A method for predicting cognitive function status based on functional subnetwork characteristics

By layering the brain functional network and using support vector regressors, the prediction uncertainty problem caused by ignoring modular features in traditional methods is solved, and the prediction accuracy of cognitive functional state is improved.

CN116712034BActive Publication Date: 2024-11-26CHANGZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310620857.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-30
Publication Date
2024-11-26
Estimated Expiration
2043-05-30

AI Technical Summary

Technical Problem

Traditional cognitive functional state prediction methods ignore the modular characteristics of the brain functional network, resulting in uncertainty in prediction results and the inability to effectively capture changes in cognitive functional state, especially in local modules that are not significant in the global network.

Method used

Using a prediction method based on the features of functional subnets, a functional network of the brain is constructed through functional magnetic resonance images, a molecular network is layered, topological attribute parameters are calculated, features with high correlation are screened, and a support vector regressor is used to optimize, and a functional subnet is established. Mapping relationship with cognitive function state.

Benefits of technology

The accuracy of cognitive functional state prediction is improved, and the differences in brain functional networks under different cognitive functional states are revealed through hierarchical clustering and statistical methods, which enhances the ability to capture changes in local modules.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116712034B_ABST
    Figure CN116712034B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for predicting cognitive function state based on functional subnetwork characteristics, comprising the following steps: obtaining a cognitive function state scale score of each subject; collecting functional magnetic resonance images of each subject and preprocessing the images to construct a brain function network; obtaining a hierarchical subnetwork by recursive clustering of the brain function network; screening out topological attribute parameters of two functional subnetworks with large correlation coefficients with cognitive function states, and using the area under the receiver operating characteristic curve corresponding to the two topological attribute parameters as a feature set of the functional subnetwork; finding the optimal value of the parameter in the Gaussian radial basis kernel function of a support vector regression machine; and fitting the feature set of the functional subnetwork and the cognitive function state scale score using support vector regression. The present invention helps to deeply understand the relationship between cognitive function states and functional subnetworks, and improve the accuracy of predicting cognitive function states.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of biomedical information, and in particular to a method for predicting cognitive function status based on functional subnetwork characteristics. Background Art

[0002] As an organ of thinking and consciousness, the brain has a very complex structure and function. Hundreds of billions of neurons form an intricate network of connectivity through synapses. The brain's executive functions, whether simple or not, always rely on extensive interactions between multiple brain regions, forming a complex functional network. Therefore, studying the brain from a network perspective provides us with a new perspective.

[0003] The continuous maturity of imaging technology and the rapid development of graph theory and complex network research have provided powerful tools for brain science research. The human brain can be divided into multiple brain regions. Abnormal changes in brain region functions will lead to changes in cognitive function states, and the connection relationship between brain regions will also change accordingly. The traditional method of predicting cognitive function states starts from the global brain function network, ignoring the modular characteristics of the brain function network; or focuses on studying the modularity at a single level, without extracting multiple levels of sub-networks to analyze cognitive function states.

[0004] Some studies have shown that in brain networks with different cognitive function states, not all brain regions have obvious differences. Changes in cognitive function states may only exist in certain local modules, which may not be significant in the global network. Ignoring the local modular structure of the brain function network will lead to uncertainty in the prediction results. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting cognitive function status based on functional subnetwork characteristics, so as to improve the accuracy of cognitive function status prediction.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] A method for predicting cognitive function state based on functional subnetwork characteristics, comprising:

[0008] Step 1: Obtain the cognitive function status scale score of each subject; collect functional magnetic resonance images of each subject and preprocess the images, and construct a brain functional network based on the preprocessed functional magnetic resonance images;

[0009] Step 2: Divide the brain functional network into multiple sub-networks, and obtain the sub-network of the next level by recursive clustering to subdivide each sub-network of the current level, and finally obtain a hierarchical sub-network;

[0010] Step 3: Use graph theory algorithms to calculate the values ​​of the topological attribute parameters of each functional subnetwork, and calculate the correlation coefficient between the area under the receiver operating characteristic curve corresponding to the topological attribute parameters and the cognitive function status scale;

[0011] Step 4: Use statistical methods to screen out topological attribute parameters with high correlation, and use the corresponding area under the receiver operating characteristic curve as the feature set of the functional subnetwork;

[0012] Step 5: Find the optimal value of the parameters in the Gaussian radial basis kernel function of the support vector regression machine, and optimize the support vector regression machine using the feature set;

[0013] Step 6: Use support vector regression to fit the functional subnetwork feature set and the cognitive function status scale score, and predict the cognitive function status based on the mapping relationship between the two.

