephalitis patients and labeling the seizure segments of each channel in each sample (step S10). Preprocess each electroencephalogram sample to obtain a plurality of sub-band signals of different frequency bands (step S20). Construct a multi-layer weighted brain network and optimize the multi-layer brain network based on an improved genetic algorithm (step S30). Train a random forest model based on the network layer weighting coefficients and feature layer weighting coefficients of the multi-layer network optimized by the improved genetic algorithm to realize the classification of seizure results (step S40); specifically, according to the feature layer and network layer position to which each feature belongs, multiply the optimized network layer weighting coefficient and the corresponding feature layer weighting coefficient to construct a feature vector for each electroencephalogram segment; multiply a plurality of segment features and a plurality of corresponding feature vectors to obtain a random forest classification model as input to obtain a recognition model.

[0049] The training of the recognition model starts from step S10, in which a large number of long-term monitoring original electroencephalogram signals of patients with different genders and ages are collected using a multi-channel electroencephalograph as electroencephalogram samples. The sampling frequency is 256 Hz, the electrode distribution uses the international 10-20 electroencephalogram collection standard, and a total of 20 channels of electroencephalogram data are collected. The seizure segments of each channel of the obtained large number of electroencephalogram samples are labeled to form a sample database. Then step S20 is performed, the labeled original electroencephalogram samples in the sample database are preprocessed, and the preprocessing steps are the same as those in the preprocessing module 20, that is, after filtering out interference signals using a band-pass filter, sub-band signal division in the frequency domain is performed to obtain six sub-band signals.

[0050] After obtaining the six sub-band signals, step S30 is performed to construct a multi-layer weighted brain network and perform feature extraction on each layer of the network based on the six sub-band signals to form a plurality of feature weighting parameters, and then optimize the weighting coefficients of the network layers and the feature layers based on a genetic optimization algorithm. Specifically, this step includes:

[0051] Step S31: Construct a multi-layer weighted brain network based on a plurality of correlation coefficients respectively, each brain network corresponding to a weighting coefficient, and the network weighting coefficients of the plurality of brain networks forming a network layer. Specifically, three correlation coefficients, namely mutual information, Pearson correlation coefficient, and standardized permutation mutual information and permutation dislocation index, are used to quantify the relationship between the channels in the electroencephalogram signal to construct three layers of weighted brain networks. Each layer of the network corresponds to a network weighting coefficient, represented by K1, K2, and K3 respectively, and the three network weighting coefficients form the network layer weighting coefficient.

[0052] Then, step S32 is performed to divide the seizure segment in the electroencephalogram signal sample into electroencephalogram segments with a certain length, and each electroencephalogram segment corresponds to a plurality of sub-band signals. Specifically, the seizure segment marked in the electroencephalogram signal sample is divided into several 4s electroencephalogram segments using the sliding window method, and the divided electroencephalogram segments are denoted as x(n), n = 1, 2, …, N; wherein N is the number of electroencephalogram segments, and in this embodiment, the sampling frequency of the electroencephalogram signal sample is 256 Hz, so N = 1024. Based on the preprocessing of step S20, each computer segment also has six sub-band signals.

[0053] Step S33: Extract the network features of the multi-layer weighted brain network on different sub-band signals, and obtain a plurality of segment features by weighted averaging the features corresponding to all sub-band signals, and the feature weighting coefficients of each segment feature form the feature layer corresponding to the network layer. In this embodiment, the plurality of features of each sub-band signal include six network features, namely the weighted degree k i , the clustering coefficient c i , the harmonic centrality hc i , the modularity Q, the close centrality clos i , and the eigenvector centrality e. For the first sub-band of 1-4 Hz, the first sub-band weighted degree k i , the first sub-band clustering coefficient c i , the first sub-band harmonic centrality hc i , the first sub-band modularity Q, the first sub-band close centrality clos i , and the first sub-band eigenvector centrality e will be obtained after extraction; for the second sub-band of 4-8 Hz, the second sub-band weighted degree k i , the second sub-band clustering coefficient c i , the second sub-band harmonic centrality hc i, second sub-band modularity Q, second sub-band closeness centrality c i and second sub-band eigenvector centrality e. Correspondingly, six features corresponding to each sub-band can be obtained in turn.

[0054] The first sub-band weighted degree k i , second sub-band weighted degree k i , sixth sub-band weighted degree k i The weighted average of the weighted degrees k i and the corresponding feature weighting coefficient H1; the first sub-band clustering coefficient c i , second sub-band clustering coefficient c i , sixth sub-band clustering coefficient c i The weighted average of the clustering coefficients c i and the corresponding feature weighting coefficient H2; the harmonic centrality hc i (corresponding feature weighting coefficient H3), modularity Q (corresponding feature weighting coefficient H4), closeness centrality clos i (corresponding feature weighting coefficient H5) and eigenvector centrality e (corresponding feature weighting coefficient H6) are obtained in turn. The multi-layer weighted brain network corresponds to six feature weighting coefficients (H1, H2, H3, H4, H5, H6) on each electroencephalogram segment.

