A method and device for ADHD case classification based on weighted sum strategy distributed algorithm

Abstract

The present invention provides an ADHD case classification method and device based on a weighted sum strategy distributed algorithm, and relates to the field of machine learning technology. The present invention uses fMRI data and classification technology, and proposes a distributed multi-objective optimization method based on a weighted strategy, and applies it to a binary ADHD soft margin support vector machine. The multi-objective support vector machine is converted into a single-objective optimization problem using the weighted sum method, and then it is iteratively solved using a distributed method. This method can effectively handle SVMs with multiple performance indicators, improve optimization efficiency through collaboration and coordination between multi-agent systems, and effectively solve the ADHD classification problem with data distributed on different service hosts. Since rs-fMRI shows its unique advantages in the analysis of mental illness, it can be used not only for the classification of ADHD, but also for the classification of schizophrenia and Alzheimer's cases.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to an ADHD case classification method and device based on a weighted sum strategy distributed algorithm. Background Art

[0002] A multi-agent system is a highly fault-tolerant and scalable network system composed of multiple agents with perception, execution, and computational capabilities, interacting with each other through information exchange. Distributed optimization emphasizes the method of technical implementation, while multi-agent systems emphasize the subject of technical implementation. Therefore, distributed optimization technology and multi-agent networks are closely related, inseparable, and mutually reinforcing. Compared with traditional optimization algorithms that rely on a central node, distributed optimization algorithms that are node-free and perform optimization decisions within the constraints of network nodes have attracted increasing research attention due to their scalability. They have been widely applied in energy communication networks, smart grids, big data, smart buildings, drones, and formation tracking. For distributed optimization problems in multi-agent systems, there are several distributed algorithms, which can be roughly divided into two categories: discrete-time distributed algorithms and continuous-time distributed algorithms. For discrete-time distributed algorithms, there are augmented Lagrangian algorithms, constrained optimal consensus algorithms, exact first-order algorithms, other algorithms based on consensus theory, and subgradient algorithms. These algorithms are first-order algorithms with constant or adaptive step sizes. For continuous-time distributed algorithms, PI (proportional integral controller) control method, primal-dual algorithm and collaborative neural dynamics algorithm have good performance under continuous-time control technology.

[0003] However, some research has shown that distributed optimization problems in multi-agent systems with local optimality consensus constraints cannot be solved using the above methods. Furthermore, these problems also involve multiple performance metrics, meaning they are essentially multi-objective optimization problems. Therefore, traditional distributed algorithms are unable to solve these problems, necessitating the introduction of multi-objective algorithms.

[0004] In multi-objective optimization, multi-objective optimization algorithms are widely used in fields such as machine learning, transportation networks, integrated circuit design, and mechanical design. Unlike single-objective optimization, multi-objective optimization requires simultaneous optimization of multiple objective functions. When solving multi-objective optimization problems, the goal is actually to find a set of Pareto optimal solutions (Pareto Optimality) that are both effective and compromise. The mapping of the effective solution set to the non-dominated set in the objective space, i.e., the Pareto frontier, fully reflects the trade-offs between multiple performance indicators and can effectively guide decision-making. Therefore, the goal of multi-objective optimization is to find the non-dominated set and provide decision-makers with as much information as possible to assist in decision-making analysis.

[0005] With the development of cloud computing, big data, and artificial intelligence technologies, multi-objective optimization problems are becoming increasingly large-scale and complex. Take machine learning, for example. The goal of machine learning is to learn models from data and then use them to perform tasks such as classification and regression. Optimization is the core of model training. Model selection criteria in machine learning are often conflicting, such as maximizing model generalization and minimizing training error. Therefore, optimization problems in machine learning are inherently multi-objective. Existing multi-objective optimization methods mostly employ scalarization techniques. By solving a single-objective optimization problem with parameters, Pareto optimal solutions can be obtained. Furthermore, by varying the parameters, multiple Pareto optimal solutions can be obtained. Compared to evolutionary algorithms, multi-objective optimization algorithms based on mathematical programming have the support of mathematical theory. These algorithms primarily include the multi-objective simplex method, interior point method, multi-weighted sum method, elastic constraint method, and objective space-based algorithms. The simplex method and interior point method are variable space-based algorithms that aim to obtain efficient solution sets. The multi-weighted sum method and elastic constraint method can obtain discrete non-dominated points by setting different weights and elastic constraint parameters. Objective-space-based algorithms include normal boundary projection, global mapping, normal vector constraint, and modified normal boundary projection. However, these multi-objective optimization algorithms are centralized. As the problem size increases, they become incapable of handling large-scale multi-objective optimization problems due to the limited computing power of a single machine. Furthermore, the objective functions and constraints in practical problem models are nonlinear, which complicates the optimization problem.

[0006] Attention deficit hyperactivity disorder (ADHD) is one of the most common psychiatric disorders in childhood, often persisting into adulthood. ADHD is defined in the DSM-5 as a neurodevelopmental disorder characterized by symptoms such as inattention, hyperactivity, and impulsive behavior. In recent years, several techniques have been developed to assist in the classification of ADHD cases, including the use of support vector machines (SVMs). However, the objectives in a support vector machine are generally conflicting, making it an inherently multi-objective optimization problem.