[0014] Furthermore, step 1 specifically includes:

[0015] Step 1.1: All subjects were given the Montreal Cognitive Assessment Scale test to assess their cognitive function status;

[0016] Step 1.2: Use a gradient echo single-shot echo-planar imaging sequence to scan the cross-sectional brain of the subject to obtain a resting-state functional magnetic resonance image;

[0017] Step 1.3: Preprocess the functional magnetic resonance images. The preprocessing steps include: time correction, head motion correction, spatial registration, spatial normalization and smoothing filtering operations;

[0018] Step 1.4: Use the standard brain partition template to divide the brain regions, extract the blood oxygen level dependence intensity of each voxel in each brain region at different time points, and then perform arithmetic averaging on the time series of all voxels in each brain region to obtain the average time series of all voxels in each brain region;

[0019] Step 1.5: Use the Pearson correlation analysis method to calculate the Pearson correlation coefficient ρ of the average time series between brain regions X,Y :

[0020]

[0021] Where N represents the number of regions in the whole brain segmentation; X i represents the value of brain region X at the i-th time point, represents the average value of brain region X at all time points during the entire scanning period; Y i represents the value of brain area Y at the i-th time point, represents the average value of brain region Y at all time points during the entire scanning period;

[0022] Step 1.6: Get the correlation matrix of the average time series based on the correlation coefficient, and take the absolute value of the correlation matrix to get the brain functional connection matrix R:

[0023]

[0024] Among them, ρ X,Y represents the Pearson correlation coefficient of the average time series between brain region X and brain region Y.

[0025] Furthermore, step 2 specifically includes:

[0026] Step 2.1: For each subject’s brain functional network, a similarity graph G = (V, E) is used to represent it, where V = {v1, v2, …, v N} represents a vertex, E represents an edge, and the adjacency matrix of the graph is A = (a ij ), i=1,…N, j=1,…N represents the similarity between vertices; let N-dimensional vector U={u1,u2,…,u N} T is the label of N nodes in graph G, let A·U=λ·U, and get the eigenvalue λ of matrix A and the corresponding eigenvector U (v) (v=1,2,…,h), h is the number of subjects;

[0027] Step 2.2: Normalize the adjacency matrix: First, multiply the matrix A by the degree matrix D -1 , further split it into two D -1 / 2 , and obtain the symmetric and normalized matrix A1 (v) =D -1 / 2 AD -1 / 2 , where D is the diagonal matrix of node degrees on the diagonal (d1, d2, …, d N );In the traditional spectral clustering algorithm, one feature vector U is optimized each time (v) , k is the number of clusters, and the feature vector U (v) Regularized to the common centroid U (*) (v=1,2,…,h), each regularization term is composed of a weighting parameter λ reflecting the importance of vertex V v Weighted; the optimized objective function is:

[0028]

[0029] For U (v) Provide a standard spectral clustering and solve U( * )Specific eigenvectors:

[0030]

[0031] Using the cyclic property of matrix trace, we get:

[0032]

[0033] Step 2.3: Perform symmetric optimization again, and the improved normalized adjacency matrix is m represents the total number of similarity graphs; the eigenvector U is calculated by eigenvalue decomposition (v) , substituting into formula (6) to obtain U (*) , continuously update and iterate until convergence, and generate the final estimated feature vector and eigenvalue EU;

[0034]

[0035] Step 2.4: When performing hierarchical clustering, the network nodes are hierarchically divided into two clusters downstream of the current level according to the positive or negative value of the Fiedler vector; if the median value of the Fiedler eigenvalues ​​of all subjects in the group is positive in each clustering, the above steps are repeated, and the two clusters obtained at the current level are recursively sent to steps 2.1-2.3, otherwise the next step is stopped; by calculating the similarity between the two types of data points, the two most similar data points among all data points are combined, and this process is iterated repeatedly to form a top-down clustering tree;

[0036] Step 2.5: During continuous iterations, the convergence is monitored by the objective difference between continuous iterations. When the numerical difference in the objective function is lower than the minimum threshold, clustering stops, otherwise clustering continues. When the clustering stop condition is reached, the optimal solution of the iteration is obtained, which constitutes the segmentation of the graph V, and finally the clustering result of the functional subnetwork is obtained;

[0037] After the first division, the first-level division results are obtained, which are divided into low-order perceptual networks and high-order cognitive networks; after the second division, the second-level division results are obtained, in which the low-order perceptual network is divided into visual network and sensorimotor network, and the high-order cognitive network is divided into default mode network and attention network; these functionally specialized sub-networks can be further divided into finer sub-networks, and finally a tree structure of hierarchical sub-networks is obtained.

