A method for predicting cognitive function state based on multimodal brain network

Abstract

The present invention relates to a method for predicting cognitive function state based on multimodal brain network, the steps of which include: constructing a brain structure connection network and a brain function connection network respectively, and standardizing the networks; allocating weights so that the variance of elements in different networks is minimized, and fusing the two networks to form a multimodal brain network; calculating the topological attribute index of each subject's multimodal brain network as a feature; establishing a mapping relationship from the feature to the neuropsychological scale score, predicting the score through the mapping relationship, the score interval segment where the score is located corresponds to the predicted cognitive function state, and the feature is input into the mapping relationship to obtain the score to achieve cognitive function state prediction. The present invention improves the prediction accuracy.

Description

Technical Field

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

[0002] In recent years, medical imaging technology has been widely used in clinical practice. Among them, brain imaging technology can reflect the anatomical structure information of the brain and explore the secrets of the brain. Manual interpretation of brain imaging data often relies on personal experience, knowledge and emotions, and is inefficient. With the continuous development of computer and artificial intelligence technology, the use of computers to intelligently analyze and process brain imaging data, such as the segmentation, extraction, three-dimensional reconstruction and display of brain structures, and qualitative and quantitative analysis, helps to understand and decipher the working mode of the brain.

[0003] The most important thing to interpret brain imaging data is to build a brain network, and the construction of a brain network includes defining nodes and defining connection edges. When defining nodes, the brain is usually divided into several functional areas according to a medical priori template, and each independent functional area is used as a node in the network. The definition of connection edges is usually defined separately according to different modal data. For example, the functional connection defined by functional magnetic resonance imaging reflects the consistency of brain activity, while the connection defined by diffusion kurtosis images mainly refers to the number of connections between white matter fibers between two brain regions. At present, there are two main types of research on brain network analysis. The first type focuses on using graph theory methods to intuitively understand the real structural connections between different brain regions, and then study the characteristics of brain networks, explore the differences in brain networks between subjects with different cognitive function states, and thus reveal the relationship between the change pattern of brain networks and cognitive function states; the second type goes a step further on the basis of the first type, explores the correlation between cognitive function states and brain network characteristics, and uses machine learning methods to predict the cognitive function state of subjects.

[0004] The above two types of research are mainly based on single-modality neuroimaging technology. For example, functional magnetic resonance imaging, positron emission tomography, diffusion tensor imaging, diffusion kurtosis imaging, etc. Since biomarkers of different modalities can provide complementary information, combining this information can improve classification performance and prediction accuracy. However, simply splicing multimodal data to perform multimodal brain image classification and prediction ignores the relationship between modalities. Therefore, effectively fusing multimodal brain imaging data and predicting cognitive function status based on multimodal brain network characteristics has become a technical problem that needs to be solved urgently in this field. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the traditional technology for predicting cognitive function status, and to provide a new method for predicting cognitive function status based on multimodal brain network, which adopts the following technical solutions:

[0006] A method for predicting cognitive function state based on multimodal brain network, comprising:

[0007] Step 1: Based on the cross-sectional brain imaging of the subjects, two modal data, diffusion kurtosis imaging data and functional magnetic resonance imaging data, are collected. A brain structural connection network matrix is ​​constructed based on the diffusion kurtosis imaging data, and a brain functional connection network matrix is ​​constructed based on the functional magnetic resonance imaging data.

[0008] Step 2: Standardize the brain structure connection network matrix and the brain function connection network matrix, assign weights to the two standardized network matrices so that the variance of the elements in the two network matrices is minimized, and fuse the two network matrices to construct a multimodal brain network matrix;

[0009] Step 3: Calculate the global and local topological attribute indicators of each subject's multimodal brain network matrix as features to form a feature matrix, and reduce the dimension of the feature matrix to extract features, delete redundant features, and retain valid features;

[0010] Step 4: Divide each subject's neuropsychological scale score into different intervals, and the score intervals correspond to the cognitive function status of each subject;

[0011] Step 5: Take the extracted features and neuropsychological scale scores as a data set, and establish a mapping relationship from features to scale scores based on the data set. The interval where the current score is located corresponds to the predicted cognitive function state. The feature is input into the mapping relationship to obtain the score to achieve cognitive function state prediction.