[0007] In the prior art, when the data set of ADHD cases is large and the data is distributed and stored on different service hosts, the existing multi-objective optimization classification method will no longer be applicable, and it is necessary to introduce a distributed optimization method for solution. Summary of the Invention

[0008] In response to the problem in the prior art that when the data set of ADHD cases is large and the data is distributed and stored on different service hosts, the existing multi-objective optimization classification method is no longer applicable. The present invention proposes an ADHD case classification method and device based on a weighted sum strategy distributed algorithm.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] In one aspect, a method for classifying ADHD cases based on a weighted and strategic distributed algorithm is provided, comprising:

[0011] S1: Obtain the test subject's brain image, preprocess the brain image, and obtain a brain region image dataset;

[0012] S2: Perform dimensionality reduction processing on the brain region image dataset to obtain a case sample dataset, which contains m data;

[0013] S3: Build a binary classification multi-objective support vector machine model; transform the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model into a single-objective support vector machine problem through the weighted sum method;

[0014] S4: Construct a distributed algorithm for the single-objective support vector machine problem and obtain a support vector machine model based on the distributed algorithm;

[0015] S5: Select z data from the case sample data set to form a training set, and train the support vector machine model based on the distributed algorithm; select mz data from the case sample data set to form a test set, and test the trained support vector machine model based on the distributed algorithm to obtain a tested support vector machine model based on the distributed algorithm; complete the classification of attention deficit hyperactivity disorder (ADHD) cases through the tested support vector machine model based on the distributed algorithm.

[0016] Optionally, in step S1, obtaining a brain image of the test subject and preprocessing the brain image to obtain a brain region image dataset includes:

[0017] S11: collecting multiple brain images of the test subject, discarding the first 10 brain images of the multiple brain images; and rearranging the remaining brain images;

[0018] S12: The realigned brain images were spatially normalized to a standard echo-planar imaging template and resampled into functional images;

[0019] S13: Processing the functional image to reduce spatial noise, eliminate low-frequency drift, and eliminate low-frequency filtering;

[0020] S14: The brain image was divided into 116 regions according to the anatomical automatic labeling (AAL) template. The functional magnetic resonance imaging (fMRI) time series of all voxels in the 116 regions were averaged using the functional connectivity (FC) method to obtain the average time series of each of the 116 regions.

[0021] S15: Calculate the Pearson correlation coefficient between the AAL template and the average time series of each region to obtain the functional connectivity matrix; the lower triangular data of the functional connectivity matrix is ​​used as the initial features of a single sample, and the initial features of all samples in the functional connectivity matrix constitute the brain region image dataset.

[0022] Optionally, in step S2, dimensionality reduction processing is performed on the brain region image dataset, including:

[0023] The principal component analysis (PCA) method was used to reduce the original feature dimension of the obtained brain region image dataset.

[0024] Optionally, step S3, constructing a binary classification multi-objective support vector machine model, includes:

[0025] S31: Based on the typical soft margin of the 2-norm, a preliminary model of a two-class multi-objective support vector machine is obtained. The preliminary model is described as follows:

[0026]

[0027] Among them, ω and b are decision variables and also the parameters of the classifier, ζ j Represents the training point x j The slack variable of x j Does not satisfy its corresponding constraint y j (ω T x j +b)≥1, y j is the training sample point x j The label used to determine x j The category to which it belongs; C represents the penalty parameter that determines the regularization cost of the two categories.

[0028] S32: Based on the characteristics that the data in the case sample dataset is distributed and stored in different intelligent agents, and the data volume is large, formula (1) is modeled as a binary classification multi-objective support vector machine model, such as the following formula (2):

[0029]

[0030] Where i∈{1,2,…,n} represents the i-th agent, and m is the number of training samples; is the slack variable.

[0031] Optionally, in step S3, the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model is converted into a single-objective support vector machine problem by using a weighted sum method, including:

[0032] By using the weighted sum method, the formula (2) of the multi-objective optimization problem is transformed into the formula (3) of the single-objective optimization problem.

[0033]

[0034] where w i is the positive weight factor corresponding to the i-th objective function, i.e. w i >0 and

[0035] Optionally, in step S4, a distributed algorithm for a single-objective support vector machine problem is constructed to obtain a support vector machine model based on the distributed algorithm, including:

[0036] S41: Based on formula (3), construct the Lagrangian function as shown in formula (4):

[0037]

[0038] Among them, λ and μ represent auxiliary variables, namely Lagrange multipliers;

[0039] S42: Design a distributed algorithm program based on the Lagerrangian function; the distributed algorithm program includes:

[0040] Input: h uniformly selected weight vectors w = {w 1 ,w 2 ,…,w h} T ,

[0041] in and the initialized decision variables and auxiliary variables x i [1],λ i [1],μ i [1],i=1,2,…,n;

[0042]

[0043] Output: x * ={x 1 [K+1],…,x h [K+1]} T and f(x * )={f(x 1 [K+1]),…,f(x h [K+1])} T ;

[0044] where x o [K+1]={x o,1 [K+1],…,x o,n [K+1]},o∈{1,2,…,h};P Ω represents the Euclidean projection of the feasible region of the decision variables; Represents the feasible region for the auxiliary variable λ (and R + ), α is a constant iteration step, and K is the maximum number of iteration steps;

[0045] S43: Applying the designed distributed algorithm program to the binary classification multi-objective support vector machine model to obtain a support vector machine model based on the distributed algorithm.

