Brain function abnormality prediction technology of tensor map neural network

Abstract

The invention discloses a brain function abnormity prediction technology based on a tensor graph neural network. The technology comprises a whole process algorithm model construction technology of five modules including data acquisition and labeling, data preprocessing, graph structure and graph tensor construction, tensor graph neural network construction, index monitoring and alarm. In addition, three graph structures are defined and constructed in the patent and used for mapping the electroencephalogram signals, the graph tensor constructed based on the graph structures can fuse the characteristics of the time sequence, the frequency domain and the space domain into the brain function abnormality prediction model, the topological structure of the brain neuron cluster and the relevance between the space and the time sequence characteristics are fully utilized, and the prediction accuracy of the brain function abnormality is improved. And the pathogenesis of brain dysfunction diseases can be more favorably explored.

Description

Technical Field

[0001] The present invention relates to the fields of neural network and artificial intelligence technology, and in particular to a brain function abnormality prediction technology based on tensor graph neural network. Background Art

[0002] At present, the commonly used technical models in the field of brain function abnormality prediction are mainly multi-layer artificial neural network algorithms, which mainly include multi-layer perceptron (Multilayer Perceptron, M1LP), convolutional neural networks (Convolutional Neural Networks, CNN), recurrent neural networks (Recurrent Neural Network, RNN), long short-term memory networks (Long Short-Term Memory, LSTM), Transformer and other neural networks and their variants.

[0003] The brain function abnormality prediction model based on MLP is too simple and has fewer model parameters, resulting in an inflexible model learning mode and poor generalization performance; the brain function abnormality prediction model based on CNN has characteristics such as local connection and weight sharing, but CNN is sensitive to the position and scale of the image in the input data, which means that when the position of the image changes, the model may not be able to correctly identify it, which is not friendly to the identification of physiological signals such as EEG signals. In addition, CNN ignores the time series characteristics of physiological data when brain function is abnormal, resulting in deficiencies in processing time series data; the brain function abnormality prediction model based on RNN, LSTM and Transformer can make full use of the time series characteristics of physiological data, but it ignores the spatial information of the data, and it is difficult to take advantage of the multi-dimensional data characteristics of the good data at the same time, and there is a risk of overfitting.

[0004] At the same time, brain neurons are distributed in a reticular pattern. Abnormal brain function is closely related to abnormal transmission of electrical signals in the brain neuron network, and different types of abnormal brain function diseases have a strong correlation with the activity of brain neurons at the lesion site. In the past, abnormal brain function prediction models often extracted EEG signal time domain, frequency domain and other features as model input for training, ignoring the topological structure of brain neuron clusters and the correlation between spatial and temporal features. Combining the non-Euclidean data structure characteristics of brain neuron distribution to construct a brain function abnormality prediction model is a new idea for studying brain function. Summary of the invention

[0005] In order to solve the problem that previous abnormal brain function prediction models ignored the topological structure of brain neuron clusters and the correlation between spatial and temporal features, the present invention discloses a brain function abnormality prediction technology based on tensor graph neural network, which can integrate the characteristics of time series, frequency domain and spatial domain into the brain function abnormality prediction model, fully utilizes the topological structure of brain neuron clusters and the correlation between spatial and temporal features, and is more conducive to exploring the pathogenesis of abnormal brain function diseases.

[0006] The technical solution extracted by the present invention mainly includes the following five modules: data collection and annotation, data preprocessing, construction of graph structure and graph tensor, construction of tensor graph neural network, monitoring indicators and alarms.

[0007] The data acquisition and annotation module is mainly used to collect and store brain neuron activities (the data is mainly in the form of EEG signals), annotate the collected EEG data according to the function and physiological stage, and divide the EEG signal into four stages: the abnormal brain function interval, the preictal period, the ictal period, and the postictal period. The annotation method used is as follows: a 5-minute time interval is reserved before the abnormal brain function to take preventive measures, and the data 5-35 minutes before the onset of abnormal brain function is marked as the preictal period; considering that there is no obvious boundary between the interictal period and the preictal period, the 4-hour interval can avoid the impact of unclear data annotation on the experimental results, so the data outside 4 hours from the start and end of each attack is marked as the interictal period. Take the advice of professional doctors and annotate the data according to the above method.