[0038] Furthermore, step 3 specifically includes:

[0039] Step 3.1: Use graph theory algorithms to calculate the values ​​of topological attribute parameters of each functional subnetwork, and calculate the area under the receiver operating characteristic curve corresponding to each topological attribute parameter, where the topological attribute parameters include degree, clustering coefficient, characteristic path length, small-world attribute, and node centrality;

[0040] Step 3.2: Use statistical methods to perform correlation analysis between the area under the receiver operating characteristic curve corresponding to the topological attribute parameters of each functional subnetwork and the cognitive function status scale score, and use the Pearson method to calculate the correlation coefficient r, as shown in the following formula:

[0041]

[0042] Among them, G represents the area under the receiver operating characteristic curve corresponding to the topological attribute parameter, W represents the cognitive function status scale score, σ G , σ W Represent the standard deviation of G and W respectively.

[0043] Furthermore, step 4 specifically includes:

[0044] Use the T test method to find the corresponding critical value in the distribution table. The formula is as follows:

[0045]

[0046] Among them, the t value is a statistic, and n represents the number of elements in the feature set;

[0047] According to the value of t, the distribution table is consulted to obtain the corresponding p-value. According to the significance level, whether the test result is significant is judged. If the p-value is less than the significance level, the null hypothesis is rejected, and it is considered that the selected topological attribute parameters have a significant effect on the cognitive function state scale score. Finally, the two topological attribute parameters with higher overall correlation are screened out, and the area under the receiver operating characteristic curve corresponding to these two topological attribute parameters is used as the feature collection of the functional sub-network.

[0048] Furthermore, step 5 specifically includes:

[0049] Step 5.1: Use the support vector machine model with Gaussian radial basis kernel function, and the decision function is as follows:

[0050]

[0051] Among them, ||x i -x|| 2 represents the square of the Euclidean distance between the model input and the center of the kernel function; h is the number of subjects, b is the bias parameter in the decision function; g is the characteristic parameter in the radial basis function; where the model input is the feature set of the functional subnetwork obtained in step 4;

[0052] Step 5.2: Find the optimal penalty factor C and characteristic parameter g: First, initialize the particle swarm and calculate the fitness of each particle. Then, in each iteration, the particle updates the data by tracking the two extreme values ​​pbest and gbest. When the global optimal position is reached, the global optimal solution is obtained.

[0053] Step 5.3: Train the support vector regression machine based on the feature set: set the size of the population and the number of iterations within a reasonable range of the algorithm, initialize the position and velocity of the population particles, and each individual is a group of (C, g) vectors; use the network mean square error of the support vector machine as the fitness function to calculate the fitness of the particle swarm in the particle swarm algorithm; bring the parameter value represented by each particle in each iteration process into the trained support vector machine model, and then use the test data to evaluate the network mean square error of the model, that is, the quality of the parameters.

[0054] Furthermore, step 6 specifically includes:

[0055] Step 6.1: Assume that the functional subnetwork feature set contains n elements, and the feature vector corresponding to each feature is represented as x i ={x i1 ,x i2 ,…,x ij}, where j is the feature dimension, x ij represents the value of the jth feature of the i-th subject; the target value of the corresponding cognitive function status scale is y i , then the training data set in the support vector machine model is expressed as:

[0056] T={(x1,y1),(x2,y2),…,(x i ,y i )},x i ∈R h ,y i ∈{1,-1}(i=1,2,…,h) (10)

[0057] The formula for the minimum distance from the support vector to the decision surface is obtained:

[0058]

[0059] Among them, ω is the normal vector of the hyperplane, ||ω|| represents the modulus of ω, and b is the intercept of the hyperplane;

[0060] Introducing the slack variable ξ i ,ξ i ≥0, the support vector machine is represented by y i (ω T x i +b)≥1-ξ i , which needs to be minimized The penalty factor C>0; the larger C is, the greater the sample penalty is; the smaller C is, the smaller the sample penalty is;

[0061] Step 6.2: Introduce the Lagrange multiplier method and obtain:

[0062]

[0063] Where ξ refers to the slack variable ξ i A vector of all values, a i ,μ i is the Lagrange multiplier, and α=(α1,α2,…,α h ) T ,μ=(μ1,μ2,…,μ h ) T , α and μ refer to the Lagrange multipliers a i ,μ i All the vectors of values; let L(ω,b,ξ,α,μ) be the value of ω,b,ξ i The partial derivative of is zero, so we can get Substituting the result into the above formula, we can get:

[0064]

[0065]

[0066] The above formula is used to solve the Lagrangian function. The goal is to make the partitioning hyperplane have the maximum interval and transform the convex quadratic programming problem of the maximum interval into the Lagrangian function solution problem; among them, when When 0≤α i ≤C;α j ,y j 、x j is another variable in the dual problem mapped to the high-dimensional space;

[0067] Solution Get the maximum value of the Lagrange multiplier Then we calculate:

[0068]

[0069] By introducing nonlinear mapping and replacing it with the kernel function, we get the decision function of the nonlinear support vector machine model, and its prediction value can be expressed as:

[0070]

[0071] Among them, K(x i ,x) is the Gaussian radial basis function, expressed as:

[0072]

[0073] Step 6.3: Use the features of the screened functional sub-network and import them into the support vector regression machine after parameter optimization to form a mapping relationship to predict the cognitive function status.

