CN111627553A - Method for constructing individualized prediction model of first-onset schizophrenia - Google Patents

Abstract

The invention belongs to the fields of psychiatry, neuroimaging and artificial intelligence, and discloses a construction method for an individualized prediction model of first-episode schizophrenia, which solves the problem of low accuracy of auxiliary diagnosis of the existing SCH brain structure network model. The method includes the following steps: A. Obtaining a diffusion tensor image of a first-episode schizophrenic patient; B. Preprocessing the obtained diffusion tensor image; C. Constructing a sparse brain structure network based on the preprocessed image; D. The similarity network fusion method is used to construct the multi-threshold fusion brain structure network after each subject's sparseness; E. Extract the topological attribute features of the multi-threshold fusion brain structure network, and then perform feature screening; F. Based on the filtered features, use a classifier to perform Perform classification training to obtain an individualized prediction model for first-episode schizophrenia; G. Perform performance verification and evaluation on the individualized prediction model for first-episode schizophrenia obtained through training.

Description

Construction method of individualized prediction model of first-episode schizophrenia

technical field

The invention belongs to the fields of psychiatry, neuroimaging and artificial intelligence, and particularly relates to a method for constructing an individualized prediction model for first-episode schizophrenia.

Background technique

Schizophrenia (SCH) is a highly disabling and fatal mental disorder. The World Health Organization ranks it as one of the top ten diseases in the global disease burden list. However, its brain mechanism has not been fully clarified. Diagnosis lacks objective criteria and cure rates are low. It has become an urgent clinical problem to seek an objective, effective, convenient and feasible biological marker for early individualized classification, diagnosis and treatment of SCH. Brain structural network changes are an important biological basis for neuroanatomical abnormalities in SCH. As a data-driven prediction and analysis tool, machine learning methods can make full use of the inherent structural information of biomarker data to build an individualized brain structural network model for SCH.

The current research status of the SCH brain structure network model is: 1) Directly extract the structural connection value in the original network as a feature, and incorporate pseudo-connections with lower weights; 2) The method of sparse network based on a single fixed threshold has different levels of noise influence , and the selection of a single threshold is subjective; 3) The original structural connection value is directly used as the low-level structural information feature, ignoring the important network properties of the brain's complex topology.

Therefore, it is very difficult to find sensitive biomarkers of schizophrenia based on the characteristics of the brain structure network in the existing SCH brain structure network model, and the accuracy of the auxiliary diagnosis of the model is low.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is: to propose a method for constructing an individualized prediction model of first-episode schizophrenia, so as to solve the problem of low accuracy of auxiliary diagnosis of the existing SCH brain structure network model.

The technical scheme adopted by the present invention to solve the above-mentioned technical problems is:

The construction method of the individualized prediction model of first-episode schizophrenia includes the following steps:

A. Obtaining diffusion tensor images of patients with first-episode schizophrenia;

B, preprocessing the acquired diffusion tensor image;

C. Build a sparse brain structure network based on the preprocessed images;

D. Using the similarity network fusion method to construct a multi-threshold fusion brain structure network after each subject's sparseness;

E. Extract multi-threshold fusion brain structure network topology attribute features, and then perform feature screening;

F. Based on the characteristics after screening, the classifier is used for classification training to obtain the individualized prediction model of first-episode schizophrenia;

G. Perform performance verification evaluation on the individualized prediction model of first-episode schizophrenia obtained by training.

As a further optimization, in step A, the diffusion tensor image of the first-episode schizophrenia patient is obtained by scanning with a single-shot echo-planar imaging (EPI) technique using an MRI scanner.

As a further optimization, in step B, the preprocessing of the acquired diffusion tensor image specifically includes:

B1. Use MRI convert (magnetic resonance imaging conversion) to convert the diffusion tensor image in DICOM data format to an image in NIFTI format;

B2. Perform eddy current correction and head motion correction on the diffusion tensor image after data format conversion;

B3. Use the Brain Extraction Tool of FSL to remove the skull and remove non-brain tissue images.