[0055] For the six network features, in each layer of brain network, the weighted degree k i characterizes the importance of node i in the brain network, and its formula is as follows:

[0056]

[0057] wherein M is the set of all nodes in each brain network, and A is the adjacency matrix; A i,j is an element in the adjacency matrix A, indicating the relationship between node i and node j; ω i,j is the weight between i and j nodes;

[0058] The clustering coefficient c i reflects the degree to which nodes tend to gather together, and its formula is as follows:

[0059]

[0060] wherein k i is the importance of node i in the network; d i is the degree of node i; j and h nodes represent the other two vertices that form a triad with node i, d p is the degree of the pth node, which is used to traverse all nodes;

[0061] Harmonic centrality hc i The difficulty of a node to reach other nodes is measured by the following formula:

[0062]

[0063] Where n is the total number of nodes in the network, dist i,j represents the shortest path between i and j nodes;

[0064] Closeness centrality clos i The formula is as follows:

[0065]

[0066] Where M is the set of all nodes in each brain network, dist i,j represents the shortest path between i and j nodes;

[0067] Modularity Q represents the strength between groups after the nodes are divided into groups, and its formula is:

[0068]

[0069] Where C i is the cluster name of i node, δ(C i C j is 1 when i and j nodes belong to the same cluster, otherwise 0; k i is the importance of node i in the network; k j is the importance of node j in the network;

[0070] The eigenvector centrality reflects the importance of neighbor nodes, and the formula is as follows:

[0071] Ae=λe

[0072] Where e=[e1,...,e i , …e n ] represents a vector composed of eigenvector centralities of all nodes, A is an adjacency matrix, and λ is the corresponding eigenvalue of the adjacency matrix.

[0073] The recognition result of the random forest classification model is related to the features extracted on each segment and the weighted coefficients of the segment. As indicated by the above formula, the calculation of the features is related to the adjacency matrix of the brain network, which is determined by the characteristics of the electroencephalogram signal. In order to improve the accuracy of recognition, it is necessary to determine the relevant weighted coefficients; for the weighted coefficients of the segment, according to the different network structures, the network layer is divided into three network weighted coefficients; according to different features, the feature layer is divided into six feature weighted coefficients.

[0074] In this embodiment, step S34 will adopt the improved genetic algorithm to independently optimize the network layer weighting coefficients and the feature layer weighting coefficients, and combine the prediction of the test set by the random forest model in step S40 to obtain the optimal network layer weighting coefficients and the optimal feature layer weighting coefficients. Specifically, this step includes:

[0075] Step S341: setting the initial population size, the evolution number of generations, the crossover probability, and the mutation probability;

[0076] Step S342: randomly generating a plurality of network weighting coefficients of the network layer and a plurality of feature weighting coefficients of the feature layer;

[0077] Step S343: performing the optimization of the network layer weighting coefficients to select the plurality of network weighting coefficients in the generation with the highest accuracy. The constraint condition is that the plurality of network weighting coefficients in the network layer are added to 1, the fitness function is to maximize the accuracy of the classification of the test set by the random forest model in step S40, and the plurality of network weighting coefficients are optimized through the crossover, recombination, and mutation operations of the genetic algorithm;

[0078] Step S344: performing the optimization of the feature layer weighting coefficients to select the weighting coefficients in the generation with the highest accuracy. The constraint condition is that the plurality of feature weighting coefficients in the feature layer are added to 1, the fitness function is to maximize the accuracy of the classification of the test set by the random forest model in step S40, and the weight optimization is performed through the crossover, recombination, and mutation operations of the genetic algorithm;

[0079] The above steps S342 to S344 are repeated until the accuracy of the test set no longer changes, the optimization of the genetic algorithm is completed, and the optimized network layer weighting coefficients and the feature layer weighting coefficients are obtained.

[0080] After obtaining the network layer weighting coefficients and the feature layer weighting coefficients of the multi-layer network based on the improved genetic algorithm optimization in step S30, step S40 is performed: training the random forest model to realize the classification of the seizure results. Specifically, this step includes:

[0081] Step S41: multiplying the optimized network layer weighting coefficients and the corresponding feature layer weighting coefficients to construct the feature vector of each electroencephalogram segment according to the feature layer and the network layer position to which each feature belongs. Specifically, based on the optimized three network layer weighting coefficients and six feature layer weighting coefficients, the feature vector of each electroencephalogram segment is constructed, and the feature vector KH of each electroencephalogram segment contains eighteen vector elements, KH = [K1*H1, K1*H2, K1*H3, K1*H4, K1*H5, K1*H6, K2*H1, K2*H2, K2*H3, K2*H4, K2*H5, K2*H6, K3*H1, K3*H2, K3*H3, K3*H4, K3*H5, K3*H6].

