A method and system for normalized modeling of EEG multi-index network

Abstract

The present invention discloses a method for normalized modeling of EEG multi-index network, comprising the following steps: S1: using a large-scale, full-life cycle multi-country EEG cross-spectrum data set as the original EEG data, calculating the cross-spectrum matrix of the original EEG data, and preprocessing the original EEG cross-spectrum matrix; S2: using the preprocessed original EEG cross-spectrum matrix to construct a weighted undirected network as a brain function network; S3: calculating the network characteristics of the three dimensions of brain network function integration, function separation and centrality of the brain function network, and establishing the evolution trajectory of the network characteristics; S4: designing a lightweight neural network regression model with interpretability, reconstructing the upper triangular elements of the brain function network to obtain a complete brain function network, that is, a normalized brain function network corresponding to this age. A system for normalized modeling of EEG multi-index network is also disclosed. The present invention can more accurately reflect the principle of the brain function network.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal processing, and in particular to an EEG multi-index network normalization modeling method and system. Background Art

[0002] Normative modeling is an emerging research topic in population neuroscience that seeks to understand population variation at the individual level rather than at the average level. Much like height or weight growth charts in pediatric medicine, normative modeling establishes relationships between neuroimaging features and demographic variables such as age or sex. Deviations are then calculated to quantify individual differences for the prevention, diagnosis, and treatment of mental disorders.

[0003] The structure and function of the brain are constantly changing throughout a person's life. Establishing a standardized evolutionary trajectory is important for understanding human development and medical disease diagnosis. Although a large number of studies have conducted normative modeling of neuroimaging features, most studies have focused on modeling a single attribute of changes in brain structure or function. From the perspective of normative modeling of brain structure, Zabihi et al. estimated a longitudinal normative model of cortical thickness and then plotted individual deviations from the typical pattern; Wolfers et al. predicted the normative evolutionary trajectory of gray matter volume in adults by age and gender.

[0004] Bethlehem et al. mapped a precise brain map of gray matter and white matter volume changes in a large cohort. Subsequently, Rutherford et al. used twisted Bayesian regression to validate medical data, showing significant clinical value. From the perspective of normative modeling of brain function, Sun et al. established a normative model of functional connectivity strength throughout the lifespan, revealing individual heterogeneity in patients with major depression. Rutherford et al. established a normative model of functional connectivity and studied the differences between schizophrenia and healthy controls. Gao et al. revealed the topological structure and growth trajectory of nine cortical functional connections in the first 2 years of life, indicating that the higher-order networks of newborns are topologically incomplete and isolated. Khundrakpam et al. found that in late childhood, topological features changed significantly, especially local efficiency and modularity decreased significantly, while global efficiency increased significantly. Supekar et al. found that the default mode network (DMN) underwent significant developmental changes in functional and structural connectivity during adolescence. In addition, Cao et al. found that modularity decreased linearly, while local efficiency and rich club structure showed an inverted U-shaped trajectory in the age range of 7-85 years.

[0005] There are already many methods for standardized modeling, such as Gaussian process regression, warped Bayesian regression, GAMLSS regression, etc. They can be applied to single attributes, but there are still many difficulties in standardized modeling of the entire brain network. Therefore, developing a tool for standardized modeling of brain networks at the whole-brain level is crucial for understanding brain development and the diagnosis and treatment of mental diseases. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for standardized modeling of electroencephalogram multi-index networks. By combining professional knowledge of electroencephalogram signal processing, statistical analysis, and machine learning methods, a tool for standardized modeling of brain functional networks at the whole-brain level with interpretability is developed, and the modeling process with interpretability enables it to more accurately reflect the principle of brain functional networks.