As a further optimization, in step C, the construction of a sparse brain structure network based on the preprocessed images specifically includes:

C1. The preprocessed brain image of each subject is registered to the b0 image by the linear registration method of rotation and translation in the diffusion tensor space; then the registered b0 image is registered to the T1 image in the standard MNI space ; By inverting the transformation matrix, using the obtained inverse matrix to transform the AAL template from the MNI space to the diffusion tensor space, and obtaining 90 brain area network nodes divided by each subject based on the AAL template;

C2. Using the probabilistic tractography method, based on the BEDPOSTX tool for sampling Bayesian estimation of the diffusion parameters, and using the Markov Chain Monte Carlo (MCMC) sampling method to establish a distribution for the diffusion parameters of each voxel , preset each voxel of the brain as a fiber crossing model, and automatically determine how many fiber bundles there are crossing;

C3. Probability tracking fiber bundle reconstruction based on PROBTRACKX tool, by repeatedly sampling the distribution of the main diffusion direction of each voxel, generating streamlines from the extracted local samples each time, and establishing the streamline position through multiple sampling The statistical diagram of the test distribution can be used to obtain the distribution of the probability of structural connection between the two brain regions; the weight of each edge is defined as the probability of fiber bundle connection between the two node regions, then each subject obtains a symmetrical 90×90 The weighted network matrix of the fiber bundle connection probability;

C4. Set the probability threshold of fiber bundle connection. There are structural connections between the two brain regions that exceed the threshold. Test the effect of different sparsity thresholds on the fusion effect. Use a relatively narrow threshold range (5% to 40%, step size is 1) %) to build a sparse structure network.

As a further optimization, in step D, the similar network fusion method is used to construct a sparse multi-threshold fusion brain structure network for each subject, which specifically includes:

D1. Define the connection matrix after sparse as a full kernel matrix W _i ^j , for the full kernel matrix W _i ^j under the jth threshold of the i-th subject, further construct the corresponding sparse kernel (sparse kernel) matrix:

Let δ _u be the k-nearest neighbors of node u in the full-kernel matrix Wi _j (including node u itself), then the sparse kernel ^matrix S _i ^j is defined as:

D2. Iteratively update the full-kernel matrix based on the sparse kernel matrix corresponding to the full-kernel matrix:

where (W _i ^c ) ^(m) represents the full kernel matrix of the i-th subject at the c-th threshold at the m-th iteration, and (W _i ^j ) ^(m+1) represents the full-kernel matrix at the m+1-th iteration Kernel matrix, N is the total number of sparse thresholds;

D3. Determine whether the iterative convergence condition is met, if so, execute step D4, otherwise continue to iterate;

Wherein, the convergence condition is: ||(W _i ^j ) ^(m+1) -(W _i ^j ) ^(m) ||≤0.01;

D4. Average the full-kernel matrix W _i ^j corresponding to the updated N thresholds, so as to construct an average full-kernel matrix for each subject:

D5. _Normalize the elements in Wi to the interval [0, 1], thereby generating the final fused network for each subject.

As a further optimization, in step E, the extraction of multi-threshold fusion brain structure network topology attribute features specifically includes:

Based on the graph theory analysis method, the AUC values of 8 global topological attributes and 3 node topological attributes of the fused network under all thresholds were calculated as the initial features of subsequent classification;

The 8 global topological properties include network strength Sp , global efficiency E _glob , local efficiency E _loc , shortest path length L _p , clustering coefficient C _p , _normalized shortest path length λ , normalized clustering coefficient γ , small-world property σ :

The calculation formula of the network strength _Sp is:

where S(i) is the sum of the weights of the edges connected to the ith node, and N is the number of brain regions in the whole-brain network;

The calculation formula of the shortest path length L _p is:

Among them, L _ij represents the shortest path between node i and node j, and L _p is the shortest path length of the entire network G;

The calculation formula of the global efficiency E _glob includes:

Among them, E _{glob_i} (G) is the global efficiency of node i, and the global efficiency of network G is the average value of the global efficiency of all nodes in the network;

The calculation formula of the local efficiency E _loc includes:

where L _jk is the shortest path length between region j and region k, G _i is the sub-network formed by nodes connected to region i, N _Gi is the number of brain regions in sub-network _Gi ; E _{loc_i} (G) is The local efficiency of node i, the local efficiency of network G E _loc (G) is the average of the local efficiencies of all nodes in the network;

The calculation formula of the agglomeration coefficient C _p includes:

Among them, C(i) is the clustering coefficient of node i, and the clustering coefficient of network G is the average value of the clustering coefficients of all nodes;

The normalized aggregation coefficient γ and the normalized shortest path length λ are calculated as follows:

Among them, C _p ^rand and L _p ^rand are the average of the clustering coefficient and the shortest path length of 100 random networks;

The calculation formula of the small-world attribute σ is as follows:

σ=γ/λ;

The three types of node topology attributes include node degree D _nodal (i), node efficiency E _nodal (i), node betweenness centrality B _nodal (i), which are respectively defined as follows:

where _est represents the number of all shortest paths from node s to node t in network G, and e _sit is the number of these shortest paths through node i.

As a further optimization, in step E, the feature screening specifically includes:

The Recursive Feature Elimination (RFE) algorithm based on support vector machine performs feature selection by continuously training the classifier and removing the feature dimension with small feature weight, including:

① Initialize the selected feature set to contain all selected features,

② Take the feature set as input, train the classifier, and get the classification effect and the weight of each feature;

③ Remove the feature with the smallest weight to form a new feature set;

选取分类效果最好的情况。④Repeat ②, ③ until

Choose the case with the best classification effect.

As a further optimization, in step F, the classifier adopts SVM classifier based on radial basis function (RBF) kernel, logistic regression classifier (Logistic Regression Classifier) or integrated learning of multiple classifiers; During training, a variety of pattern recognition classification methods are applied to find the optimal classifier model and screen the key multi-threshold fusion brain structure network features.

As a further optimization, in step G, cross-validation is used to evaluate the performance of various prediction models of schizophrenia training set and test set, including accuracy, sensitivity, specificity, ROC curve and AUC value, etc. Classification algorithms, applying permutation tests, optimizing brain imaging feature selection and individual subtype prediction models, and identifying subtype-related core brain imaging group features to improve individualized prediction accuracy.

The beneficial effects of the present invention are:

(1) Constructing a multi-threshold fusion network, by integrating the complementary information provided by the original network under different topological views, the topological attribute features of the fusion network that do not depend on a single threshold are generated as the initial features of classification. Classification training, which can not only improve the classification accuracy, but also consider the components of structural indicators, which is interpretable;

(2) The artificial intelligence model based on the fusion of multi-threshold brain structure network biological data can automatically collect data to update the model and establish objective biological markers for early diagnosis of schizophrenia patients. The model is accurate and qualitative for the classification of first-episode schizophrenia patients. Shows extremely high reliability and high stability at the same time.

Description of drawings

FIG. 1 is a flowchart of the construction method of the individualized prediction model of first-episode schizophrenia in the present invention.

Detailed ways

The invention aims to propose a method for constructing an individualized prediction model of first-episode schizophrenia, so as to solve the problem of low accuracy of auxiliary diagnosis of the existing SCH brain structure network model. The core idea is to obtain the single-shot echo plane imaging of the diffusion tensor of patients with first-episode schizophrenia; to preprocess the diffusion tensor images; to construct a sparse brain structure network based on the preprocessed images; Construct a sparse multi-threshold structure network for each subject; extract the topological attribute features based on the processed fusion multi-threshold brain structure network, perform classification training after feature screening, and obtain an individualized prediction model for first-episode schizophrenia. Performance validation evaluation of an individualized prediction model for first-episode schizophrenia.