[0012] Furthermore, step 1 specifically includes:

[0013] Step 1.1: Use a spin echo single shot echo planar imaging sequence to scan the cross-sectional imaging of the brain to obtain diffusion kurtosis imaging data; use a gradient echo single shot echo planar imaging sequence to scan the cross-sectional imaging of the brain to obtain functional magnetic resonance imaging data;

[0014] Step 1.2: Preprocess the two modalities of data: diffusion kurtosis imaging data and functional magnetic resonance imaging data;

[0015] Step 1.3: Construct a brain structural connection network, specifically: first divide the preprocessed diffusion kurtosis imaging data into several brain regions, then perform white matter fiber tracking to construct a deterministic structural network, use the deterministic fiber bundle tracking method to calculate the number of fibers between brain regions, and finally obtain a fiber bundle number matrix, i.e., the brain structural connection network matrix;

[0016] Step 1.4: Construct a brain functional connection network, specifically: first, divide the preprocessed functional magnetic resonance imaging data into several brain regions, average the time series of all voxels contained in each brain region, then use the multivariate linear regression analysis method to eliminate some errors caused by head movement and global brain signals, and finally calculate the Pearson correlation coefficient of the average time series between brain regions to obtain a correlation matrix, namely the brain functional connection network matrix.

[0017] Furthermore, in step 1.2, the method for preprocessing the diffusion kurtosis imaging data is as follows: first, convert the original data format of the subjects and construct the parameter maps; second, use head motion and eddy current correction to ensure the consistency between the data and remove the data with large head motion parameters; finally, perform smoothing and noise reduction to eliminate the subtle differences between brain structures of different subjects and reduce spatial noise;

[0018] The method for preprocessing functional magnetic resonance imaging data is as follows: first, convert the subject's original data format; second, delete the data at the initial time point and perform time slice correction and head motion correction on the data; then align, smooth, remove linear drift and low-frequency filter the image; finally, remove the subject's head motion parameters, whole brain signals, white matter and cerebrospinal fluid to obtain the final time series.

[0019] Furthermore, step 2 specifically includes:

[0020] Step 2.1: Perform Min-Max normalization on the brain structure connection network matrix and the brain function connection network matrix respectively. The calculation process is shown in formula (1):

[0021]

[0022] Where: z is the element in the standardized brain structure connection network matrix and the standardized brain function connection network matrix; x is the element corresponding to z in the brain structure connection network matrix and the brain function connection network matrix; min is the minimum value of the elements in the brain structure connection network matrix and the brain function connection network matrix; max is the maximum value of the elements in the brain structure connection network matrix and the brain function connection network matrix;

[0023] Step 2.2: The obtained standardized brain structure connection network matrix and the standardized brain functional connection network matrix are fused to obtain a multimodal brain network matrix. The calculation process is shown in formula (2):

[0024]

[0025] Where: C is the multimodal brain network matrix; A is the standardized brain structure connection network matrix; B is the standardized brain function connection network matrix; ω is the weight, so that cov(ω·A,(1-ω)·B) takes the minimum value, that is: when the covariance of the standardized brain structure connection matrix and the standardized brain function connection matrix is ​​the smallest, it is expected that the elements in the standardized brain structure connection matrix and the standardized brain function connection matrix are as similar as possible.

[0026] Furthermore, step 3 specifically includes:

[0027] Step 3.1: Calculate the global and local topological property parameters of each subject’s multimodal brain network matrix as features, and record these parameters as x1~x n , forming a feature matrix X;

[0028] X=[x1,x2,x3......x n ] (3)

[0029] x j =[x 1j ,x 2j ,x 3j ......x mj ] T (4)

[0030] Where: n represents the number of features; x j represents the vector composed of the values ​​of the jth feature of all subjects, where 1≤j≤n; x ij represents the value of the jth feature of the i-th subject, where 1≤i≤m; m represents the number of subjects;

[0031] Step 3.2: Normalize the feature matrix X to get the matrix Y:

[0032] Y=[y1,y2,y3......y n ] (5)

[0033] y j =[y 1j, y 2j ,y 3j ......y mj ] T (6)

[0034]

[0035] Where: n represents the number of standardized features, which is the same as the number of original features; y j represents the vector consisting of the values ​​of the jth standardized feature of all subjects, where 1≤j≤n; Represents xj The average value of all elements in y ij Represents x ij The corresponding element, i.e., the value of the jth standardized feature of the i-th subject, is 1≤i≤m; m represents the number of subjects;