[0046] Optionally, in step S5, z data in the case sample data set are selected to form a training set, and a support vector machine model based on a distributed algorithm is trained, including:

[0047] S51: Preset the positive weight factors to equal weights, i.e., ω1=ω2=…=ω n , step size α = 0.002, and initial vector = 0;

[0048] S52: Select z data from the case sample data set to form a training set, input the training set into a support vector machine model based on a distributed algorithm; run the distributed algorithm to obtain an iterative trajectory diagram of the decision variable;

[0049] S53: The optimal hyperplane parameters are solved through the iteration of the distributed algorithm, the training set is classified, and the trained classifier is obtained, and the model training is ended.

[0050] Optionally, in step S5, mz data in the case sample data set are selected to form a test set, and the trained support vector machine model based on the distributed algorithm is tested to obtain a tested support vector machine model based on the distributed algorithm; and the classification optimization of attention deficit hyperactivity disorder (ADHD) cases is completed by using the tested support vector machine model based on the distributed algorithm, including:

[0051] mz data in the case sample data set are selected to form a test set. The trained classifier is used to perform classification tests on the test set of ADHD data distributed and stored in multiple service hosts. The classification of ADHD cases is completed through the tested support vector machine model based on the distributed algorithm.

[0052] In one aspect, an ADHD case classification device based on a weighted and strategy distributed algorithm is provided, the device comprising:

[0053] The data acquisition module is used to obtain the test subject's brain image, pre-process the brain image, and obtain a brain region image dataset;

[0054] A data dimension reduction module is used to perform dimension reduction processing on the brain region image dataset to obtain a case sample dataset, which contains m data;

[0055] The model building module is used to build a binary classification multi-objective support vector machine model; through the weighted sum method, the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model is converted into a single-objective support vector machine problem;

[0056] Distributed algorithm design module, used to construct a distributed algorithm for single-objective support vector machine problems and obtain a support vector machine model based on the distributed algorithm;

[0057] The model training module is used to select z data from the case sample data set to form a training set and train the support vector machine model based on the distributed algorithm; select mz data from the case sample data set to form a test set and test the trained model to complete the classification of ADHD cases.

[0058] Optionally, the data acquisition module is further configured to acquire brain images of the test subject, discard the first ten brain images, and rearrange the remaining brain images;

[0059] The realigned images were spatially normalized to a standard echo-planar imaging template and resampled into functional images;

[0060] Perform spatial noise reduction, low-frequency drift elimination, and low-frequency filtering operations on the functional image;

[0061] The brain images were divided into 116 regions according to the AAL template. The functional magnetic resonance imaging time series of all voxels in the regions were averaged using the functional connectivity FC method to obtain the average time series of each of the 116 regions.

[0062] The Pearson correlation coefficient between the AAL template and each average time series was calculated to obtain the functional connectivity matrix; the lower triangular data of the functional connectivity matrix was used as the initial features of a single sample, namely the brain region image dataset.

[0063] On the one hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned ADHD case classification method based on the weighted and strategy distributed algorithm.

[0064] On the one hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned ADHD case classification method based on the weighted and strategy distributed algorithm.

[0065] The above technical solutions of the embodiments of the present invention have at least the following beneficial effects:

[0066] In the above scheme, 1. The distributed multi-objective optimization method based on the weighted sum strategy proposed in this invention can be used to solve the multi-objective support vector machine problem in a multi-agent system. When the classification data is distributed and stored in different agents, the multi-objective problem is converted into a single-objective problem by selecting uniformly distributed weight factors that are all greater than 0, and a distributed solution is performed to obtain the Pareto optimal solution, thereby obtaining an approximately discrete and representative Pareto frontier of the problem, and this frontier has good uniformity and coverage indicators. This is something that traditional methods cannot achieve;

[0067] 2. This paper applies the proposed method to a binary support vector machine for ADHD classification in a multi-agent system, i.e., a binary support vector machine where data is distributed and stored in different agents and the objectives conflict with each other. By selecting different weight vectors, multiple classifiers can be trained. Decision makers can select the most appropriate classifier based on their needs or the classification performance indicators of the classifier.

[0068] 3. The convergence rate of the method proposed in the present invention reaches linear convergence, that is, when the data sets are distributed and stored in different intelligent agents, the classification optimization efficiency of the multi-objective support vector machine is also high enough. At the same time, the error level also proves that the classification accuracy of the method can be guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0070] Figure 1 is a flow chart of an ADHD case classification method based on a weighted sum strategy distributed algorithm provided by an embodiment of the present invention;

[0071] Figure 2 is a flow chart of an ADHD case classification method based on a weighted sum strategy distributed algorithm provided by an embodiment of the present invention;

[0072] Figure 3 This is a functional connectivity matrix acquisition flow chart of an ADHD case classification method based on a weighted sum strategy distributed algorithm provided by an embodiment of the present invention;

[0073] Figure 4 Statistical differences between the healthy control group and ADHD patients using the FC method provided by the embodiment of the present invention (left: active area detection results; right: 3D brain surface activation map);

[0074] Figure 5 The conventional support vector machine classification method based on regularized hyperparameter selection provided by the embodiment of the present invention;

[0075] Figure 6 A multi-objective support vector machine classification method in a multi-agent system provided by an embodiment of the present invention;

[0076] Figure 7 is an undirected graph used in the binary classification support vector machine experiment provided in an embodiment of the present invention;