Networks with different sparsity thresholds can be regarded as different types of feature expressions for the same subject's brain network. The present invention fuses networks with multiple sparsity thresholds to obtain richer topology information, which is beneficial for subsequent The classification work also avoids the influence of different levels of noise in the method of using a single fixed threshold to sparse the network, and the subjectivity of the selection of a single threshold; and the network topology attributes obtained by the present invention based on graph theory analysis can reflect the high level of the network. Hierarchical attributes, using these attributes as classification features can obtain better classification results than using the original network connection values, which can only reflect low-level information as classification features.

In terms of specific implementation, the construction method flow of the first-episode schizophrenia individualized prediction model in the present invention is shown in Figure 1, and it includes the following implementation steps:

1. Obtain diffusion tensor images of first-episode schizophrenia patients;

In this step, the data collected by Philip 3.0T and GE 3.0T MRI scanners are used as the training module test set and the test module data set. As a specific implementation method, the scanning parameters are as follows: in the 32 axial plane directions, TR= 10295ms, TE=91ms, FOV=128mm×128mm ² , flip angle=90°, slice thickness=4mm, matrix=256×256, single voxel size is 2×2×2mm ³ , b=1000m/s. Collect 3DT1 structural image data to optimize DTI data registration. The scanning parameters are as follows: TR=8.4ms, TE=3.8ms, FOV=256×256mm ² , flip angle=90°, slice thickness=lmm, continuous scan without interval, matrix= 256×256, the size of a single voxel is 1×1×1 mm ³ , and a total of 188 slices of images are collected in the whole brain.

2. Preprocessing the acquired diffusion tensor image;

In this step, as a specific means of implementation, the preprocessing process is as follows:

①Use MRI convert to convert the diffusion tensor image in DICOM data format to NIFTI format image;

② Perform eddy current correction and head motion correction on the diffusion tensor image after data format conversion;

③ Apply FSL's Brain Extraction Tool to remove the skull and remove the non-brain tissue images.

3. Build a sparse brain structure network based on the preprocessed images;

In this step, as a specific implementation method, the process of constructing a sparse brain structure network is as follows:

①Register the preprocessed brain image of each subject to the b0 image in the diffusion tensor space using the linear registration method of rotation and translation; then register the registered b0 image to the T1 image in the standard MNI space; By inverting the transformation matrix, using the obtained inverse matrix to transform the AAL template from the MNI space to the diffusion tensor space, and obtaining 90 brain area network nodes divided by the AAL template for each subject;

②Using the probabilistic tractography method, based on the BEDPOSTX tool for sampling Bayesian estimation of the diffusion parameters, and using the Markov chain Monte Carlo sampling method to establish a distribution for the diffusion parameters of each voxel, and preset each voxel in the brain It is a fiber crossing model, and automatically determines how many types of fiber bundles cross through;

③Probability tracing fiber bundle reconstruction based on the PROBTRACKX tool, by repeatedly sampling the distribution of the main diffusion direction of each voxel, generating streamlines from the extracted local samples each time, and establishing the posterior position of the streamlines through multiple sampling The statistical diagram of the distribution can obtain the distribution of the probability of structural connection between the two brain regions; the weight of each edge is defined as the probability of fiber bundle connection between the two node regions, then each subject obtains a symmetrical 90×90 Weighted network matrix of fiber bundle connection probabilities;

(4) Set the fiber bundle connection probability threshold, and there are structural connections between the two brain regions that exceed the threshold. Test the effect of different sparsity thresholds on the fusion effect, using a relatively narrower threshold range (5% to 40%, with a step size of 1%). ) to build a sparse structure network.