[0036] Step 3.3: Calculate any two features y in the standardized features p and q The covariance between pq for:

[0037]

[0038] Where: n represents the number of standardized features, y ip Indicates the i-th subject in the standardized feature vector y p The value of , where 1≤p≤n; y iq Indicates the i-th subject in the standardized feature vector y q The value of , where 1≤q≤n; Represents y p The average value of all elements in ; Represents y q The average value of all elements in ; m represents the number of subjects;

[0039] The resulting covariance matrix R is:

[0040]

[0041] Where: n represents the number of standardized features;

[0042] Step 3.4: Calculate the eigenvalues ​​and eigenvectors of the R matrix:

[0043] Rv=λv (10)

[0044] Where: λ is the eigenvalue of the R matrix, and there are n eigenvalues ​​in the n-order R matrix; v is the eigenvector corresponding to λ, and v is an n-dimensional column vector;

[0045] Step 3.5: Sort the n eigenvalues ​​λ in descending order, select the largest first k eigenvalues, where k < n, and extract their corresponding k eigenvectors:

[0046] {(λ1,v1),(λ2,v2),...,(λ k ,v k )} (11)

[0047] K=Y·(v1,v2,...,v k ) (12)

[0048] Where: Y is the standardized feature matrix; K represents the feature matrix after dimensionality reduction.

[0049] Furthermore, step 5 is specifically as follows:

[0050] Step 5.1: Construct a dataset based on the extracted features and neuropsychological scale scores, and divide the dataset into k subsets of equal size. Each subset is used as a test set, and the rest are used as training sets.

[0051] Step 5.2: Use the same algorithm to learn k groups of training sets in turn, and establish a mapping relationship from features to scale scores on the training sets. The features in each test set obtain scale scores through this mapping relationship, and the interval segment where the current score is located corresponds to the predicted cognitive function state.

[0052] Furthermore, step 5.1 is specifically as follows:

[0053] Construct dataset D:

[0054] D={(K1,y1),(K2,y2),......,(K m ,y m )} (13)

[0055] In the formula, K i represents the i-th row element of the matrix K, that is, the value of the i-th subject's characteristic, where 1≤i≤m; y i represents the neuropsychological scale score of the i-th subject, where 1≤i≤m; m is the number of subjects;

[0056] Divide the data set D into k subsets of equal size, each subset is used as a test set, and the rest are used as training sets.

[0057] Furthermore, step 5.2 specifically includes:

[0058] Step 5.21: Based on the principle of support vector regression, establish a mapping relationship from features to scale scores on the training set:

[0059] f(K)=ω T K+b (14)

[0060] Where: K represents the feature matrix after dimensionality reduction; ω represents the weight vector; b represents the bias value;

[0061] Step 5.22: With f(K) as the center, construct an interval band with a width of 2ε, where ε is a constant. If y falls into this interval band, the prediction is successful. The constructed relationship is as follows:

[0062]

[0063]

[0064] Where: represents the search for parameters ω, b such that Take the minimum value; ||ω|| represents the modulus of the weight vector; C represents the regularization constant; l ε is the loss function; K i represents the data of the ith row of the feature matrix K, that is, the value of the eigenvector of the ith subject; y i represents the neuropsychological scale score of the i-th subject; m is the number of subjects;

[0065] Step 5.23: Introduce the slack variable ξ i and Formula (15) can be modified as follows:

[0066]

[0067] Where: It means to find the parameters ω, b, ξ under the constraints i , Make Take the minimum value; st is followed by constraints; m is the number of subjects;

[0068] Step 5.24: Introduce the Lagrange multiplier μ i ≥0, a i ≥0, By Lagrangian multiplication, we can obtain the Lagrangian function of equations (17) and (18):

[0069]

[0070] Step 5.25: Order For ω, b, ξ i , The partial derivative of is zero, so:

[0071]

[0072] Step 5.26: Substituting equation (20) into equation (14) yields a new mapping relationship:

[0073]

[0074] Step 5.27: Introduce the Gaussian kernel function in equation (21), and the final mapping relationship is transformed into:

[0075]

[0076] Where: σ is the width factor of the Gaussian kernel function;

[0077] Step 5.28: Input the characteristics of the subjects in the test set into formula (22) to obtain the corresponding scale score. The interval of the current score corresponds to the predicted cognitive function state.