[0077] Figure 8is a detailed description of the data set provided by the embodiment of the present invention;

[0078] Figure 9 It is a classifier parameter variable iteration trajectory diagram under a set of weights provided by an embodiment of the present invention;

[0079] Figure 10 is the classification result on the test set provided by the embodiment of the present invention;

[0080] Figure 11 It is the discrete approximate Pareto frontier obtained by solving the binary classification support vector machine provided by the embodiment of the present invention;

[0081] Figure 12 It is an undirected graph used in a binary classification support vector machine experiment on a brain region image dataset of ADHD provided by an embodiment of the present invention;

[0082] Figure 13 ω1, ω2, ..., ω5, ω trajectory diagram of classification hyperparameter decision variables obtained by classification training of a brain region image dataset under the first set of weight vectors provided by an embodiment of the present invention;

[0083] Figure 14 is a trajectory diagram of classification hyperparameter decision variables ω1, ω2, ..., ω5, ω obtained by classification training of a brain region image dataset under the second set of weight vectors provided by an embodiment of the present invention;

[0084] Figure 15 This is a device block diagram of an ADHD case classification device based on a weighted sum strategy distributed algorithm provided by an embodiment of the present invention;

[0085] Figure 16 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0086] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0087] like Figure 1 As shown, an embodiment of the present invention provides an ADHD case classification method based on a weighted sum strategy distributed algorithm, comprising:

[0088] S101: Acquire a brain image of the test subject, pre-process the brain image, and obtain a brain region image dataset;

[0089] S102: performing dimensionality reduction processing on the brain region image dataset to obtain a case sample dataset, where the case sample dataset includes m data;

[0090] S103: constructing a binary classification multi-objective support vector machine model; transforming the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model into a single-objective support vector machine problem through a weighted sum method;

[0091] S104: constructing a distributed algorithm for the single-objective support vector machine problem, and obtaining a support vector machine model based on the distributed algorithm;

[0092] S105: Select z data from the case sample data set to form a training set, and train the support vector machine model based on the distributed algorithm; select mz data from the case sample data set to form a test set, and test the trained support vector machine model based on the distributed algorithm to obtain a tested support vector machine model based on the distributed algorithm; complete the classification of attention deficit hyperactivity disorder (ADHD) cases through the tested support vector machine model based on the distributed algorithm.

[0093] Optionally, in step S101, obtaining a brain image of a test subject and preprocessing the brain image to obtain a brain region image dataset includes:

[0094] S111: collecting multiple brain images of the test subject, discarding the first 10 brain images of the multiple brain images; and rearranging the remaining brain images;

[0095] S112: The realigned brain images are spatially normalized to a standard echo-planar imaging template and resampled into functional images;

[0096] S113: performing processing on the functional image to reduce spatial noise, eliminate low-frequency drift, and eliminate low-frequency filtering;

[0097] S114: The brain image was divided into 116 regions according to the anatomical automatic labeling (AAL) template. The functional magnetic resonance imaging (fMRI) time series of all voxels in the 116 regions were averaged using the functional connectivity (FC) method to obtain the average time series of each of the 116 regions.

[0098] S115: Calculate the Pearson correlation coefficient between the AAL template and the average time series of each region to obtain a functional connectivity matrix; the lower triangular data of the functional connectivity matrix is ​​used as the initial features of a single sample, and the initial features of all samples in the functional connectivity matrix constitute the brain region image dataset.

[0099] Optionally, in step S102, dimensionality reduction processing is performed on the brain region image dataset, including:

[0100] The principal component analysis (PCA) method was used to reduce the original feature dimension of the obtained brain region image dataset.

[0101] Optionally, step S103, constructing a binary classification multi-objective support vector machine model, includes:

[0102] S131: Based on the typical soft margin of the 2-norm, a preliminary model of a binary multi-objective support vector machine is obtained. The preliminary model is described as follows:

[0103]

[0104] Among them, ω and b are decision variables and also the parameters of the classifier, ζ j Represents the training point x j The slack variable of x j Does not satisfy its corresponding constraint y j (ω T x j +b)≥1, y j is the training sample point x j The label used to determine x j The category to which it belongs; C represents the penalty parameter that determines the regularization cost of the two categories.

[0105] S132: Based on the characteristics that the data in the case sample dataset is distributed and stored in different intelligent agents, and the data volume is large, formula (1) is modeled as a binary classification multi-objective support vector machine model, such as the following formula (2):

[0106]

[0107] Where i∈{1,2,…,n} represents the i-th agent, and m is the number of training samples; is the slack variable.

[0108] Optionally, in step S103, the multi-objective support vector machine problem in the binary multi-objective support vector machine model is converted into a single-objective support vector machine problem by using a weighted sum method, including:

[0109] By using the weighted sum method, the formula (2) of the multi-objective optimization problem is transformed into the formula (3) of the single-objective optimization problem.