[0074] The present invention has the following beneficial effects:

[0075] The present invention hierarchically clusters the global network into sub-networks from the brain functional network, and provides a hierarchical network topology structure across multiple levels; performs multi-angle analysis on the cognitive function state in the sub-network, and more specifically describes the differences and similarities of the brain functional network under different cognitive function states; judges different cognitive function states according to the changes in the brain functional sub-network, and uses the optimized support vector regression machine to perform prediction, thereby improving the accuracy of the prediction of the cognitive function state. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of the implementation of a method for predicting cognitive function status based on functional subnetwork characteristics of the present invention. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] Please refer to Figure 1 The present invention is a method for predicting cognitive function state based on functional subnetwork characteristics, which comprises the following steps:

[0079] Step 1: Obtain the cognitive function status scale score of each subject; collect functional magnetic resonance images of each subject and preprocess the images, and construct the brain functional network based on the preprocessed functional magnetic resonance images; specifically including: Step 1.1: Before the functional magnetic resonance imaging examination, all subjects are tested with the Montreal Cognitive Function Assessment Scale to assess their cognitive function status; according to the scoring standard of the Montreal Cognitive Function Assessment Scale, the full score is 30 points, and a score of 26 points or above is considered normal, 20-26 points are considered mild cognitive impairment, and below 20 points are considered severe cognitive impairment. Step 1.2: Use a gradient echo single-shot echo planar imaging sequence to scan the cross-sectional brain of the subject to obtain a resting-state functional magnetic resonance image; in this embodiment, when using a gradient echo single-shot echo planar imaging sequence to scan the cross-sectional brain, the repetition time is 2000ms, the echo time is 40ms, the field of view is 240mm×240mm, the matrix size is 64×64, the layer thickness is 6mm, and the flip angle is 90°; the scanning time for each subject is 6min; before collecting data, ensure that the subject's physical condition is in a stable state, and remind the subject to stay as still as possible during the collection to avoid the influence of head movement on the collected data.

[0080] Step 1.3: Use the DPARSF toolbox to preprocess the functional magnetic resonance images to remove the effects of the equipment, the subject's head movement, and noise during the acquisition process; the preprocessing steps include: time correction, head movement correction, spatial registration, spatial normalization, and smoothing filtering operations;

[0081] Step 1.4: Use the standard brain partition template to divide the brain regions, extract the blood oxygen level dependence intensity of each voxel in each brain region at different time points, and then perform arithmetic averaging on the time series of all voxels in each brain region to obtain the average time series of all voxels in each brain region; In this embodiment, an internationally accepted anatomical marker template such as the AAL template is used for brain region segmentation; Step 1.5: Use the Pearson correlation analysis method to calculate the Pearson correlation coefficient ρ of the average time series between brain regions X,Y :

[0082]

[0083] Where N represents the number of regions in the whole brain segmentation; X i represents the value of brain region X at the i-th time point, represents the average value of brain region X at all time points during the entire scanning period; Y i represents the value of brain area Y at the i-th time point, represents the average value of brain region Y at all time points during the entire scanning period;

[0084] Step 1.6: Get the correlation matrix of the average time series based on the correlation coefficient, and take the absolute value of the correlation matrix to get the brain functional connection matrix R:

[0085]

[0086] Among them, ρ X,Y represents the Pearson correlation coefficient of the average time series between brain region X and brain region Y.

[0087] Step 2: Divide the brain functional network into multiple sub-networks, and obtain the sub-network of the next level by recursive clustering to subdivide each sub-network of the current level, and finally obtain a hierarchical sub-network; specifically, it includes:

[0088] Step 2.1: For each subject’s brain functional network, a similarity graph G = (V, E) is used to represent it, where V = {v1, v2, …, v N} represents a vertex, E represents an edge, and the adjacency matrix of the graph is A = (a ij ), i=1,…N, j=1,…N represents the similarity between vertices; let N-dimensional vector U={u1,u2,…,u N} Tis the label of N nodes in graph G, let A·U=λ·U, and get the eigenvalue λ of matrix A and the corresponding eigenvector U (v) (v=1,2,…,h), h is the number of subjects;

[0089] Step 2.2: Normalize the adjacency matrix: First, multiply the matrix A by the degree matrix D -1 , further split it into two D -1 / 2 , and obtain the symmetric and normalized matrix A1 (v) =D -1 / 2 AD -1 / 2 , where D is the diagonal matrix of node degrees on the diagonal (d1, d2, …, d N );In the traditional spectral clustering algorithm, one feature vector U is optimized each time (v) , k is the number of clusters, and the feature vector U (v) Regularized to the common centroid U (*) (v=1,2,…,h), each regularization term is composed of a weighting parameter λ reflecting the importance of vertex V v Weighted; the optimized objective function is:

[0090]

[0091] For U (v) Provide a standard spectral clustering solution U (*) Specific eigenvectors:

[0092]

[0093] Using the cyclic property of matrix trace, we get:

[0094]

[0095] The objective function is to regularize each subject-specific feature vector into a common centroid, making full use of the complementary information between subjects, and each regularization term is weighted by a weighting parameter reflecting the importance of vertex V.