[0078] Furthermore, the method further comprises the following steps:

[0079] Step 6: Compare the cognitive function status of the subjects predicted by the algorithm with the actual cognitive function status of the subjects to verify the prediction accuracy. If the predicted status is the same as the actual status, the prediction is correct; if the predicted status is different from the actual status, the prediction is wrong. The proportion of the number of subjects with correct predictions in the total number of subjects is taken as the prediction accuracy of the algorithm.

[0080] The present invention has the following beneficial effects:

[0081] 1. The present invention adopts the covariance minimization principle to fuse the brain structure connection network and the brain function connection network, making the elements of the two different networks similar and reducing the impact of individual differences. The fusion method based on the covariance minimization principle effectively fuses the two modal data and concentrates the effective connection information of the structure connection network and the functional connection network in the multimodal brain network, which is better than the simple splicing processing of multimodal data and improves the prediction accuracy.

[0082] Second, the present invention predicts cognitive function status based on multimodal brain networks, making full use of the complementary information provided by structural connection networks and functional connection networks, and the prediction accuracy is higher than the previous cognitive function status prediction method based on single-modal brain networks;

[0083] 3. The present invention performs precise numerical analysis on abstract cognitive function states, converts abstract states into precise numerical values, uses the score intervals of the neuropsychological scale to correspond to cognitive function states, obtains regularities from the numerical values, and uses the regularities to make predictions; this is beneficial for exploring the inherent regularities of cognitive function states and predicting the development trend of cognitive function states. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 This is a flowchart of an implementation of a method for predicting cognitive function status based on a multimodal brain network according to the present invention; DETAILED DESCRIPTION

[0085] 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.

[0086] Please refer to Figure 1 The present invention is a method for predicting cognitive function state based on multimodal brain network, comprising the following steps:

[0087] Step 1: Based on the cross-sectional imaging of the subject's brain, two modal data, diffusion kurtosis imaging data and functional magnetic resonance imaging data, are collected. The brain structural connection network matrix is ​​constructed based on the diffusion kurtosis imaging data, and the brain functional connection network matrix is ​​constructed based on the functional magnetic resonance imaging data: specifically including:

[0088] Step 1.1: Diffusion kurtosis imaging data were obtained by scanning the cross-sectional brain using a single-shot spin echo planar imaging sequence, with b values ​​of b = 0 as the reference and b = 1000 s / mm 2 , 2000s / mm 2 , 30 diffusion gradient encoding directions, repetition time of 6500ms, echo time of 95.8ms, field of view of 240mm×240mm, matrix size of 128×128, slice thickness of 3.6mm, number of excitations of 2 times, number of scanning layers of 30 layers, and scanning time of 14min43s for each subject;

[0089] Functional magnetic resonance imaging data were obtained using a gradient echo single-shot echo-planar imaging sequence to scan cross-sectional brain images, with a repetition time of 2000 ms, an echo time of 40 ms, a field of view of 240 mm × 240 mm, a matrix size of 64 × 64, a slice thickness of 6 mm, and a flip angle of 90°; the scanning time for each subject was 6 min;

[0090] Step 1.2: Preprocess the two modalities of data, diffusion kurtosis imaging data and functional magnetic resonance imaging data; specifically:

[0091] The collected diffusion kurtosis imaging data were preprocessed using the PANDA toolbox. The preprocessing steps were as follows: first, the original data of the subjects were converted to construct the parameter maps; second, head motion and eddy current correction were used to ensure the consistency between the data and remove the data with large head motion parameters; finally, smoothing and noise reduction were performed to eliminate the subtle differences in brain structures between different subjects and reduce spatial noise;

[0092] The acquired functional magnetic resonance imaging data were preprocessed using the DPARSF toolbox. The preprocessing steps were as follows: first, the original image format of the subject was converted; second, the data of the first 10 time points were deleted and the time points and image positions were corrected; then, the images were registered, smoothed, linear drift removed, and low-frequency filtered, with a filtering frequency range of 0.01 to 0.08 Hz; finally, the subject's head motion parameters, whole brain signals, white matter, and cerebrospinal fluid were removed to obtain the final time series;

[0093] Step 1.3: Construct the brain structural connection network. The process is as follows: first, the preprocessed diffusion kurtosis image is divided into 90 brain regions using the Automated Anatomical Labeling (AAL) template; second, the PANDA toolbox is used for white matter fiber tracking to construct a deterministic structural network; finally, the deterministic fiber bundle tracking method is used to calculate the number of fibers between brain regions to obtain a 90×90 fiber bundle number matrix, i.e., the brain structural connection network matrix;