4. Using the similarity network fusion method to construct a multi-threshold fusion brain structure network after each subject's sparse;

In this step, as a specific implementation method, the process of constructing a multi-threshold fusion brain structure network is as follows:

① The sparse post-structure connection matrix is defined as the full-kernel matrix W _i ^j , for the full-kernel matrix Wi j under the j-th threshold of the _i - ^th subject, the corresponding sparse kernel matrix is further constructed. The sparse-kernel matrix is used for coding Strong connections that remain after the network is sparse:

Let δ _u be the k-nearest neighbors of node u in the full-kernel matrix Wi _j (including node u itself), then the sparse kernel ^matrix S _i ^j is defined as:

2. Iteratively update the full kernel matrix based on the sparse kernel matrix corresponding to the full kernel matrix:

where (W _i ^c ) ^(m) represents the full kernel matrix of the i-th subject at the c-th threshold at the m-th iteration, and (W _i ^j ) ^(m+1) represents the full-kernel matrix at the m+1-th iteration Kernel matrix, N is the total number of sparse thresholds;

By interacting with all other threshold networks except its own, the all-kernel matrix W _i ^j can integrate the complementary information provided by the original network under other topological views. Meanwhile, the sparse kernel matrix S _i ^j guides the iterative process through the strongest connection in the corresponding full kernel matrix W _i ^j , so the noise can be effectively suppressed. Looking at the entire iterative process from the perspective of matrix multiplication, the above formula means that the connection value of any two nodes in the full-kernel matrix W _i ^j also depends on the k-nearest neighbors of the corresponding nodes in other threshold networks. In particular, if the respective k-nearest neighbors of two nodes are the strongest connections in other thresholded networks, the connections between them will strengthen after iterative updates (although they may themselves be weak connections), and vice versa.

3. Determine whether the iterative convergence conditions are met, if so, perform step 4, otherwise continue to iterate;

Wherein, the convergence condition is: ||(W _i ^j ) ^(m+1) -(W _i ^j ) ^(m) ||≤0.01;

④. Average the full-kernel matrix W _i ^j corresponding to the updated N thresholds to construct an average full-kernel matrix for each subject:

⑤. _Normalize the elements in Wi to the interval [0, 1] to generate the final fused network for each subject.

5. Extract multi-threshold fusion brain structure network topology attribute features, and then perform feature screening;

In this step, as a specific implementation method, the AUC values of 8 global topology attributes and 3 node topology attributes of the fused network under all thresholds are calculated based on the graph theory analysis method as the initial features of subsequent classification. Among them, the 8 global topological properties include network strength (node degree) S _p , global efficiency E _glob , local efficiency E _loc , shortest path length L _p , clustering coefficient C _p , normalized shortest path length λ, normalized agglomeration coefficient γ, Small world property σ. The specific definitions are as follows:

The network strength (node degree) _Sp reflects important network evolution characteristics. The degree of a node is defined as the sum of the weights of the edges directly connected to the node. The greater the degree of a node, the more connections the node has, and the more important its position in the network is. The definition formula is:

where S(i) is the sum of the weights of edges connected to the ith node, and N is the number of brain regions in the whole-brain network. The average of the degrees of all nodes in a network is the strength of that network.

The shortest path length L _p : the average value of the shortest paths of all nodes in the network is the shortest path of the network, reflecting the operation efficiency of the entire network. Information can be transmitted faster over the shortest path, saving system resources. The definition formula is:

Among them, L _i,j represents the shortest path between node i and node j, and L _p is the shortest path length of the entire network G. It can be seen that the calculation of the shortest path length must be based on the full connectivity of the network: if node i cannot reach node j through any way, then _{Li, j} does not exist or is infinite, and L _p also does not exist.

Global efficiency E _glob : describe the information transmission efficiency in the network, the global efficiency of node i is defined by the following formula on the basis of the shortest path length

It can be seen from the above formula that the smaller the shortest path of a node, the faster the information transfer between the node and other nodes, that is, the higher the global efficiency of the node.