[0007] To solve the above technical problem, the first technical solution adopted by the present invention is: to provide a method for standardized modeling of electroencephalogram multi-index networks, including the following steps:

[0008] S1: Using a large-scale multi-country electroencephalogram cross-spectrum dataset covering the whole life cycle as the original electroencephalogram data, calculating the cross-spectrum matrix of the original electroencephalogram data, and preprocessing the original electroencephalogram cross-spectrum matrix;

[0009] S2: Using the preprocessed original electroencephalogram cross-spectrum matrix to construct a weighted undirected network as the brain functional network;

[0010] S3: Calculating the network characteristics of three dimensions, namely brain network functional integration, functional separation, and centrality of the brain functional network, and establishing the evolutionary trajectory of the network characteristics;

[0011] S4: Designing an interpretable lightweight neural network regression model, where the input of the model is age and the network feature values under the corresponding evolutionary trajectory, and the output is the upper triangular elements of the brain functional network at this age. Reconstructing the upper triangular elements of the brain functional network to obtain a complete brain functional network, that is, the standardized brain functional network corresponding to this age.

[0012] In a preferred embodiment of the present invention, in step S1, the steps of preprocessing the original electroencephalogram cross-spectrum matrix include:

[0013] S101: For the original electroencephalogram cross-spectrum matrix Perform average reference, and the average reference formula is:

[0014] ;

[0015] Among them, is the frequency, is the average reference transformation matrix:

[0016] ;

[0017] Where Nc is the number of electrodes used in the EEG data.

[0018] S102: Globally adjust the entire EEG signal by multiplying it by a global gain factor , to change the amplitude of the signal, and use the cross-spectral matrix corrected by the global gain factor for:

[0019] .

[0020] In a preferred embodiment of the present invention, in step S2, the brain functional network uses the coherence matrix as the functional network matrix, the electrodes as nodes, and the pairwise coherence values ​​to represent the connectivity strength between the electrodes to establish a weighted undirected network, and the coherence is obtained by normalizing the cross spectrum. The coherence matrix under frequency is:

[0021] ;

[0022] in, represents the cross spectral density of x and y, and Represents the power spectral density of signals x and y. The coherence value is normalized to satisfy the range of 0 to 1, where 1 means that the two signals are completely coherent and 0 means they are incoherent.

[0023] In a preferred embodiment of the present invention, in step S3, the network characteristics include characteristic path length, global efficiency, clustering coefficient, local efficiency, modularity coefficient, betweenness centrality, and participation coefficient.

[0024] In a preferred embodiment of the present invention, in step S3, the evolution trajectory of the network characteristics is constructed using the generalized additive model GAMLSS method, and several percentile lines are set to understand the distribution of the evolution trajectory at different percentiles.

[0025] In a preferred embodiment of the present invention, in step S4, the network structure of the lightweight neural network regression model consists of an input layer, three hidden layers h1, h2, h3 and an output layer, the input is the number of neurons of the network characteristics under the age and the corresponding evolution trajectory, the number of neurons in the first hidden layer h1 indicates that each input element will correspond to a parameter to learn the relationship with each output element, the number of neurons in the second hidden layer h2 is determined by correlation analysis between input and output, and the number of neurons in the third hidden layer h3 is the number of triangular elements on the brain function network.

[0026] In a preferred embodiment of the present invention, in step S4, after the lightweight neural network regression model is designed, the large-scale, full-life cycle, multi-country EEG cross-spectrum data set is divided into training data and test data in proportion, and put into the lightweight neural network regression model for training and evaluation.

[0027] In a preferred embodiment of the present invention, in step S4, the age and the network characteristic value at a certain quantile after the corresponding generalized additive model GAMLSS are modeled are input into the trained lightweight neural network regression model to obtain a normalized brain function network for the age.

[0028] In order to solve the above technical problems, the second technical solution adopted by the present invention is to provide an EEG multi-index network normalized modeling system using the EEG multi-index network normalized modeling method as described above, comprising:

[0029] The EEG data preprocessing module is used to use a large-scale, full-life cycle, multi-country EEG cross-spectrum data set as the original EEG data, calculate the cross-spectrum matrix of the original EEG data, and preprocess the original EEG cross-spectrum matrix;

[0030] A brain function network construction module is used to construct a weighted undirected network using the preprocessed raw EEG cross-spectrum matrix as a brain function network;

[0031] A brain network feature evolution trajectory module is used to calculate the network features of the brain function network in three dimensions: brain network function integration, function separation and centrality, and establish the evolution trajectory of the network features;

[0032] The normalized modeling module of the brain functional network is used to design an interpretable lightweight neural network regression model. The input of the model is the age and the network characteristic values ​​under the corresponding evolutionary trajectory. The output is the upper triangular elements of the brain functional network at this age. The upper triangular elements of the brain functional network are reconstructed to obtain a complete brain functional network, that is, the normalized brain functional network corresponding to this age.