[0110]

[0111] where w i is the positive weight factor corresponding to the i-th objective function, i.e. w i >0 and

[0112] Optionally, in step S104, a distributed algorithm for a single-objective support vector machine problem is constructed to obtain a support vector machine model based on the distributed algorithm, including:

[0113] S141: Based on formula (3), construct the Lagrangian function as shown in formula (4):

[0114]

[0115] Among them, λ and μ represent auxiliary variables, namely Lagrange multipliers;

[0116] S142: Design a distributed algorithm program based on the Lagerrangian function; the distributed algorithm program includes:

[0117] Input: h uniformly selected weight vectors w = {w 1 ,w 2 ,…,w h} T ,

[0118] in and the initialized decision variables and auxiliary variables x i [1],λ i [1],μ i [1],i=1,2,…,n;

[0119]

[0120] Output: x * ={x 1 [K+1],…,x h [K+1]} T and f(x * )={f(x 1 [K+1]),…,f(x h [K+1])} T ;

[0121] where x o [K+1]={x o,1 [K+1],…,x o,n [K+1]},o∈{1,2,…,h};P Ω represents the Euclidean projection of the feasible region of the decision variables; Represents the feasible region for the auxiliary variable λ (and R + ), α is a constant iteration step, and K is the maximum number of iteration steps;

[0122] S143: Applying the designed distributed algorithm program to the binary classification multi-objective support vector machine model to obtain a support vector machine model based on the distributed algorithm.

[0123] Optionally, in step S105, z data in the case sample data set are selected to form a training set, and a support vector machine model based on a distributed algorithm is trained, including:

[0124] S151: Preset the positive weight factors to equal weights, i.e., ω1=ω2=…=ω n , step size α = 0.002, and initial vector = 0;

[0125] S152: Select z data from the case sample data set to form a training set, input the training set into a support vector machine model based on a distributed algorithm; run the distributed algorithm to obtain an iterative trajectory diagram of the decision variables;

[0126] S153: The optimal hyperplane parameters are solved through the iteration of the distributed algorithm, the training set is classified, and a trained classifier is obtained, thereby ending the model training.

[0127] Optionally, in step S105, mz data in the case sample data set are selected to form a test set, and the trained support vector machine model based on the distributed algorithm is tested to obtain a tested support vector machine model based on the distributed algorithm; and the classification optimization of attention deficit hyperactivity disorder (ADHD) cases is completed by using the tested support vector machine model based on the distributed algorithm, including:

[0128] mz data in the case sample data set are selected to form a test set. The trained classifier is used to perform classification tests on the test set of ADHD data distributed and stored in multiple service hosts. The classification of ADHD cases is completed through the tested support vector machine model based on the distributed algorithm.

[0129] In an embodiment of the present invention, rs-fMRI (resting state functional magnetic resonance imaging) data and classification technology are used, and a distributed multi-objective optimization method based on a weighted sum strategy is proposed, and applied to a binary ADHD soft margin support vector machine. The multi-objective support vector machine is converted into a single-objective optimization problem using the weighted sum method, and then it is iteratively solved using a distributed method. This method can effectively handle SVMs with multiple performance indicators, improve optimization efficiency through collaboration and coordination between multi-agent systems, and solve the ADHD classification problem with data distributed on different service hosts. Since rs-fMRI shows its unique advantages in the analysis of mental illness, it can be used not only for the classification of ADHD, but also for the classification of schizophrenia and Alzheimer's cases.

[0130] The embodiment of the present invention provides an ADHD case classification method based on a weighted sum strategy distributed algorithm, which can be implemented by an electronic device, which can be a terminal or a server. Figure 2 The flowchart of the ADHD case classification method based on the weighted sum strategy distributed algorithm is shown. The processing flow of the method may include the following steps:

[0131] S201: Acquire multiple brain images of the test subject, discard the first 10 brain images among the multiple brain images; and rearrange the remaining brain images.

[0132] In one feasible embodiment, the brain data is pre-processed based on the Data Processing Assistant for Resting State Functional Resonance (DPARSF) toolbox. Due to the instability of the initial signal and the subject's adaptation to the situation, the first 10 brain images need to be discarded.

[0133] S202: spatially normalizing the rearranged brain image to a standard echo planar imaging template and resampling it into a functional image;

[0134] S203: performing processing on the functional image to reduce spatial noise, eliminate low-frequency drift, and eliminate low-frequency filtering;

[0135] S204: Divide the brain image into 116 regions based on the anatomical automatic labeling (AAL) template; average the functional magnetic resonance imaging time series of all voxels in the 116 regions using the functional connectivity (FC) method to obtain the average time series of each of the 116 regions;

[0136] S205: Calculate the Pearson correlation coefficient between the AAL template and the average time series of each region to obtain a functional connectivity matrix; the lower triangular data of the functional connectivity matrix is ​​used as the initial features of a single sample, and the initial features of the samples in all functional connectivity matrices constitute a brain region image dataset.

[0137] In one implementation, the realigned images were first spatially normalized to a standard echo-planar imaging template and resampled to 3 × 3 × 3 mm³. Next, the functional images were spatially smoothed using a 4 × 4 × 4 mm³ full-width-high-width Gaussian kernel to reduce spatial noise. Subsequently, the fMRI data were temporally filtered from 0.01 to 0.08 Hz to eliminate the effects of low-frequency drift and high-frequency noise. After the above preprocessing, the brain images were segmented into 116 brain regions according to the AAL (Anatomical Automatic Labeling) template, with 90 regions located in the cerebrum and 26 in the cerebellum. The fMRI time series of all voxels in the region were averaged using the FC (functional connectivity) method to obtain the mean time series for each of the 116 regions. The Pearson correlation coefficient was calculated between each pair to obtain the functional connectivity matrix. The functional connectivity matrix is ​​a symmetric matrix. The lower triangular data of the matrix were used as the initial features for each sample. The dimension of the initial eigenvector is 6670(116×116 / 2)-116 / 2). Figure 3 As shown in Figure 1, a schematic diagram of obtaining the functional connectivity matrix is ​​given. The initial feature matrix size of the brain region image dataset is 85×6670.