[0082] Step S42: randomly divide the plurality of EEG segments of the plurality of EEG signal samples into a training set and a test set, multiply the plurality of network features of each EEG segment in the training set by the corresponding feature vector, and input the result into the random forest classifier to train the plurality of decision trees in the random forest classifier, thereby forming a random forest model. Multiply the plurality of network features of each EEG segment in the test set by the corresponding feature vector, and input the result into the trained random forest model to obtain a test result of the seizure detection, and determine the trained random forest model based on the test result. The random forest classifier uses the integration of a plurality of weak classifiers to obtain a classification result by voting, and uses sampling with replacement to construct a sub-data set equal in size to the original data set. Then, the sub-data set is used to build a sub-decision tree, that is, each decision tree obtains a classification result according to the sub-data set. Finally, the output of the random forest classifier is obtained by voting on all the results of the sub-decision trees. Each decision tree has the same distribution, but the classification ability depends on the input data.

[0083] At this point, the recognition model in the detection module 30 based on the multi-layer network structure and the random forest model has been trained. When performing automatic detection of seizures in patients with encephalitis, the EEG signal processed by the preprocessing module 20 is input into the detection module 30. After signal segmentation, the segment features are extracted, the extracted segment features are associated with the optimized network layer correlation coefficient and the feature layer correlation coefficient, and then input into the random forest model. After classification by the random forest model, the seizure detection result is obtained to realize automatic detection.

[0084] In this embodiment, the automatic detection device for seizures in patients with severe encephalitis based on brain network optimization further includes an optimization training module 40. The optimization training module performs multi-layer network construction, double-layer weighting coefficient optimization, and random forest model training to obtain the recognition model in the detection module 30. The training module 40 includes an improved genetic algorithm optimized multi-layer network unit 43 and a model training unit 44. The training steps S10 and S20 can be implemented based on the acquisition module 10 and the preprocessing module 20 in the automatic detection device for seizures in patients with severe encephalitis based on brain network optimization provided in this embodiment. The improved genetic algorithm optimized multi-layer network unit 43 will execute S30 in the training steps, and the model training unit 44 will execute S30 in the above training steps, which will not be repeated here. However, the present application does not make any limitation in this regard. In other embodiments, the recognition model in the detection module can also be trained by an external computer device.

[0085] To sum up, the recognition model set in the detection module of the automatic detection device for seizures of patients with severe encephalitis based on brain network optimization fuses multiple brain networks and multiple features, more comprehensively reflects the relationship between each node in the brain network from multiple related dimensions; and the network layer and feature layer optimization based on the improved genetic algorithm realizes the optimal determination of multiple network weighting coefficients and multiple network weighting coefficients, takes the optimal network layer and feature layer as the machine learning input, thereby greatly improving the accuracy of epilepsy recognition.

[0086] Although the present application has been disclosed by the preferred embodiments as above, it is not intended to limit the present application, and any person skilled in the art can make some changes and modifications without departing from the spirit and scope of the present application, and the protection scope of the present application shall be subject to the scope required by the claims.

Claims

1. A severe encephalitis patient seizure automatic detection device based on brain network optimization, characterized in that, The method comprises a collection module, a preprocessing module and a detection module, the collection module collects original electroencephalogram signals of a patient with encephalitis to be detected; The preprocessing module inputs the collected electroencephalogram signals to be detected into an identification model in the detection module after preprocessing to obtain a detection result of a seizure; the identification model in the detection module is obtained by training in the following manner: Obtain original electroencephalogram signal samples of multiple severe encephalitis patients and label the seizure segments of each channel in each sample; Preprocess each electroencephalogram signal sample to obtain multiple sub-band signals of different frequency bands; Construct a multi-layer weighted brain network and optimize the multi-layer brain network based on an improved genetic algorithm, which comprises: constructing a multi-layer weighted brain network based on multiple correlation coefficients respectively, each brain network corresponding to a weighted coefficient, and the network weighted coefficients of the multiple brain networks forming a network layer; Divide the seizure segments in the electroencephalogram signal samples into electroencephalogram segments with a certain length, each electroencephalogram segment corresponding to multiple sub-band signals; extract network features of the multi-layer weighted brain network on different sub-band signals, including weighted degree, clustering coefficient, harmonic centrality, modularity, close centrality and feature vector centrality, and weight average all the features corresponding to the sub-band signals to obtain multiple segment features, the feature weighting coefficients of each segment feature forming a feature layer corresponding to the network layer; independently optimize multiple network weighted coefficients of the network layer and multiple feature weighted coefficients of the feature layer using an improved genetic algorithm; Multiply the optimized network layer weighted coefficients and the corresponding feature layer weighted coefficients according to the feature layer and the network layer position to which each feature belongs to construct a feature vector of each electroencephalogram segment; multiply the multiple segment features and the corresponding multiple feature vectors to obtain an identification model by training a random forest classification model.