[0033] In order to solve the above technical problems, the third technical solution adopted by the present invention is: to provide an application of the EEG multi-index network normalization modeling system as described above in the field of diagnosis of mental illnesses.

[0034] The beneficial effects of the present invention are:

[0035] (1) This invention combines EEG signal processing expertise, statistical analysis, and machine learning methods to develop an interpretable full-brain level brain functional network standardization modeling tool, which makes up for the limitation of existing standardization modeling methods that can only model a single attribute. The interpretable modeling process enables it to more accurately reflect the principles of brain functional networks;

[0036] (2) The present invention has been verified on self-collected data sets (HY-AHU) and open source data sets (BrainLat), making it applicable in the field of disease diagnosis. This tool breaks through the limitations of machine learning neural networks as black box models. By incorporating professional knowledge in the field of EEG signal processing, the tool is interpretable and can provide a deeper insight into the construction mechanism of brain networks and the process of brain development. It also provides a more accurate and convincing basis for the diagnosis of mental illnesses.

[0037] (3) The invention establishes a normalized brain network model at the whole brain level, which can reveal the interaction between neurons and help us to deeply understand the functions of the brain in cognition, perception, learning, etc. At the same time, by establishing a normalized brain network model at the whole brain level, we can more accurately examine the impact of diseases on the brain neural network, thereby better understanding the pathogenesis of diseases and improving the accuracy and predictability of diagnosing mental illnesses, making certain contributions to the medical field. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the EEG multi-index network normalization modeling method of the present invention;

[0039] Figure 2 is an architecture diagram of the lightweight neural network;

[0040] Figure 3 is the individual deviation map validated across datasets;

[0041] Figure 4 It is a structural block diagram of the EEG multi-index network normalization modeling system. DETAILED DESCRIPTION

[0042] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0043] See also Figure 1 , the embodiment of the present invention includes:

[0044] A method for normalizing modeling of EEG multi-index network includes the following steps:

[0045] S1: A large-scale, full-life multinational cross-spectra dataset (MN-CS) is used as the original EEG data to calculate the cross-spectra matrix of the original EEG data and pre-process the original EEG cross-spectra matrix.

[0046] The MN-CS dataset includes 1966 healthy subjects from 14 sites in 9 countries, covering almost the entire life cycle, with an even gender distribution. The original EEG recording is based on the 10 / 20 international electrode placement system, with a total of 19 electrodes: Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3 / T7, T4 / T8, T5 / P7, T6 / P8, Fz, Cz, Pz. The frequency range is 1.17-19.14Hz, with an interval of 0.39Hz.

[0047] The process of preprocessing the original EEG cross spectrum matrix includes:

[0048] S101: Raw EEG cross-spectrum matrix An average reference is performed to eliminate or reduce the effects caused by improper position or poor contact of a single electrode. The average reference formula is:

[0049]

[0050] in, is the frequency, is the average reference transformation matrix:

[0051]

[0052] Where Nc is the number of electrodes used in the EEG data.

[0053] S102: Due to the differences in amplifiers, recording conditions and equipment during the acquisition process, the overall amplitude of the actual EEG signal may vary greatly. In order to ensure that the amplitude of the signal is within the appropriate range, it is necessary to globally adjust the entire EEG signal by multiplying it by a global gain factor , to change the amplitude of the signal, and use the cross-spectral matrix corrected by the global gain factor for:

[0054]

[0055] S2: Use the preprocessed raw EEG cross-spectrum matrix to construct a weighted undirected network as the brain functional network;

[0056] The cross spectrum is the representation of two signals in the frequency domain, reflecting the relationship between the two signals at different frequencies. By calculating the cross spectrum of two signals, their coherence at different frequencies can be obtained. Coherence is an indicator to measure the synchronization of two signals in a specific frequency band. In EEG research, coherence reflects the multivariate relationship between multiple signals in the frequency domain through Fourier transform. The coherence matrix is ​​used as a functional network matrix, with electrodes as nodes and pairwise coherence values ​​to represent the connectivity strength between electrodes, to establish a weighted undirected network. The coherence is obtained by normalizing the cross spectrum. The coherence matrix under frequency is:

[0057]

[0058] in, represents the cross spectral density of x and y, and Represents the power spectral density of signals x and y. The coherence value is normalized to satisfy the range of 0 to 1, where 1 means that the two signals are completely coherent and 0 means they are incoherent.