[0138] In a feasible factual method, statistical analysis is used to show the differences in the brain between ADHD (Attention Deficit Hyperactivity Disorder) patients and healthy controls. We collected ten patient samples and ten control samples from the brain region image dataset and performed a random effect two-tailed two-sample t-test in the MATLAB simulation software SPM5. Using the FC analysis of the software SPM5 (Statistical Parametric Mapping), the statistical difference results were obtained, such as Figure 4 As shown in the figure, the red highlighted regions corresponding to the PCC (Posterior Cingulate Cortex), dPCC (dorsalside Posterior Cingulate Cortex), and dACC (dorsal side anterior Cingulate Cortex) were considered to have statistically significant differences between the two groups. This result further validates the feasibility of using the feature matrix to classify ADHD data.

[0139] S206: performing dimensionality reduction processing on the brain region image dataset to obtain a case sample dataset, where the case sample dataset includes m data;

[0140] In a feasible implementation, PCA (Principal Component Analysis) is used to reduce the original feature dimension of the obtained brain region image dataset.

[0141] In a feasible implementation, for the brain region image dataset, the number of original brain features used for classification is 6670, and the sample size is 85. The dimension of the original brain features is much higher than the number of samples. Therefore, dimensionality reduction is required to improve the classification performance of the classifier. For these small sample high-dimensional data, we compared three traditional methods, namely PCA, isometric mapping Isomap and LLE (Locally Linear Embedding) methods to reduce the dimension of the original features. After comparison, the present invention selects the PCA method as the dimensionality reduction method. Finally, the size of the feature matrix of the dataset is reduced to 85×20.

[0142] S207: Based on the typical soft margin of the 2-norm, a preliminary model of a binary multi-objective support vector machine is obtained. The preliminary model is described as follows:

[0143]

[0144] Among them, ω and b are decision variables and also the parameters of the classifier, ξ j Represents the training point x j The slack variable of x j Does not satisfy its corresponding constraint y j (ω T x j +b)≥1, y j is the training sample point x j The label used to determine x j The category to which it belongs; C represents the penalty parameter that determines the regularization cost of the two categories.

[0145] S208: Based on the characteristics that the data in the case sample dataset is distributed and stored in different intelligent agents, and the data volume is large, formula (1) is modeled as a binary classification multi-objective support vector machine model, such as the following formula (2):

[0146]

[0147] Where i∈{1,2,…,n} represents the i-th agent, and m is the number of training samples; is the slack variable.

[0148] In one feasible implementation, due to the contradiction between maximizing the separating hyperplane margin and minimizing regularization, Equation (1) can be viewed as a multi-objective optimization problem. However, when the data is distributed and stored in different agents, and the amount of data is very large, the traditional centralized method cannot handle it. Therefore, the multi-objective optimization problem is modeled as in Equation (2).

[0149] In an embodiment of the present invention, since the support vector machine requires a parameter adjustment process to achieve a good balance between point regularization and margin maximization. According to the performance indicators of the support vector machine, training the support vector machine on a two-class dataset must optimize three different objectives, namely, margin maximization, minimization of positive class regularization, and minimization of negative class regularization. Margin maximization is inconsistent with the other objectives, because improving the performance of the classifier in terms of margin maximization may cause the classifier's performance in terms of regularization to deteriorate. In addition, reducing the regularization of one of the classes will still result in poor regularization performance of the other class, so the latter two objectives are also contradictory. Therefore, training such a classifier can be regarded as a multi-objective optimization problem.

[0150] S209: Using the weighted sum method, transform the formula (2) of the multi-objective optimization problem into the formula (3) of the single-objective optimization problem.

[0151]

[0152] where w i is the positive weight factor corresponding to the i-th objective function, i.e. w i >0 and

[0153] In a feasible implementation, multiple Pareto optimal solutions can be obtained by solving formula (3) with different weight factors.

[0154] S210: Based on formula (3), construct the Lagrangian function as shown in formula (4):

[0155]

[0156] Among them, λ and μ represent auxiliary variables, namely Lagrange multipliers;

[0157] S211: Design a distributed algorithm based on the Lagerrangian function; the distributed algorithm includes:

[0158] Input: h uniformly selected weight vectors w = {w 1 ,w 2 ,…,w h} T ,

[0159] in and the initialized decision variables and auxiliary variables x i [1],λ i [1],μ i [1],i=1,2,…,n;

[0160]

[0161]

[0162] Output: x * ={x 1 [K+1],…,x h [K+1]} T and f(x * )={f(x 1 [K+1]),…,f(x h [K+1])} T ;

[0163] where x o [K+1]={x o,1 [K+1],…,x o,n [K+1]},o∈{1,2,…,h};P Ω represents the Euclidean projection of the feasible region of the decision variables; Represents the feasible region for the auxiliary variable λ (and R + ), α is a constant iteration step, and K is the maximum number of iteration steps;

[0164] S212: Applying the designed distributed algorithm program to the binary classification multi-objective support vector machine model to obtain a support vector machine model based on the distributed algorithm.