[0096] Step 2.3: Perform symmetric optimization again, and the improved normalized adjacency matrix is m represents the total number of similarity graphs; the eigenvector U is calculated by eigenvalue decomposition (v) , substituting into formula (6) to obtain U (*) , continuously update and iterate until convergence, and generate the final estimated feature vector and eigenvalue EU;

[0097]

[0098] Step 2.4: When performing hierarchical clustering, the network nodes are hierarchically divided into two clusters downstream of the current level according to the positive or negative value of the Fiedler vector; if the median value of the Fiedler eigenvalues ​​of all subjects in the group is positive in each clustering, the above steps are repeated, and the two clusters obtained at the current level are recursively sent to steps 2.1-2.3, otherwise the next step is stopped; by calculating the similarity between the two types of data points, the two most similar data points among all data points are combined, and this process is iterated repeatedly to form a top-down clustering tree;

[0099] Step 2.5: During successive iterations, the convergence is monitored by the objective difference between successive iterations. When the numerical difference in the objective function falls below a minimum threshold ε = 10 -4 When the clustering stops, otherwise the clustering continues; when the clustering stop condition is reached, the optimal solution of the iteration is obtained, which constitutes the segmentation of the graph V, and finally the clustering result of the functional subnetwork is obtained;

[0100] After the first division, the first-level division results are obtained, which are divided into low-order perceptual networks and high-order cognitive networks; after the second division, the second-level division results are obtained, in which the low-order perceptual network is divided into visual network and sensorimotor network, and the high-order cognitive network is divided into default mode network and attention network; these functionally specialized sub-networks can be further divided into finer sub-networks, and finally a tree structure of hierarchical sub-networks is obtained.

[0101] Step 3: Use graph theory algorithms to calculate the values ​​of topological attribute parameters of each functional subnetwork, and calculate the correlation coefficient between the area under the receiver operating characteristic curve corresponding to the topological attribute parameters and the cognitive function status scale; specifically including:

[0102] Step 3.1: Use graph theory algorithms to calculate the values ​​of topological attribute parameters of each functional subnetwork, and calculate the area under the receiver operating characteristic curve corresponding to each topological attribute parameter, where the topological attribute parameters include degree, clustering coefficient, characteristic path length, small-world attribute, and node centrality;

[0103] Step 3.2: Use statistical methods to perform correlation analysis between the area under the receiver operating characteristic curve corresponding to the topological attribute parameters of each functional subnetwork and the cognitive function status scale score, and use the Pearson method to calculate the correlation coefficient r, as shown in the following formula:

[0104]

[0105] Among them, G represents the area under the receiver operating characteristic curve corresponding to the topological attribute parameter, W represents the cognitive function status scale score, σ G , σ W Represent the standard deviation of G and W respectively.

[0106] Step 4: Use statistical methods to screen out topological attribute parameters with high correlation, and use the corresponding area under the receiver operating characteristic curve as the feature set of the functional subnetwork; specifically, use the T test method to find the corresponding critical value in the distribution table, and the formula is as follows:

[0107]

[0108] Among them, the t value is a statistic, and n represents the number of elements in the feature set;

[0109] According to the value of t, the distribution table is consulted to obtain the corresponding p-value. According to the significance level, whether the test result is significant is judged. If the p-value is less than the significance level, the null hypothesis is rejected, and it is considered that the selected topological attribute parameters have a significant effect on the cognitive function status scale score. Finally, the two topological attribute parameters with higher overall correlation are screened out, and the area under the receiver operating characteristic curve corresponding to these two topological attribute parameters is used as the feature collection of the functional sub-network.

[0110] Step 5: Find the optimal value of the parameters in the Gaussian radial basis kernel function of the support vector regression machine, optimize the support vector regression machine using the feature set, optimize the prediction performance using the representative and discriminative feature set, improve the prediction accuracy of the support vector regression machine, and accelerate the convergence speed of the algorithm; specifically include:

[0111] Step 5.1: Use the support vector machine model with Gaussian radial basis kernel function, and the decision function is as follows:

[0112]

[0113] Among them, ||x i -x|| 2 represents the square of the Euclidean distance between the model input and the center of the kernel function; h is the number of subjects, b is the bias parameter in the decision function; g is the characteristic parameter in the radial basis function; where the model input is the feature set of the functional subnetwork obtained in step 4;

[0114] Step 5.2: Find the optimal penalty factor C and characteristic parameter g: First, initialize the particle swarm and calculate the fitness of each particle. Then, in each iteration, the particle updates the data by tracking the two extreme values ​​pbest and gbest. When the global optimal position is reached, the global optimal solution is obtained.