[0059] S3: Calculate the network characteristics of the brain functional network in three dimensions: brain network functional integration, functional separation and centrality, and establish the evolution trajectory of the network characteristics; the specific steps include:

[0060] S301: Calculation of network characteristics;

[0061] In this example, seven network features of three dimensions, namely, functional integration (characteristic path length, global efficiency), functional separation (clustering coefficient, local efficiency, modularity coefficient), and centrality (betweenness centrality, participation coefficient) of the brain network were calculated. Specifically, characteristic path length (CPL) refers to the average length of all shortest paths between all node pairs, which indicates the efficiency of information transmission in the entire network. The smaller the value, the higher the efficiency. Global efficiency (GE) is inversely proportional to the average shortest path length and is used to measure the communication efficiency of the network. Clustering coefficient (CC) is the ratio of the actual number of connected edges to the most likely number of connected edges, which is used to quantify the neighbor strength of connected nodes in the network as a local network feature. Local efficiency (LE) measures the local information transmission capacity of the network and reflects, to a certain extent, the ability of the network to resist random attacks. Modularity coefficient (M) is the degree to which the network can be subdivided into such clearly divided and non-overlapping groups. Betweenness centrality (BC) measures the bridge role of nodes in the network. Nodes with high betweenness centrality are key nodes for the flow of information and resources. Nodes with high participation coefficient (PC) can promote global modular integration. All network features were calculated using scripts from the Brain Connectivity Tool (BCT).

[0062] S302: Evolution of network characteristics;

[0063] The generalized additive model (GAMLSS) provides a comprehensive description of the data distribution by constructing a flexible regression model for multiple distribution parameters of the response variable and allows modeling of nonlinear relationships within the data. The evolutionary trajectories of the frequency bands. Using logarithmically transformed age as the independent variable, cubic smoothing splines and the Box-Cox-t distribution family were used to establish the evolutionary trajectories of the seven network features. In addition, the 5%, 25%, 50%, 75%, and 95% percentile lines were set to better understand the distribution of the evolutionary trajectories at different percentiles.

[0064] S4: A lightweight neural network regression model (LNN) with interpretability is designed. The input of the model is the age and the network feature value under the corresponding evolution trajectory. The output is the upper triangular elements of the brain function network under this age. The upper triangular elements of the brain function network are reconstructed to obtain a complete brain function network, that is, the normalized brain function network corresponding to this age. The specific steps include:

[0065] S401: Design lightweight neural network regression model LNN;

[0066] like Figure 2 As shown, the lightweight neural network regression model LNN consists of an input layer, three hidden layers (h1, h2, h3) and an output layer, and the determination of the number of neurons in each layer has a certain significance. The goal of the modeling of the present invention is to use the actual age of the subject and the age-related network characteristics to predict the entire brain function network matrix. Therefore, the input layer of LNN is eight neurons, including age and seven network eigenvalues. The output layer is 171 neurons, corresponding to the number of elements in the triangle on the brain function network, and a complete brain function network can be obtained after reconstruction. LNN contains three hidden layers (h1, h2, h3) with ReLU activation function. The first hidden layer h1 has 1368 neurons (8*171), indicating that each input element will correspond to a parameter to learn the relationship with each output element; the number of neurons in the second hidden layer h2 is determined by correlation analysis between input and output. Specifically, by calculating the Pearson correlation coefficient of 8 input elements and 171 output elements and performing a significance test, for the input element matrix X The columns and output elements in the matrix Y , the calculation formula of Pearson correlation coefficient (rho) is:

[0067]

[0068] in, and denote the i-th observation value of variables a and b respectively, represents the j-th observation value of variable b, and denote the mean values ​​of variables a and b respectively, , , n represents the number of observations. According to statistics, the number of elements with strong correlation between input and output elements (correlation>0.6 and p value<0.05) is 325, so the number of neurons in h2 is 325; the number of neurons in the third hidden layer h3 is 171, which is the number of triangular elements on the brain function network.