[0008] The data preprocessing module is mainly used to standardize the data processed by the previous module. Since the signal frequencies collected from different EEG signal sources may be inconsistent, and EEG signals with too high a frequency will increase the consumption of computing resources, it is necessary to perform unified downsampling. In addition, noise and artifacts need to be removed to improve the accuracy of EEG signal analysis. The continuous EEG signal is then divided into time slices using a sliding window. Considering that the data in the early stage of brain function abnormality and the interval of brain function abnormality are greatly unbalanced, overlapping time windows are used to divide the data in the early stage of brain function abnormality, and non-overlapping cutting is performed on the data in the interval of brain function abnormality. Finally, the divided data segments are processed by fast Fourier transform to better extract frequency domain features.

[0009] This module defines three graph structures for graph processing of EEG signals so that the brain function abnormality prediction model can make full use of the topological structure of the brain's neuron clusters.

[0010] Specifically, the graph is defined as follows: G = {V, E, A}, where V represents the set of nodes in the graph structure. In this paper, each graph node represents a channel of the EEG signal. E represents the set of edges in the graph structure, which in this paper means the correlation between channel signals. A represents the adjacency matrix of the graph structure, which can represent the structural information of the graph.

[0011] Furthermore, the present invention defines a distance-based graph structure (distance graph). The graph structure regards the electrode channels during EEG sampling as nodes of the graph structure, calculates the physical distance between the channels as the edge length between the nodes, and constructs the graph structure. At the same time, considering the differences in the absolute distances of intracranial electrodes of different patients, the relative distance (percentage as distance) is taken as the weight of the edge when calculating the distance.

[0012] The present invention defines a cross correlation graph. The graph regards the electrode channels during EEG sampling as nodes of the graph structure, calculates the correlation between channels, and normalizes them as the edge length between nodes to construct the graph structure. The correlation definition formula is as follows:

[0013]

[0014] Where x and y are one-dimensional vectors, z is the cross-correlation calculation result of the two vectors, k∈(0, 1, ..., ||x||+||y||-2), ||x||, ||y|| are the lengths of the x and y vectors respectively, N=max(||x||, ||y||)), in the application scenario of this article, ||x||=||y||, and z is a constant.

[0015] NCC (Normalized Cross Correlation) is the normalized cross correlation, and its value is taken as the weight of the edge in the graph structure. Only the edges with the largest weights are retained for each node. The normalized cross correlation is defined as follows:

[0016]

[0017] The present invention defines a graph structure based on mutual information. The graph structure regards the electrode channels during EEG sampling as nodes of the graph structure, calculates the mutual information (MI) between the channels, and uses it as the edge length between the nodes to construct the graph structure. The advantage of mutual information over the correlation coefficient is that it can capture a wider range of relationship types, including linear and nonlinear relationships. It measures the amount of information shared between two variables. For variables X and Y, the calculation formula for calculating their mutual information is as follows:

[0018]

[0019] The absolute value is taken as the weight of the edge in the graph structure, and only the edges with the largest weights are retained for each node. In this paper, k is taken as 3, and a graph structure based on mutual information is constructed.

[0020] This module is mainly used to process the graph structure constructed in the previous module into a tensor form, which is convenient for use as the input of the tensor graph neural network. The definition of the graph tensor is as follows:

[0021] τ=(G 1, G2,…,G r )

[0022] Among them G i = {V i ,E i ,A i} represents the i-th graph in the graph tensor τ, i∈(1, 2, ..., r), and r represents the total number of graphs in the graph tensor.