And the global efficiency E _glob of the network G is defined as the average value of the global efficiency of all nodes in the network

The local efficiency E _loc is an important indicator to measure the compactness of the "clique" composed of adjacent nodes in the network, and also to describe the redundancy of the network and the tolerance of external attacks. The local efficiency of node i and the local efficiency of network G are defined as:

where L _jk is the shortest path length between region j and region k, G _i is the sub-network composed of nodes connected to region i, N _Gi is the number of brain regions in sub-network _Gi , and N is G in the whole-brain network the number of nodes in the .

The clustering coefficient _Cp is an important indicator to measure the cliquishness and the degree of local interconnectivity of the network. The clustering coefficient C(i) of node i is defined as the “other nodes” directly connected to node i in network G. The ratio between the number of edges between these "other nodes" and the maximum possible number of edges between these "other nodes", as defined by the following formula. The clustering coefficient _Cp of the network G is defined as the average of the clustering coefficients of all nodes.

A network is considered to have small-worldness if it has both a high clustering coefficient and a short shortest path length. In order to quantitatively determine whether the network has the small-world property, the clustering coefficient and the shortest path length of the network are generally compared with the corresponding properties of the random network.

The normalized aggregation coefficient γ and normalized shortest path length λ are calculated respectively according to the following formulas:

where C _p ^rand and L _p ^rand are the mean values of the clustering coefficients and the shortest path lengths of 100 random networks, respectively.

The small-world property σ is defined as σ=γ/λ. If γ>1 and λ≈1, that is, σ>1, the network is judged to have small-world properties.

The three node topological properties include node degree D _nodal (i), node efficiency E _nodal (i), node betweenness centrality B _nodal (i), which are defined as follows:

Among them, betweenness centrality defines the centrality of nodes from the perspective of information flow. In the formula _est represents the number of all shortest paths from node s to node t in network G, and e _sit is these shortest paths. The number of passing nodes i in .

In this step, the feature screening specifically includes:

The Recursive Feature Elimination (RFE) algorithm based on support vector machine performs feature selection by continuously training the classifier and removing the feature dimension with small feature weight, including:

① Initialize the selected feature set to contain all selected features,

② Take the feature set as input, train the classifier, and get the classification effect and the weight of each feature;

③ Remove the feature with the smallest weight to form a new feature set;

选取分类效果最好的情况。④Repeat ②, ③ until

Choose the case with the best classification effect.

6. Based on the characteristics after screening, the classifier is used for classification training to obtain the individualized prediction model of first-episode schizophrenia;

In this step, as a specific implementation means, the classifier adopts SVM classifier based on radial basis function (radial basis function, RBF) kernel, logistic regression classifier (Logistic Regression Classifier) or integrated learning of multiple classifiers; During classification training, a variety of pattern recognition classification methods are applied to find the optimal classifier model and screen the key multi-threshold fusion brain structure network features.

7. Perform performance verification evaluation on the individualized prediction model of first-episode schizophrenia obtained by training.

In this step, as a specific implementation method, cross-validation is used to evaluate the performance of various prediction models of schizophrenia training set and test set, including accuracy, sensitivity, specificity, ROC curve and AUC value, etc. and classification algorithms, apply permutation tests, optimize brain imaging feature selection and individual subtype prediction models, and identify subtype-related core brain imaging group features to improve individualized prediction accuracy.

To sum up, the present invention adopts the technology of artificial intelligence and machine learning, and constructs an individualized prediction model of first-episode schizophrenia with good robustness by analyzing and mining the magnetic resonance imaging data of the brain structure of first-episode schizophrenia. Early diagnosis and identification of schizophrenia can achieve accurate and objective auxiliary diagnosis and improve curative effect.