[0094] Step 1.4: Construct a brain functional connection network. The process is as follows: first, use the AAL template to divide the preprocessed functional magnetic resonance images into 90 brain regions; second, average the time series of all voxels contained in each brain region, and use the multivariate linear regression analysis method to eliminate some errors caused by head movement and global brain signals; finally, calculate the Pearson correlation coefficient of the average time series between brain regions to obtain a 90×90 correlation matrix, namely the brain functional connection network matrix.

[0095] Step 2: Standardize the brain structure connection network matrix and the brain function connection network matrix, assign weights to the two standardized network matrices so that the variance of the elements in the two network matrices is minimized, and fuse the two network matrices to form a multimodal brain network matrix; specifically, it includes:

[0096] Step 2.1: Perform Min-Max normalization on the brain structure connection network matrix and the brain function connection network matrix respectively. The calculation process is shown in formula (23):

[0097]

[0098] Where: z is the element in the standardized brain structure connection network matrix and the standardized brain function connection network matrix; x is the element corresponding to z in the brain structure connection network matrix and the brain function connection network matrix; min is the minimum value of the elements in the brain structure connection network matrix and the brain function connection network matrix; max is the maximum value of the elements in the brain structure connection network matrix and the brain function connection network matrix;

[0099] Step 2.2: The obtained standardized brain structure connection network matrix and the standardized brain functional connection network matrix are fused to obtain a multimodal brain network matrix. The calculation process is shown in formula (24):

[0100]

[0101] Where: C is the multimodal brain network matrix; A is the standardized brain structure connection network matrix; B is the standardized brain function connection network matrix; ω is the weight, so that cov(ω·A,(1-ω)·B) takes the minimum value, that is: when the covariance of the standardized brain structure connection network matrix and the standardized brain function connection network matrix is ​​the smallest, it is expected that the elements in the standardized brain structure connection network matrix and the standardized brain function connection network matrix are as similar as possible.

[0102] Step 3: Calculate the global and local topological attribute indicators of each subject's multimodal brain network matrix as features to form a feature matrix, reduce the dimension of the feature matrix, delete redundant features, and retain valid features;

[0103] Step 3.1: Calculate the node degree, betweenness centrality, standardized clustering coefficient, standardized path length, and small-world property of each subject's multimodal brain network matrix as features, and record these parameters as x1 to x5 in turn to form a feature matrix X;

[0104] X=[x1,x2,x3,x4,x5] (25)

[0105] x j =[x 1j ,x 2j ,x 3j ......x 100j ] T (26)

[0106] Where: x j represents the vector composed of the values ​​of the jth feature of 100 subjects, where 1≤j≤5; x ij represents the value of the jth feature of the i-th subject, where 1≤i≤100;

[0107] Step 3.2: Normalize the feature matrix X to get the matrix Y:

[0108] Y=[y1,y2,y3,y4,y5] (27)

[0109] y j =[y 1j, y 2j ,y 3j ......y 100j ] T (28)

[0110]

[0111] Where: y j represents the vector of the values ​​of the jth standardized feature of 100 subjects, where 1≤j≤5; Represents x j The average value of all elements in y ij Represents x ij The corresponding element is the value of the jth standardized feature of the i-th subject;

[0112] Step 3.3: Calculate any two features y in the standardized features p and q The covariance between pq for:

[0113]

[0114] Where: y ip Indicates the i-th subject in the standardized feature vector y p The value of y, where 1≤p≤5; iq Indicates the i-th subject in the standardized feature vector y q The value of , where 1≤q≤5; Represents y p The average value of all elements in ; Represents y q The average value of all elements in ;

[0115] The resulting covariance matrix R is:

[0116]

[0117] Step 3.4: Calculate the eigenvalues ​​and eigenvectors of the R matrix:

[0118] Rv=λv (32)

[0119] Where: λ is the eigenvalue of the R matrix, and there are 5 eigenvalues ​​in the 5-order R matrix; v is the eigenvector corresponding to λ, and v is a 5-dimensional column vector;

[0120] Step 3.5: Sort the five eigenvalues ​​λ in descending order, select the first two largest eigenvalues, and extract their corresponding two eigenvectors:

[0121] {(λ1,v1),(λ2,v2)} (33)

[0122] K=Y·(v1,v2) (34)

[0123] Where: Y is the standardized feature matrix; K represents the feature matrix after dimensionality reduction, which contains 2 retained features.