[0115] Step 5.3: Train the support vector regression machine based on the feature set: set the size of the population and the number of iterations within a reasonable range of the algorithm, initialize the position and velocity of the population particles, and each individual is a group of (C, g) vectors; use the network mean square error of the support vector machine as the fitness function to calculate the fitness of the particle swarm in the particle swarm algorithm; bring the parameter value represented by each particle in each iteration process into the trained support vector machine model, and then use the test data to evaluate the network mean square error of the model, that is, the quality of the parameters.

[0116] Step 6: Use support vector regression to fit the functional subnetwork feature set and the cognitive function status scale score, and predict the cognitive function status based on the mapping relationship between the two; specifically, it includes:

[0117] Step 6.1: Assume that the functional subnetwork feature set contains n elements, and the feature vector corresponding to each feature is represented as x i ={x i1 ,x i2 ,…,x ij}, where j is the feature dimension, x ij represents the value of the jth feature of the i-th subject; the target value of the corresponding cognitive function status scale is y i , then the training data set in the support vector machine model is expressed as:

[0118] T={(x1,y1),(x2,y2),…,(x i ,y i )},x i ∈R h ,y i ∈{1,-1}(i=1,2,…,h) (10)

[0119] The formula for the minimum distance from the support vector to the decision surface is obtained:

[0120]

[0121] Among them, ω is the normal vector of the hyperplane, ||ω|| represents the modulus of ω, and b is the intercept of the hyperplane;

[0122] Introducing the slack variable ξ i ,ξ i ≥0, the support vector machine is represented by y i (ω T x i +b)≥1-ξ i , which needs to be minimized The penalty factor C>0; the larger C is, the greater the sample penalty is; the smaller C is, the smaller the sample penalty is;

[0123] Step 6.2: Introduce the Lagrange multiplier method and obtain:

[0124]

[0125] Where ξ refers to the slack variable ξ i A vector of all values, a i ,μ i is the Lagrange multiplier, and α=(α1,α2,…,α h ) T ,μ=(μ1,μ2,…,μ h ) T , α and μ refer to the Lagrange multipliers a i ,μ i All the vectors of values; let L(ω,b,ξ,α,μ) be the value of ω,b,ξ i The partial derivative of is zero, so we can get Substituting the result into the above formula, we can get:

[0126]

[0127]

[0128] The above formula is used to solve the Lagrangian function. The goal is to make the partitioning hyperplane have the maximum interval and transform the convex quadratic programming problem of the maximum interval into the Lagrangian function solution problem; among them, when When 0≤α i ≤C;α j ,y j 、x j is another variable in the dual problem mapped to the high-dimensional space;

[0129] Solution Get the maximum value of the Lagrange multiplier Then we calculate:

[0130]

[0131] By introducing nonlinear mapping and replacing it with the kernel function, we get the decision function of the nonlinear support vector machine model, and its prediction value can be expressed as:

[0132]

[0133] Among them, K(x i ,x) is the Gaussian radial basis function, expressed as:

[0134]

[0135] Step 6.3: Use the feature set of the functional sub-network selected in step 4 and import it into the support vector regression machine after parameter optimization, i.e., formula (16), to form a mapping relationship to predict the cognitive function state.

[0136] The parts not involved in the present invention are the same as the prior art or are implemented by using the prior art.