Claims (9)

1. The method for constructing the individualized prediction model of the first-onset schizophrenia is characterized by comprising the following steps of:

A. acquiring a diffusion tensor image of a patient with first schizophrenia;

B. preprocessing the acquired diffusion tensor image;

C. constructing a sparse brain structure network based on the preprocessed image;

D. constructing a multi-threshold fusion brain structure network after each tested sparse by adopting a similar network fusion method;

E. extracting network topology attribute features of the multi-threshold fusion brain structure, and then performing feature screening;

F. based on the screened features, performing classification training by adopting a classifier to obtain an individualized prediction model of the first schizophrenia;

G. and (4) performing performance verification evaluation on the trained individualized prediction model of the first schizophrenia.

2. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in the step A, a nuclear magnetic resonance imaging scanner is used for scanning and obtaining a diffusion tensor image of the first schizophrenia patient by adopting a single-shot plane echo imaging technology.

3. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in step B, the preprocessing of the acquired diffusion tensor image specifically includes:

b1, converting the dispersion tensor image in the DICOM data format into an image in an NIFTI format by using MRI convert;

b2, performing eddy current correction and head movement correction on the diffusion tensor image after data format conversion;

b3, removing skull by Brain Extraction Tool of FSL, and removing non-Brain tissue image.

4. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in step C, the constructing a sparse brain structure network based on the preprocessed image specifically includes:

c1, registering the brain images after the pre-processing of each test to b0 images in a diffusion tensor space by a rotation and translation linear registration method; the registered b0 images are then registered to T1 images in standard MNI space; inverting the conversion matrix, and transforming the AAL template from the MNI space to the diffusion tensor space by using the obtained inverse matrix to obtain 90 brain area network nodes divided based on the AAL template;

c2, using a probabilistic fiber bundle imaging method, carrying out Bayesian estimation of dispersion parameters based on a BEDPOSTX tool, establishing distribution of the dispersion parameters of each voxel by using a Markov chain Monte Carlo sampling method, presetting each voxel of a brain as a fiber cross model, and automatically judging how many kinds of cross-passing fiber bundles pass;

c3, carrying out probability tracking fiber bundle reconstruction based on a PROBTRACKX tool, repeatedly sampling the distribution of each voxel in the main diffusion direction, generating a streamline from the extracted local sample each time, and establishing a statistical graph of streamline position posterior distribution through multiple sampling to obtain the distribution situation of the structural connection probability between every two brain regions; defining the weight of each edge as the fiber bundle connection probability between every two node areas, and obtaining a symmetrical 90 x 90 fiber bundle connection probability weighting network matrix for each tested object;

and C4, setting a fiber bundle connection probability threshold, wherein structural connection exists between the two brain areas exceeding the threshold, testing the influence of different sparsity thresholds on the fusion effect, and constructing a sparse structure network by adopting a relatively narrower threshold range.

5. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in step D, the constructing of the multi-threshold fusion brain structure network after each tested sparse by using the similar network fusion method specifically includes:

d1, defining the sparse structure connection matrix as a full-core matrix

For the ith tested full-core matrix under the jth threshold

Further constructing a corresponding sparse kernel matrix:

order to_uIs a full-core matrix

K adjacent to the middle node u, then sparse kernel matrix

Is defined as:

d2, carrying out iterative updating of the full-kernel matrix based on the sparse kernel matrix corresponding to the full-kernel matrix:

wherein (W)_i ^c)^(m)Represents the full-kernel matrix under the ith tested c threshold value in the mth iteration, (W)_i ^j)^(m+1)Representing a full-kernel matrix in the (m + 1) th iteration, wherein N is the total sparse threshold number;

d3, judging whether the iteration convergence condition is met, if so, executing a step D4, otherwise, continuing the iteration;

wherein the convergence condition is: i (W)_i ^j)^(m+1)-(W_i ^j)^(m)||≤0.01；

D4, and corresponding the updated N thresholds to the full-core matrix

Averaging was performed to construct an averaged full kernel matrix for each test:

d5, mixing W_iNormalized to the interval [0, 1 ]]Thus generating a final fused network for each test.

6. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in step E, the extracting of the multi-threshold fusion brain structure network topology attribute feature specifically includes:

calculating AUC values of the fused network under all thresholds of 8 global topological attributes and 3 node topological attributes based on a graph theory analysis method as initial features of subsequent classification;

the 8 global topology attributes include network strength S_pGlobal efficiency E_globLocal efficiency E_locLength of shortest path L_pCoefficient of aggregation C_pNormalized shortest path length λ, normalized clustering coefficient γ, small world attribute σ:

the network strength S_pThe calculation formula of (2) is as follows:

wherein, S (i) is the weighted sum of edges connected with the ith node, and N is the number of brain areas in the whole brain network;

the shortest path length L_pThe calculation formula of (2) is as follows:

wherein L is_ijRepresenting the shortest path between node i and node j, L_pThe shortest path length for the entire network G;

the global efficiency E_globThe calculation formula (2) includes:

wherein E is_{glob_i}(G) The global efficiency of the node i is the average value of the global efficiency of all nodes in the network G;

the local efficiency E_locThe calculation formula (2) includes:

wherein L is_jkIs the shortest path length between region j and region k, G_iIs a sub-network of nodes connected to area i, N_GiIs a sub-network G_iThe number of midbrain regions; e_{loc_i}(G) Is the local efficiency of node i, the local efficiency E of network G_loc(G) The average value of the local efficiency of all nodes in the network is obtained;

the agglomeration coefficient C_pThe calculation formula (2) includes:

wherein, C (i) is the aggregation coefficient of the node i, and the aggregation coefficient of the network G is the average value of the aggregation coefficients of all the nodes;

the normalized aggregation coefficient γ and the normalized shortest path length λ are calculated as follows:

wherein, C_p ^randAnd L_p ^randThe average values of the aggregation coefficients and the shortest path lengths of 100 random networks are respectively;

the formula for calculating the small world attribute sigma is as follows:

σ＝γ/λ；

the 3 kinds of node topology attributes comprise a node degree D_nodal(i) Node efficiency E_nodal(i) Center degree of node betweenness B_nodal(i) Respectively defined as follows:

wherein e is_stRepresenting the number of all shortest paths from node s to node t in network G, e_sitIs the number of passing nodes i in these shortest paths.

7. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in step E, the performing feature screening specifically includes:

the recursive feature elimination algorithm based on the support vector machine performs feature selection by continuously training a classifier and removing feature dimensions with smaller feature weights, and specifically comprises the following steps:

① initializes the selected feature set to contain all of the selected features,

secondly, taking the feature set as input, training a classifier, and obtaining a classification effect and the weight of each feature;

removing the features with the minimum weight to form a new feature set;

④ repeat ②, ③ until

And selecting the condition with the best classification effect.

8. The method for constructing an individualized model for predicting first-onset schizophrenia according to claim 1, wherein the first-onset schizophrenia is a single-dose schizophrenia,

in the step F, the classifier adopts an SVM classifier, a logistic regression classifier or a plurality of classifiers integrated learning based on a radial basis function kernel; during classification training, a plurality of pattern recognition classification methods are applied to find an optimal classifier model and screen key multi-threshold fusion brain structure network characteristics.

9. The method for constructing the individualized prediction model for the first-onset schizophrenia according to any one of claims 1 to 8, wherein in the step G, cross validation is adopted to evaluate the performances of various prediction models of the training set and the test set of schizophrenia, including accuracy, sensitivity, specificity, ROC curve, AUC value and the like, and the system optimizes the brain image feature selection and the individual subtype prediction models by using the test feature screening and classification algorithm and applies the permutation test to identify the core brain image group features related to the subtypes so as to improve the individualized prediction accuracy.

Images