[0124] Step 4: Divide each subject's neuropsychological scale score into three different intervals: a score greater than or equal to 26 corresponds to normal cognitive function; a score between 20 and 25 corresponds to mild cognitive impairment; and a score less than 20 corresponds to severe cognitive impairment.

[0125] Step 5: Take the extracted features and neuropsychological scale scores as a data set, and establish a mapping relationship from features to scale scores based on the data set. The interval segment where the current score is located corresponds to the predicted cognitive function state. Input the features into the mapping relationship to obtain the score to achieve cognitive function state prediction; specifically including:

[0126] Step 5.1: Construct a data set based on the extracted features and neuropsychological scale scores, and divide the data set into k subsets of equal size. Each subset is used as a test set, and the rest are used as training sets; specifically:

[0127] The retained features and neuropsychological scale scores were used as the data set, and the data set was divided into k = 10 subsets of equal size, each of which was used as a test set, and the rest were used as training sets;

[0128] Construct dataset D:

[0129] D={(K1,y1),(K2,y2),......,(K 100 ,y 100 )} (35)

[0130] In the formula, K i represents the i-th row element of matrix K, i.e., the value of the i-th subject’s characteristic (1≤i≤100); i represents the neuropsychological scale score of the i-th subject (1≤i≤100);

[0131] Divide the dataset D into 10 subsets of equal size, each subset is used as a test set, and the rest are used as training sets.

[0132] Step 5.2: Use the same algorithm to learn 10 training sets in turn, establish a mapping relationship from features to scale scores on the training sets, and obtain scale scores from the features in each test set through the mapping relationship. The interval segment where the current score is located corresponds to the predicted cognitive function state; Step 5.2 specifically includes:

[0133] Step 5.21: Based on the principle of support vector regression, establish a mapping relationship from features to scale scores on the training set:

[0134] f(K)=ω T K+b (36)

[0135] Where: K represents the feature matrix after dimensionality reduction; ω represents the weight vector; b represents the bias value;

[0136] Step 5.22: With f(K) as the center, construct an interval band with a width of 2ε, where ε is a constant. If y falls into this interval band, the prediction is successful. The constructed relationship is as follows:

[0137]

[0138] Where: represents the search for parameters ω, b such that Take the minimum value; ||ω|| represents the modulus of the weight vector; C represents the regularization constant; l ε is the loss function; K i represents the data of the ith row of the feature matrix K, that is, the value of the eigenvector of the ith subject; y i represents the neuropsychological scale score of the ith subject;

[0139] Step 5.23: Introduce the slack variable ξ i and Formula (37) can be modified as follows:

[0140]

[0141] Where: It means to find the parameters ω, b, ξ under the constraints i , Make Take the minimum value; st is followed by the constraint condition;

[0142] Step 5.24: Introduce the Lagrange multiplier μ i ≥0, a i ≥0, By Lagrangian multiplication, we can obtain the Lagrangian functions of equations (39) and (40):

[0143]

[0144] Step 5.25: Order For ω, b, ξ i , The partial derivative of is zero, so:

[0145]

[0146] Step 5.26: Substituting equation (42) into equation (36) yields a new mapping relationship:

[0147]

[0148] Step 5.27: Introduce the Gaussian kernel function in equation (43), and the final mapping relationship is transformed into:

[0149]

[0150] Where: σ is the width factor of the Gaussian kernel function;

[0151] Step 5.28: Input the characteristics of the subjects in the test set into formula (44) to obtain the corresponding scale score, and the interval of the score corresponds to the predicted cognitive function state.

[0152] Step 6: Compare the cognitive function status of the subjects predicted by the algorithm with the actual cognitive function status of the subjects to evaluate the prediction accuracy. If the predicted status is the same as the actual status, the prediction is correct; if the predicted status is different from the actual status, the prediction is wrong. The proportion of the number of subjects with correct predictions in the total number of subjects is taken as the prediction accuracy of the algorithm.

[0153] The parts not involved in the present invention are the same as the prior art or are implemented by using the prior art.

[0154] 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 protection scope of the present invention.