[0023] The tensor graph neural network construction module mainly expands the ordinary graph convolutional neural network structure and defines the main structure of the tensor graph neural network. The definition of the ordinary graph convolutional neural network is as follows:

[0024]

[0025] in, is the feature matrix of the lth layer, A∈R n×n is the adjacency matrix of the graph, Where I is the identity matrix, d l is the number of features of each node in the first layer, is a matrix The diagonal node degree matrix of is the trainable weight matrix of this layer. Combining the intra-graph attention mechanism and the inter-graph attention mechanism, it is expanded into a tensor graph neural network as follows:

[0026]

[0027]

[0028]

[0029] The monitoring index and alarm module are mainly used to evaluate the degree of abnormal brain function and issue timely alarms. This patent uses commonly used evaluation indicators of classification models, such as accuracy (Acc), sensitivity (Sen), specificity (Spe), F1 score (F1-Score), etc. The calculation formula is as follows:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] Among them, TP is the number of pre-ictal samples predicted as pre-ictal samples; TN is the number of inter-ictal samples predicted as inter-ictal samples; FP is the number of inter-ictal samples predicted as pre-ictal samples; FN is the number of pre-ictal samples predicted as inter-ictal samples.

[0036] At the same time, in order to warn of abnormal brain function throughout the whole process, it is also necessary to use evaluation indicators for continuous processes in order to issue alarms. Here, the evaluation indicators based on the attack event, event sensitivity and false alarm rate, are used. Event sensitivity (sensitivity-event, Sen-E) refers to the ratio of the number of abnormal brain function accurately predicted by the alarm to the number of all abnormal brain function attacks. The false prediction alarm rate (FPR) refers to the number of false alarms per hour. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to better understand the drawings cited in the detailed description of the present invention, a brief description of each drawing is provided.

[0038] Figure 1 The present invention is a schematic diagram of a model for predicting abnormal brain function using a tensor graph neural network-based technology based on the technical concept of the present invention.

[0039] Figure 2 This is a schematic diagram of a partial process of preprocessing data of the pre-onset and inter-onset period of abnormal brain function according to an embodiment of the present invention.

[0040] Figure 3 A schematic diagram of a graph structure generated based on three graph structure generation methods proposed in the technology of the present invention according to an embodiment of the present invention

[0041] Figure 4 A schematic diagram of a tensor graph neural network according to an embodiment of the present invention DETAILED DESCRIPTION

[0042] The present invention is susceptible of various transformations and embodiments, and specific embodiments will be illustrated in the drawings and described in detail in the detailed description.

[0043] Figure 1 The present invention is a schematic diagram of a model for predicting abnormal brain function using a tensor graph neural network-based technology based on the technical concept of the present invention.

[0044] like Figure 1 As shown, the brain function abnormality prediction technology model based on tensor graph neural network disclosed in the present invention is divided into five modules: data collection and annotation, data preprocessing, construction of graph structure and graph tensor, construction of tensor graph neural network, monitoring indicators and alarms.

[0045] First, use EEG signal acquisition equipment to collect and store EEG signals. EEG signal acquisition methods are mainly divided into scalp EEG signal acquisition and intracranial EEG signal acquisition. The main method is: placing electrodes on the scalp / intracranial to record brain electrical activity. Usually a 10-20 electrode system or other system is used to standardize the position of the electrodes. The electrodes usually collect data at a rate of several hundred hertz per second to capture the rapid changes in brain neuronal activity. This acquisition method can monitor brain neuronal activity in real time, and store the acquisition results for subsequent data annotation processes.

[0046] After completing the data collection, the EEG signals are labeled according to the data labeling standards and divided into four labels: interictal period, preictal period, ictal period, and postictal period. The specific labeling method is as follows: a 5-minute time interval is reserved before the abnormal brain function to take preventive measures, and the data 5-35 minutes before the onset of abnormal brain function is labeled as preictal period; considering that there is no obvious boundary between the interictal period and the preictal period, the 4-hour interval can avoid the impact of unclear data labeling on the experimental results, so the data outside 4 hours from the start and end of each attack is labeled as interictal period. Take the advice of professional physicians and label the data according to the above method.

[0047] After being processed by the data acquisition and annotation modules, the original data that can be used to build a model for brain function abnormality prediction technology based on tensor graph neural network is obtained. In order to further improve the data quality and model effect, the original data is preprocessed. The preprocessing process is as follows Figure 2 shown.

[0048] like Figure 2 As shown, in order to reduce the amount of calculation, the data is uniformly downsampled to 256Hz.

[0049] like Figure 2 As shown, a Butterworth bandpass filter with a frequency range of 0.1-64 Hz is used to remove high frequency noise.