[0137] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for predicting cognitive function status based on functional subnetwork characteristics, Features: include Step 1: Obtain the cognitive function status scale score of each subject; collect functional magnetic resonance images of each subject and preprocess the images, and construct a brain functional network based on the preprocessed functional magnetic resonance images; Step 2: Divide the brain functional network into multiple sub-networks, and obtain the sub-network of the next level by recursive clustering to subdivide each sub-network of the current level, and finally obtain a hierarchical sub-network; specifically, it includes: Step 2.1: For each subject’s brain functional network, a similarity graph G = (V, E) is used to represent it, where V = {v 1 ,v 2 ,…,v N } represents a vertex, E represents an edge, and the adjacency matrix of the graph is A = (a ij ), i = 1, ... N, j = 1, ... N represents the similarity between vertices; let N-dimensional vector U = {u 1 ,u 2 ,…,u N } T is the label of N nodes in graph G, let A·U=λ·U, and get the eigenvalue λ of matrix A and the corresponding eigenvector U (v) (v=1,2,…,h), h is the number of subjects; Step 2.2: Normalize the adjacency matrix: First, multiply the matrix A by the degree matrix D -1 , further split it into two D -1 / 2 , and obtain a symmetric and normalized matrix A 1 (v) =D -1 / 2 AD -1 / 2 , where D is the diagonal matrix of the node degrees on the diagonal (d 1 ,d 2 ,…,d N );In the traditional spectral clustering algorithm, one feature vector U is optimized each time (v) , k is the number of clusters, and the feature vector U (v) Regularized to the common centroid U (*) (v=1,2,…,h), each regularization term is composed of a weighting parameter λ reflecting the importance of vertex V v Weighted; the optimized objective function is: For U (v) Provide a standard spectral clustering and solve U( * )Specific eigenvectors: Using the cyclic property of matrix trace, we get: Step 2.3: Perform symmetric optimization again, and the improved normalized adjacency matrix is m represents the total number of similarity graphs; the eigenvector U is calculated by eigenvalue decomposition (v) , substituting into formula (6) to obtain U (*) , continuously update and iterate until convergence, and generate the final estimated feature vector and eigenvalue EU; Step 2.4: When performing hierarchical clustering, the network nodes are hierarchically divided into two clusters downstream of the current level according to the positive or negative value of the Fiedler vector; if the median value of the Fiedler eigenvalues ​​of all subjects in the group is positive in each clustering, the above steps are repeated, and the two clusters obtained at the current level are recursively sent to steps 2.1-2.3, otherwise the next step is stopped; by calculating the similarity between the two types of data points, the two most similar data points among all data points are combined, and this process is iterated repeatedly to form a top-down clustering tree; Step 2.5: During continuous iterations, the convergence is monitored by the objective difference between continuous iterations. When the numerical difference in the objective function is lower than the minimum threshold, clustering stops, otherwise clustering continues. When the clustering stop condition is reached, the optimal solution of the iteration is obtained, which constitutes the segmentation of the graph V, and finally the clustering result of the functional subnetwork is obtained; After the first division, the first-level division results are obtained, which are divided into low-order perceptual networks and high-order cognitive networks; after the second division, the second-level division results are obtained, in which the low-order perceptual network is divided into visual network and sensorimotor network, and the high-order cognitive network is divided into default mode network and attention network; these functionally specialized sub-networks can be further divided into finer sub-networks, and finally a tree structure of hierarchical sub-networks is obtained; Step 3: Use graph theory algorithms to calculate the values ​​of the topological attribute parameters of each functional subnetwork, and calculate the correlation coefficient between the area under the receiver operating characteristic curve corresponding to the topological attribute parameters and the cognitive function status scale; Step 4: Use statistical methods to screen out topological attribute parameters with high correlation, and use the corresponding area under the receiver operating characteristic curve as the feature set of the functional subnetwork; Step 5: Find the optimal value of the parameters in the Gaussian radial basis kernel function of the support vector regression machine, and optimize the support vector regression machine using the feature set; Step 6: Use support vector regression to fit the functional subnetwork feature set and the cognitive function status scale score, and predict the cognitive function status based on the mapping relationship between the two.

2. The method for predicting cognitive function status based on functional subnetwork features according to claim 1, Features: Step 1 specifically includes: Step 1.1: All subjects were given the Montreal Cognitive Assessment Scale test to assess their cognitive function status; Step 1.2: A gradient echo single-shot echo-planar imaging sequence was used to scan the cross-sectional brain of the subjects to obtain resting-state functional magnetic resonance images; Step 1.3: Preprocess the functional magnetic resonance images. The preprocessing steps include: time correction, head motion correction, spatial registration, spatial normalization and smoothing filtering operations; Step 1.4: Use the standard brain partition template to divide the brain regions, extract the blood oxygen level dependence intensity of each voxel in each brain region at different time points, and then perform arithmetic averaging on the time series of all voxels in each brain region to obtain the average time series of all voxels in each brain region; Step 1.5: Use the Pearson correlation analysis method to calculate the Pearson correlation coefficient ρ of the average time series between brain regions X,Y : Where N represents the number of regions in the whole brain segmentation; X i represents the value of brain region X at the i-th time point, represents the average value of brain region X at all time points during the entire scanning period; Y i represents the value of brain area Y at the i-th time point, represents the average value of brain region Y at all time points during the entire scanning period; Step 1.6: Obtain the correlation matrix of the average time series based on the correlation coefficient, and take the absolute value of the correlation matrix to obtain the brain functional connection matrix R: Among them, ρ X,Y represents the Pearson correlation coefficient of the average time series between brain region X and brain region Y.

3. The method for predicting cognitive function status based on functional subnetwork characteristics according to claim 1, Features: Step 3 specifically includes: Step 3.1: Use graph theory algorithms to calculate the values ​​of topological attribute parameters of each functional subnetwork, and calculate the area under the receiver operating characteristic curve corresponding to each topological attribute parameter, where the topological attribute parameters include degree, clustering coefficient, characteristic path length, small-world attribute, and node centrality; Step 3.2: Use statistical methods to perform correlation analysis between the area under the receiver operating characteristic curve corresponding to the topological attribute parameters of each functional subnetwork and the cognitive function status scale score, and use the Pearson method to calculate the correlation coefficient r, as shown in the following formula: Among them, G represents the area under the receiver operating characteristic curve corresponding to the topological attribute parameter, W represents the cognitive function status scale score, σ G , σ W Represent the standard deviation of G and W respectively.