Claims

1. A method for predicting cognitive function status based on multimodal brain networks, characterized in that: include Step 1: Based on the cross-sectional brain imaging of the subjects, two modal data, diffusion kurtosis imaging data and functional magnetic resonance imaging data, are collected. A brain structural connection network matrix is ​​constructed based on the diffusion kurtosis imaging data, and a brain functional connection network matrix is ​​constructed based on the functional magnetic resonance imaging data. Step 2: Standardize the brain structure connection network matrix and the brain function connection network matrix, assign weights to the two standardized network matrices so that the variance of the elements in the two network matrices is minimized, and fuse the two network matrices to construct a multimodal brain network matrix; Step 3: Calculate the global and local topological attribute indicators of each subject's multimodal brain network matrix as features to form a feature matrix, and reduce the dimension of the feature matrix to extract features; specifically include: Step 3.1: Calculate the global and local topological property parameters of each subject’s multimodal brain network matrix as features, and record these parameters as x1~x n , forming a feature matrix X; X=[x1,x2,x3......x n ] (3) x j =[x 1j ,x 2j ,x 3j ……x mj ] T (4) Where: n represents the number of features; x j represents the vector composed of the values ​​of the jth feature of all subjects, where 1≤j≤n; x ij represents the value of the jth feature of the i-th subject, where 1≤i≤m; m represents the number of subjects; Step 3.2: Normalize the feature matrix X to get the matrix Y: <h2 style=";text-align:left;direction:ltr">Y = [y1,y2,y3......y<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> ] (5) and j =[and 1j, and 2j ,and 3j ......and mj ] T (6) Where: y j Represents the vector consisting of the values ​​of the jth standardized feature of all subjects; Represents x j The average value of all elements in y ij Represents x ij The corresponding element is the value of the jth standardized feature of the i-th subject; Step 3.3: Calculate any two features y in the standardized features p and q The covariance between pq for: Where: y ip Indicates the i-th subject in the standardized feature vector y p The value of , where 1≤p≤n; y iq Indicates the i-th subject in the standardized feature vector y q The value of , where 1≤q≤n; Represents y p The average value of all elements in ; Represents y q The average value of all elements in ; The resulting covariance matrix R is: Step 3.4: Calculate the eigenvalues ​​and eigenvectors of the R matrix: Rv=λv (10) Where: λ is the eigenvalue of the R matrix, and there are n eigenvalues ​​in the n-order R matrix; v is the eigenvector corresponding to λ, and v is an n-dimensional column vector; Step 3.5: Sort the n eigenvalues ​​λ in descending order, select the largest first k eigenvalues, where k < n, and extract their corresponding k eigenvectors: {(λ1,v1),(λ2,v2),...,(λ k ,v k )} (11) K=Y·(v1,v2,...,v k ) (12) Where: Y is the standardized feature matrix; K represents the feature matrix after dimensionality reduction; Step 4: Divide each subject's neuropsychological scale score into different intervals, and the score intervals correspond to the cognitive function status of each subject; Step 5: The extracted features and neuropsychological scale scores are used as a data set. A mapping relationship from features to scale scores is established based on the data set. The interval segment where the current score is located corresponds to the predicted cognitive function state. The features are input into the mapping relationship to obtain the score to realize the cognitive function state prediction; Step 5.1: Construct a data set based on the extracted features and neuropsychological scale scores, and divide the data set into k subsets of equal size. Each subset is used as a test set, and the rest are used as training sets. Specifically: Construct dataset D: D={(K1,y1),(K2,y2),......,(K m ,y m )} (13) In the formula, K i represents the i-th row element of the matrix K, that is, the value of the i-th subject’s characteristic; y i represents the neuropsychological scale score of the i-th subject; Divide the data set D into k subsets of equal size, each subset is used as a test set, and the rest are used as training sets; Step 5.2: Study the k training sets in turn, establish a mapping relationship from features to scale scores on the training sets, and obtain scale scores from the features in each test set through this mapping relationship. The interval segment where the current score is located corresponds to the predicted cognitive function state, including: Step 5.21: Based on the principle of support vector regression, establish a mapping relationship from features to scale scores on the training set: f(K)=ω T K+b (14) Where: ω represents the weight vector; b represents the bias value; Step 5.22: With f(K) as the center, construct an interval band with a width of 2ε, where ε is a constant. If y falls into this interval band, the prediction is successful. The constructed relationship is as follows: Where: represents the search for parameters ω, b such that Take the minimum value; ||ω|| represents the modulus of the weight vector; C represents the regularization constant; l ε is the loss function; Step 5.23: Introduce the slack variable ξ i and Modify formula (15) as follows: Where: It means to find the parameters ω, b, ξ under the constraints i , Make Take the minimum value; st is followed by the constraint condition; Step 5.24: Introduce the Lagrange multiplier μ i ≥0, a i ≥0, The Lagrangian functions of equations (17) and (18) are obtained by Lagrangian multiplication: Step 5.25: Order For ω, b, ξ i , The partial derivative of is zero, and we get: Step 5.26: Substitute equation (20) into equation (14) to obtain a new mapping relationship: Step 5.27: Introduce the Gaussian kernel function in equation (21), and the final mapping relationship is transformed into: Where: σ is the width factor of the Gaussian kernel function; Step 5.28: Input the characteristics of the subjects in the test set into formula (22) to obtain the corresponding scale score. The interval of the current score corresponds to the predicted cognitive function state.