2. The device for automatic detection of seizures in patients with severe encephalitis based on brain network optimization according to claim 1, characterized in that, Three correlation coefficients, including mutual information, Pearson correlation coefficient and standardized permutation mutual information and permutation dislocation index, are used to quantify the relationship between the channels in the electroencephalogram signal to construct a three-layer brain network. 3.The brain network optimization based automatic seizure detection device for severe encephalitis patients according to claim 1, wherein, When training the random forest classification model to obtain the epilepsy recognition result, the feature vectors of each electroencephalogram segment are constructed based on the optimized three network layer weighted coefficients and six feature layer weighted coefficients to form the input of the random forest model, and each electroencephalogram segment contains eighteen vector elements.

4. The device for automatic detection of seizures in patients with severe encephalitis based on brain network optimization according to claim 1, characterized in that, In each layer brain network, the weighted degree k i The importance degree of node i in brain network is represented by the following formula: Wherein, M is the set of all nodes in each brain network, A is the adjacency matrix; A i,j is an element in the adjacency matrix A, which represents the relationship between node i and node j; ω i,j is the weight between i and j nodes; Clustering coefficient c i Reflects the extent to which nodes tend to cluster together, formula as follows: where k i is the importance of node i in the network; d i is the degree of node i; j and h nodes represent the other two vertices that form a triad with node i, d p is the degree of the pth node, used to traverse all nodes; harmonic centrality hc i The ease of a node to reach other nodes is measured by the following formula: where n is the total number of nodes in the network, dist i,j denotes the shortest path between i and j nodes; Closeness centrality c i The formula is as follows: where M is the set of all nodes in each brain network, dist i,j denotes the shortest path between i and j nodes; Modularity Q represents the strength between groups after the nodes are divided into groups, and its formula is: where C i is the cluster name of i-node, δ(C i C j ) is 1 when i and j nodes belong to the same cluster, otherwise 0; k i is the importance degree of node i in the network; k j is the importance degree of node j in the network; Feature vector centrality reflects the importance of neighbor nodes, and its formula is as follows: Ae=λe Where, e = [e1, ..., e] i , ...e n Let λ represent the vector composed of the eigenvector centralities of all nodes, where A is the adjacency matrix and λ is the corresponding eigenvalue of the adjacency matrix. 5.The brain network optimization based automatic seizure detection device for severe encephalitis patients according to claim 1, wherein, When preprocessing the electroencephalogram signals to be detected and the electroencephalogram signal samples, the preprocessing module: Uses a band-pass filter to filter out frequency components above 48 Hz and below 1 Hz; Performs five-layer wavelet packet decomposition on the filtered electroencephalogram signals to obtain six sub-band signals of 1-4 Hz, 4-8 Hz, 8-12 Hz, 12-16 Hz, 16-24 Hz and 24-32 Hz. 6.The brain network optimization based automatic seizure detection device for severe encephalitis patients according to claim 1, wherein, The steps of independently optimizing the network layer and the feature layer based on the improved genetic algorithm comprise: Set the initial population size, evolution generation number, crossover probability and mutation probability; Randomly generate multiple network weighted coefficients of the network layer and multiple feature weighted coefficients of the feature layer; The network layer weighting coefficient optimization is performed to select multiple network weighting coefficients in a generation with the highest accuracy; a constraint condition is set as the sum of the multiple network weighting coefficients in the network layer being 1, an fitness function is set as maximizing the accuracy of the random forest classification model in classifying the test set, and the multiple network weighting coefficients are optimized through the crossover, recombination and mutation operations of the genetic algorithm; The feature layer coefficient optimization is performed to select the weighting coefficient in a generation with the highest accuracy; a constraint condition is set as the sum of the multiple feature weighting coefficients in the feature layer being 1, an fitness function is set as maximizing the accuracy of the random forest classification model in classifying the test set, and the weight is optimized through the crossover, recombination and mutation operations of the genetic algorithm; The above steps are repeated until the test set classification accuracy no longer changes, and the genetic algorithm optimization ends. 7.The brain network optimization based automatic seizure detection device for severe encephalitis patients according to claim 1, wherein, The sampling frequency of the collected electroencephalogram signals of the multiple severe encephalitis patients is 256 Hz, the electrode distribution adopts the international 10-20 electroencephalogram collection standard, and 20-channel electroencephalogram data is collected.

Images