[0050] like Figure 2 As shown in the figure, a sliding window is used to divide the continuous EEG signal into time slices, and the sliding window size is 30s.

[0051] like Figure 2 As shown, the data of the early stage of brain function abnormality and the period between brain function abnormality are extremely unbalanced, so the data of the early stage of brain function abnormality are divided into overlapping time windows with an overlapping rate of 25%, and the data of the period between brain function abnormality are cut without overlapping.

[0052] The segmented data segments are processed by fast Fourier transform to better extract frequency domain features.

[0053] like Figure 3 As shown, according to the generation method of three graph structures proposed by the present invention, three graph structures are generated, namely: and spliced ​​into graph tensor form as the input of the model of brain function abnormality prediction technology based on tensor graph neural network.

[0054] like Figure 4 As shown in the figure, the intra-graph attention mechanism and the inter-graph attention mechanism are introduced to build a model of brain function abnormality prediction technology based on tensor graph neural network. The intra-graph attention mechanism focuses on the relationship between nodes within a single graph, and can effectively learn the local features and interdependencies between nodes. The inter-graph attention mechanism focuses on the relationship between different graphs, and can learn global features at the graph level, such as the similarity and difference between graphs. The graph tensor model integrates multiple graphs into one tensor, which can more comprehensively express the relationship between graphs and use the inter-graph attention mechanism to learn the interaction between different graphs. The brain function abnormality prediction model built based on the tensor graph neural network can capture the graph structure information more comprehensively, make full use of the topological structure of the brain neuron cluster and the correlation between spatial and temporal features, and is more conducive to exploring the pathogenesis of brain function abnormalities.

Claims

1. A brain function abnormality prediction technology based on tensor graph neural network, characterized in that It includes the following five modules: data collection and annotation, data preprocessing, construction of graph structure and graph tensor, construction of tensor graph neural network, monitoring indicators and alarms. It makes full use of the topological structure of brain neuron clusters and the correlation between spatial and temporal characteristics, which is more conducive to exploring the pathogenesis of abnormal brain function diseases. The data acquisition and annotation module is mainly used to collect and store brain neuron activities, and annotate the collected EEG data according to the functions and physiological stages based on the annotation standards. The data preprocessing module is mainly used to standardize the data processed by the previous module. The tensor graph neural network construction module mainly combines the intra-graph attention mechanism and the inter-graph attention mechanism to expand the ordinary graph convolutional neural network structure and defines the main structure of the tensor graph neural network. The monitoring index and alarm module are mainly used to evaluate the degree of abnormal brain function and issue timely alarms.

2. The brain function abnormality prediction technology based on tensor graph neural network according to claim 1, characterized in that: It also includes data collection and annotation modules, which divide the data into types and annotation standards after data collection.

3. The brain function abnormality prediction technology based on tensor graph neural network according to claim 1, characterized in that: It also includes data preprocessing module, data preprocessing process and steps, including downsampling, bandpass filtering to remove noise, using sliding windows to cut into time slices, overlapping time windows to cut into equalization, and fast Fourier transform processing.

4. The brain function abnormality prediction technology based on tensor graph neural network according to claim 1, characterized in that: It also includes methods for extracting three graphical EEG signals and designs three graph structures: a distance-based graph structure, a correlation-based graph structure, and a mutual information-based graph structure. The three graph structures are combined in the form of graph tensors to organically integrate the characteristics of time series, frequency domain, and spatial domain.

5. The brain function abnormality prediction technology based on tensor graph neural network according to claim 4 is characterized in that: It also includes three methods for constructing graph structures. The electrode channels during EEG sampling are regarded as nodes of the graph structure. The distance, correlation, and mutual information between channels are calculated respectively. After appropriate processing, they are used as the edge lengths between nodes to construct the graph structure.

6. The brain function abnormality prediction technology based on tensor graph neural network according to claim 1, characterized in that: It also includes the construction method of tensor graph neural network, combining the intra-graph attention mechanism and the inter-graph attention mechanism, taking the graph tensor information as input, extracting and fusing the features of time series, frequency domain and spatial domain, and applying it to the field of brain function abnormality prediction.

Images