[0165] In the embodiment of the present invention, the main difference between the model classification method based on distributed algorithm and the traditional classification method based on regularized hyperparameter selection is shown in the attached figure. Figure 5 as well as Figure 6 As shown in .

[0166] In a feasible implementation, 1) for error analysis: For the convenience of proof, (5)-(9) in the above method can be written as the following compact form

[0167]

[0168] Where x and λ represent the decision variable and auxiliary variable, respectively, and Ω and U represent the feasible regions of x and λ, respectively. To further prove this, the present invention assumes certain properties that apply to the sequences and obtained from equations (10) and (11). These properties are crucial when constructing near-optimal solutions and analyzing error estimates.

[0169] Assumption 1: The subgradient used in equations (10) and (11) and is uniformly bounded (boundedness of subgradient), that is, there exists a constant d>0, for any k≥0,

[0170]

[0171] This assumption holds when the sets Ω and U are compact and the Lagrangian is convex or concave. Similarly, this assumption holds when the Lagrangian is an affine function of (x,λ).

[0172] Lemma 1 The sequences {x[k]} and {λ[k]} are generated by equations (9) and (10). For any x∈Ω,

[0173]

[0174] Furthermore, for any λ∈U,

[0175]

[0176] Under the assumption that subgradients are bounded, we establish some additional properties of the iterated sequences {x[k]} and {λ[k]}.

[0177] Lemma 2 The sequences {x[k]} and {λ[k]} are generated by equations (10) and (11). Assume that the subgradient is bounded and and is the iterative average given by:

[0178]

[0179] For any x∈Ω,λ∈U,k>1,

[0180]

[0181] The error level αd that appears in equations (12) and (13) 2 / 2 is actually a tight error bound on the performance of methods with constant step size α. This can be seen by considering a minimization problem of the form And f(x)=dmax|x i |, and by applying a The subgradient method, where e i is a column vector whose i-th entry is 1 and the other entries are 0.

[0182] The relationships mentioned in the previous lemmas are the key to finding the approximate saddle point using the average value of each iteration and the approximate saddle point bounds obtained by this method. Using these lemmas, we establish the average function value And the saddle point value L(x * ,λ * ). It also shows the relationship between and The function value at this average number of iterations and the saddle point value L(x * ,λ * ) between them.

[0183] 2) Convergence analysis of approximate saddle points

[0184] We now show how to construct an approximate saddle point using the sequences generated by (9) and (10). We use a simple averaging method to generate the approximate solution. In particular, we consider the iterative average and These averages provide approximate solutions to saddle point problems.

[0185] Proposition 1 Let Assumption 1 hold, and the sequences {x[k]} and {λ[k]} are generated by equations (10) and (11). (x * ,λ * ) is a saddle point of L(x,λ), for any k>1,

[0186]

[0187] For any k>1, and There are the following relationships,

[0188]

[0189] In formula (14), it shows that the limit of the average function value is It is based on the initial iteration variables x0 and λ0 from the Lagrangian function L to the saddle point (x * ,λ * ) changes. It is worth mentioning that the average function value Converges to the saddle point value L(x * ,λ * ), whose error level is αd 2 / 2.

[0190] S213: Preset the positive weight factors to equal weights, i.e., ω1=ω2=…=ωn , the iteration step is set to α = 0.002 and the initial vector (both are initially set to 0);

[0191] S214: Select z data from the case sample data set to form a training set, input the training set into a support vector machine model based on a distributed algorithm; run the distributed algorithm to obtain an iterative trajectory diagram of the decision variables;

[0192] S215: The optimal hyperplane parameters are solved through the iteration of the distributed algorithm, the training set is classified, and a trained classifier is obtained, thereby ending the model training.

[0193] S216: Select mz data from the case sample data set to form a test set, use the trained classifier to perform classification tests on the test set of ADHD data distributed and stored in multiple service hosts, and complete the classification of ADHD cases through the tested support vector machine model based on the distributed algorithm.

[0194] In a feasible implementation, in order to prove the feasibility of this method, we first optimize the classification of a binary support vector machine problem, assuming that all data are evenly distributed and stored in two agents, that is, an undirected graph consisting of two agents, n = 2 (as shown in the attached figure). Figure 8 The description of the dataset is shown in the Appendix. Figure 9 As shown, there are 3000 samples, 2500 of which are used as training sets, and the remaining 500 samples are used as test sets.

[0195] First, set the weight factors to equal weights w1 = 0.5, w2 = 0.5, the step size to α = 0.002 and the initial vector (they can be set to 0) and run the algorithm. The iterative trajectory of all decision variables is shown in the attached figure. Figure 10 Then the optimal hyperplane parameters (i.e. Pareto optimal solution) are used to classify the test set. The classification results are shown in the attached Figure 11 As shown in Figure 2, the classification accuracy is 91.2%. Finally, the obtained approximate non-dominated set is shown in the attached figure. Figure 12 shown.

[0196] After the previous solution was passed, the present invention optimized the classification of the binary support vector machine problem of the brain region image dataset in ADHD. The samples with label attributes of 1 and 3 in the dataset were marked as case samples (the label was reset to -1), and the samples with label attribute of 0 were marked as healthy samples (the label was reset to +1). Then, 21 samples were randomly selected from 85 samples as test samples, and the remaining 64 sample data were evenly divided into two categories, that is, assuming that these samples were distributed and stored in two intelligent bodies, n = 2 (as shown in the attached figure). Figure 13 shown).