2. The method for predicting cognitive function status based on multimodal brain network according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Use a spin echo single shot echo planar imaging sequence to scan the cross-sectional imaging of the brain to obtain diffusion kurtosis imaging data; use a gradient echo single shot echo planar imaging sequence to scan the cross-sectional imaging of the brain to obtain functional magnetic resonance imaging data; Step 1.2: Preprocess the two modalities of data: diffusion kurtosis imaging data and functional magnetic resonance imaging data; Step 1.3: Construct a brain structural connection network, specifically: first divide the preprocessed diffusion kurtosis imaging data into several brain regions, then perform white matter fiber tracking to construct a deterministic structural network, use the deterministic fiber bundle tracking method to calculate the number of fibers between brain regions, and finally obtain a fiber bundle number matrix, i.e., the brain structural connection network matrix; Step 1.4: Construct a brain functional connection network, specifically: first, divide the preprocessed functional magnetic resonance imaging data into several brain regions, average the time series of all voxels contained in each brain region, then use the multivariate linear regression analysis method to eliminate some errors caused by head movement and global brain signals, and finally calculate the Pearson correlation coefficient of the average time series between brain regions to obtain a correlation matrix, namely the brain functional connection network matrix.

3. The method for predicting cognitive function status based on multimodal brain network according to claim 2, characterized in that: In step 1.2, the method for preprocessing the diffusion kurtosis imaging data is as follows: first, convert the original data format of the subject and construct each parameter map; second, perform head motion and eddy current correction; and finally, perform smoothing and noise reduction processing; The method for preprocessing functional magnetic resonance imaging data is as follows: first, convert the subject's original data format; second, delete the data at the initial time point and perform time slice correction and head motion correction on the data; then align, smooth, remove linear drift and low-frequency filter the image; finally, remove the subject's head motion parameters, whole brain signals, white matter and cerebrospinal fluid to obtain the final time series.

4. The method for predicting cognitive function status based on multimodal brain network according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1: Perform Min-Max normalization on the brain structure connection network matrix and the brain function connection network matrix respectively. The calculation process is shown in formula (1): Where: z is the element in the standardized brain structure connection network matrix and the standardized brain function connection network matrix; x is the element corresponding to z in the brain structure connection network matrix and the brain function connection network matrix; min is the minimum value of the elements in the brain structure connection network matrix and the brain function connection network matrix; max is the maximum value of the elements in the brain structure connection network matrix and the brain function connection network matrix; Step 2.2: The obtained standardized brain structure connection network matrix and the standardized brain functional connection network matrix are fused to obtain a multimodal brain network matrix. The calculation process is shown in formula (2): Where: C is the multimodal brain network matrix; A is the standardized brain structure connection network matrix; B is the standardized brain function connection network matrix; ω is the weight, so that cov(ω·A,(1-ω)·B) takes the minimum value.

5. The method for predicting cognitive function status based on multimodal brain network according to claim 1, characterized in that: The method further comprises the following steps: Step 6: Compare the predicted cognitive function status of the subjects with their actual cognitive function status to evaluate the prediction accuracy. If the predicted status is the same as the actual status, the prediction is correct; if the predicted status is different from the actual status, the prediction is wrong. The proportion of the number of subjects with correct prediction in the total number of subjects is taken as the prediction accuracy.

Images