4. The method for predicting cognitive function status based on functional subnetwork characteristics according to claim 1, Features: Step 4 specifically includes: Use the T test method to find the corresponding critical value in the distribution table. The formula is as follows: Among them, the t value is a statistic, and n represents the number of elements in the feature set; According to the value of t, the distribution table is consulted to obtain the corresponding p-value. According to the significance level, whether the test result is significant is judged. If the p-value is less than the significance level, the null hypothesis is rejected, and it is considered that the selected topological attribute parameters have a significant effect on the cognitive function status scale score. Finally, the two topological attribute parameters with higher overall correlation are screened out, and the area under the receiver operating characteristic curve corresponding to these two topological attribute parameters is used as the feature collection of the functional sub-network.

5. The method for predicting cognitive function status based on functional subnetwork characteristics according to claim 1, Features: Step 5 specifically includes: Step 5.1: Use the support vector machine model with Gaussian radial basis kernel function, and the decision function is as follows: Among them, ||x i -x|| 2 represents the square of the Euclidean distance between the model input and the center of the kernel function; h is the number of subjects, b is the bias parameter in the decision function; g is the characteristic parameter in the radial basis function; where the model input is the feature set of the functional subnetwork obtained in step 4; Step 5.2: Find the optimal penalty factor C and characteristic parameter g: First, initialize the particle swarm and calculate the fitness of each particle. Then, in each iteration, the particle updates the data by tracking the two extreme values ​​pbest and gbest. When the global optimal position is reached, the global optimal solution is obtained. Step 5.3: Train the support vector regression machine based on the feature set: set the size of the population and the number of iterations within a reasonable range of the algorithm, initialize the position and velocity of the population particles, and each individual is a group of (C, g) vectors; use the network mean square error of the support vector machine as the fitness function to calculate the fitness of the particle swarm in the particle swarm algorithm; bring the parameter value represented by each particle in each iteration process into the trained support vector machine model, and then use the test data to evaluate the network mean square error of the model, that is, the quality of the parameters.

6. The method for predicting cognitive function status based on functional subnetwork characteristics according to claim 1, Features: Step 6 specifically includes: Step 6.1: Assume that the functional subnetwork feature set contains n elements, and the feature vector corresponding to each feature is represented as x i ={x i1 ,x i2 ,…,x ij }, where j is the feature dimension, x ij represents the value of the jth feature of the i-th subject; the target value of the corresponding cognitive function status scale is y i , then the training data set in the support vector machine model is expressed as: T={(x 1 ,y 1 ),(x 2 ,y 2 ),…,(x i ,y i )},x i ∈R h ,y i ∈{1,-1}(i=1,2,…,h) (10) The formula for the minimum distance from the support vector to the decision surface is obtained: Among them, ω is the normal vector of the hyperplane, ||ω|| represents the modulus of ω, and b is the intercept of the hyperplane; Introducing the slack variable ξ i , ξi≥0, then the support vector machine is represented by y i (ω T x i +b)≥1-ξ i , which needs to be minimized The penalty factor C>0; the larger C is, the greater the sample penalty is; the smaller C is, the smaller the sample penalty is; Step 6.2: Introduce the Lagrange multiplier method and obtain: Where ξ refers to the slack variable ξ i A vector of all values, a i ,μ i is the Lagrange multiplier, and α=(α 1 ,α 2 ,…,α h ) T ,μ=(μ 1 ,μ 2 ,…,μ h ) T , α and μ refer to the Lagrange multipliers a i ,μ i All the values ​​of the vector; let L(ω,b,ξ,α,μ) be the value of ω,b,ξ i The partial derivative of is zero, so we can get Substituting the result into the above formula, we can get: The above formula is used to solve the Lagrangian function. The goal is to make the partitioning hyperplane have the maximum interval and transform the convex quadratic programming problem of the maximum interval into the Lagrangian function solution problem; among them, when When 0≤α i ≤C;α j ,y j 、x j is another variable in the dual problem mapped to the high-dimensional space; Solution Get the maximum value of the Lagrange multiplier Then we calculate: By introducing nonlinear mapping and replacing it with the kernel function, we get the decision function of the nonlinear support vector machine model, and its prediction value can be expressed as: Among them, K(x i ,x) is the Gaussian radial basis function, expressed as: Step 6.3: Use the feature collection of the screened functional sub-network to import it into the support vector regression machine after parameter optimization to form a mapping relationship to predict the cognitive function status.