[0069] S402: Establish a normalized model of brain function network;

[0070] After designing the lightweight neural network, the MN-CS dataset was divided into training data and test data in a ratio of 7:3 and put into LNN for training evaluation. LNN uses the Adam optimization algorithm to adaptively adjust the learning rate of each parameter and uses five-fold cross validation to evaluate the generalization ability of the model. , mean absolute error (MAE), and root mean square error (RMSE) are used to evaluate the model performance.

[0071] Then, the seven network feature values ​​at the 50% quantile after age and corresponding GAMLSS modeling are input into LNN to obtain the normalized brain function network at that age. Similarly, the brain function network at the corresponding quantile can be obtained by inputting the 5%, 25%, 75%, and 95% network feature values ​​into LNN, which can more accurately calculate individual deviations.

[0072] The present invention establishes a normalized brain functional network at the whole-brain level based on the large-scale health dataset MN-CS and LNN. In order to verify the accuracy of the established normalized model, 58 healthy young subjects from Anhui University (self-collected dataset HY-AHU) were collected, and all informed consent forms were signed by the participants. In order to apply it in the field of diagnosis of mental illnesses, it is necessary to obtain the open source dataset BrainLat, which includes 530 patients with neurodegenerative diseases, including Alzheimer's disease (AD), behavioral variant frontotemporal dementia (bvFTD), multiple sclerosis (MS), Parkinson's disease (PD) and 250 healthy controls (HC). This example performed a unified preprocessing of all data, and calculated their brain functional networks and network characteristics. Subsequently, the individual deviations between the subjects and the established normative brain network were calculated from the two perspectives of mean functional connectivity strength (MFCS) and network characteristics, and statistical analysis was performed. The results show that, Figure 3As shown, the normalized model established by the present invention has a high accuracy, and the deviation of healthy subjects (indicated by green) is significantly smaller than that of the disease group (indicated by red). Therefore, this normalized model has been proven to be applicable to the diagnosis of mental illnesses, and can help diagnose mental illnesses more accurately, diagnose them earlier, and treat them earlier.

[0073] See also Figure 4 The present invention also provides an EEG multi-index network normalization modeling system, including:

[0074] The EEG data preprocessing module is used to use a large-scale, full-life cycle, multi-country EEG cross-spectrum data set as the original EEG data, calculate the cross-spectrum matrix of the original EEG data, and preprocess the original EEG cross-spectrum matrix;

[0075] A brain function network construction module is used to construct a weighted undirected network using the preprocessed raw EEG cross-spectrum matrix as a brain function network;

[0076] A brain network feature evolution trajectory module is used to calculate the network features of the brain function network in three dimensions: brain network function integration, function separation and centrality, and establish the evolution trajectory of the network features;

[0077] The normalized modeling module of the brain functional network is used to design an interpretable lightweight neural network regression model. The input of the model is the age and the network characteristic values ​​under the corresponding evolutionary trajectory. The output is the upper triangular elements of the brain functional network at this age. The upper triangular elements of the brain functional network are reconstructed to obtain a complete brain functional network, that is, the normalized brain functional network corresponding to this age.

[0078] An EEG multi-index network normalized modeling system in this example can execute an EEG multi-index network normalized modeling method provided by the present invention, can execute any combination of implementation steps of the method example, and has the corresponding functions and beneficial effects of the method.

[0079] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for normalizing EEG multi-index network modeling, characterized in that: The following steps are involved: S1: A large-scale, full-life, multi-country EEG cross-spectrum dataset is used as the original EEG data, the cross-spectrum matrix of the original EEG data is calculated, and the original EEG cross-spectrum matrix is ​​preprocessed; The preprocessing steps include: S101: Raw EEG cross-spectrum matrix Perform average reference, the average reference formula is: ； in, is the frequency, is the average reference transformation matrix: ； Where Nc is the number of electrodes of the EEG data used; S102: Globally adjust the entire EEG signal by multiplying it by a global gain factor , to change the amplitude of the signal, using the cross-spectral matrix corrected by the global gain factor for: ； S2: Use the preprocessed raw EEG cross-spectrum matrix to construct a weighted undirected network as the brain functional network; S3: Calculate the network characteristics of the brain functional network in three dimensions: brain network functional integration, functional separation and centrality, and establish the evolution trajectory of the network characteristics; S4: Design an interpretable lightweight neural network regression model, the input of the model is the network feature value under the age and the corresponding evolutionary trajectory, the output is the upper triangular elements of the brain function network under this age, and the upper triangular elements of the brain function network are reconstructed to obtain a complete brain function network, that is, a normalized brain function network corresponding to this age; the network structure of the lightweight neural network regression model consists of an input layer, three hidden layers h1, h2, h3 and an output layer, the input is the number of neurons of the network feature under the age and the corresponding evolutionary trajectory, the number of neurons in the first hidden layer h1 means that each input element will correspond to a parameter to learn the relationship with each output element, the number of neurons in the second hidden layer h2 is determined by the correlation analysis between the input and output, and the number of neurons in the third hidden layer h3 is the number of upper triangular elements in the brain function network.

2. The EEG multi-index network normalization modeling method according to claim 1 is characterized in that: In step S2, the brain functional network uses the coherence matrix as the functional network matrix, the electrodes as nodes, and the pairwise coherence values ​​to represent the connectivity strength between the electrodes to establish a weighted undirected network. The coherence is obtained by normalizing the cross spectrum. The coherence matrix under frequency is: ； in, represents the cross spectral density of x and y, and Represents the power spectral density of signals x and y. The coherence value is normalized to satisfy the range of 0 to 1, where 1 means that the two signals are completely coherent and 0 means they are incoherent.

3. The EEG multi-index network normalization modeling method according to claim 1 is characterized in that: In step S3, the network characteristics include characteristic path length, global efficiency, clustering coefficient, local efficiency, modularity coefficient, betweenness centrality, and participation coefficient.

4. The EEG multi-index network normalization modeling method according to claim 1 is characterized in that: In step S3, the evolution trajectory of the network feature is constructed using the generalized additive model GAMLSS method, and several percentile lines are set to understand the distribution of the evolution trajectory at different percentiles.

5. The EEG multi-index network normalization modeling method according to claim 1 is characterized in that: In step S4, after the lightweight neural network regression model is designed, the large-scale, full-life cycle, multi-country EEG cross-spectrum data set is divided into training data and test data in proportion, and put into the lightweight neural network regression model for training and evaluation.

6. The EEG multi-index network normalization modeling method according to claim 1 is characterized in that: In step S4, the age and the network characteristic value at a certain quantile after the corresponding generalized additive model GAMLSS are modeled are input into the trained lightweight neural network regression model to obtain the normalized brain function network at the age.

7. An EEG multi-index network normalized modeling system using the EEG multi-index network normalized modeling method according to any one of claims 1 to 6, characterized in that: include: The EEG data preprocessing module is used to use a large-scale, full-life cycle, multi-country EEG cross-spectrum data set as the original EEG data, calculate the cross-spectrum matrix of the original EEG data, and preprocess the original EEG cross-spectrum matrix; A brain function network construction module is used to construct a weighted undirected network using the preprocessed raw EEG cross-spectrum matrix as a brain function network; A brain network feature evolution trajectory module is used to calculate the network features of the brain function network in three dimensions: brain network function integration, function separation and centrality, and establish the evolution trajectory of the network features; The normalized modeling module of the brain functional network is used to design an interpretable lightweight neural network regression model. The input of the model is the age and the network characteristic values ​​under the corresponding evolutionary trajectory. The output is the upper triangular elements of the brain functional network at this age. The upper triangular elements of the brain functional network are reconstructed to obtain a complete brain functional network, that is, the normalized brain functional network corresponding to this age.

Images