[0197] First, set the weight factors to w1 = 0.2, w2 = 0.8, the step size to α = 0.0005 and the initial vector (both set to 0) and run the algorithm. The iterative trajectory of all decision variables is shown in the attached figure. Figure 14 As shown (the figure only shows the trajectories of the six classifier hyperparameter decision variables ω1, ω2,…, ω5, b).

[0198] Then change the weight factors to w1 = 0.6, w2 = 0.4, the step size to α = 0.0005 and the initial vector (both set to 0) and run the algorithm. The iterative trajectory of all decision variables is shown in the attached figure. Figure 13 The trained classifier was used to classify the test samples, achieving a classification accuracy of 95.23%. This demonstrates that even when ADHD data is distributed and stored across multiple service hosts, the proposed method not only effectively solves the multi-objective support vector machine problem, but also maintains high classification efficiency for the trained classifier.

[0199] In the embodiments of the present invention, 1. the proposed distributed multi-objective optimization method based on a weighted sum strategy can be used to solve the multi-objective support vector machine problem in a multi-agent system. When the classification data is distributed and stored in different agents, the multi-objective problem is converted into a single-objective problem by selecting uniformly distributed weight factors that are all greater than 0, and a distributed solution is performed to obtain a Pareto optimal solution, thereby obtaining an approximately discrete and representative Pareto frontier of the problem, and this frontier has good uniformity and coverage indicators. This is something that traditional methods cannot achieve;

[0200] 2. This paper applies the proposed method to a binary support vector machine for ADHD classification in a multi-agent system, i.e., a binary support vector machine where data is distributed and stored in different agents and the objectives conflict with each other. By selecting different weight vectors, multiple classifiers can be trained. Decision makers can select the most appropriate classifier based on their needs or the classification performance indicators of the classifier.

[0201] 3. The convergence rate of the method proposed in the present invention reaches linear convergence, that is, when the data sets are distributed and stored in different intelligent agents, the classification optimization efficiency of the multi-objective support vector machine is also high enough. At the same time, the error level also proves that the classification accuracy of the method can be guaranteed.

[0202] Figure 15 1 is a block diagram of an ADHD case classification device based on a weighted sum strategy distributed algorithm according to an exemplary embodiment. Figure 15 , the apparatus 300 comprises:

[0203] The data acquisition module 301 is used to obtain the brain image of the test subject, pre-process the brain image, and obtain a brain region image dataset;

[0204] The data dimension reduction module 302 is used to perform dimension reduction processing on the brain region image dataset to obtain a case sample dataset, where the case sample dataset includes m data;

[0205] The model building module 303 is used to build a binary classification multi-objective support vector machine model; by using the weighted sum method, the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model is converted into a single-objective support vector machine problem;

[0206] A distributed algorithm design module 304 is used to construct a distributed algorithm for a single-objective support vector machine problem and obtain a support vector machine model based on the distributed algorithm;

[0207] The model training module 305 is used to select z data from the case sample data set to form a training set, and train the support vector machine model based on the distributed algorithm; select mz data from the case sample data set to form a test set, test the trained model, and complete the classification of ADHD cases.

[0208] Optionally, the data acquisition module 301 is further configured to acquire brain images of the test subject, discard the first ten brain images, and rearrange the remaining brain images;

[0209] The realigned images were spatially normalized to a standard echo-planar imaging template and resampled into functional images;

[0210] Perform spatial noise reduction, low-frequency drift elimination, and low-frequency filtering operations on the functional image;

[0211] The brain images were divided into 116 regions according to the AAL template. The functional magnetic resonance imaging time series of all voxels in the regions were averaged using the functional connectivity FC method to obtain the average time series of each of the 116 regions.

[0212] The Pearson correlation coefficient between the AAL template and each average time series was calculated to obtain the functional connectivity matrix; the lower triangular data of the functional connectivity matrix was used as the initial features of a single sample, namely the brain region image dataset.

[0213] Optionally, the data dimension reduction module 302 is configured to use a principal component analysis (PCA) method to reduce the original feature dimension of the obtained brain region image dataset.

[0214] Optionally, the model building module 303 is used to obtain a preliminary model of a binary multi-objective support vector machine based on a typical soft margin of the 2-norm. The preliminary model is described as follows:

[0215]

[0216] Among them, ω and b are decision variables and also the parameters of the classifier, ζ j Represents the training point x j The slack variable of x j Does not satisfy its corresponding constraint y j (ω T x j +b)≥1, y j is the training sample point x j The label used to determine x j The category to which it belongs; C represents the penalty parameter that determines the regularization cost of the two categories;

[0217] According to the characteristics that the data in the case sample dataset is distributed and stored in different intelligent agents, and the data volume is large, formula (1) is modeled as a binary classification multi-objective support vector machine model, as shown in the following formula (2):

[0218]

[0219]

[0220] Where i∈{1,2,…,n} represents the i-th agent, and m is the number of training samples; is the slack variable.

[0221] Optionally, the model building module 303 is further configured to transform the formula (2) of the multi-objective optimization problem into the formula (3) of the single-objective optimization problem by using a weighted sum method.

[0222]

[0223] where w i is the positive weight factor corresponding to the i-th objective function, i.e. w i >0 and

[0224] Optionally, the distributed algorithm design module 304 is used to construct a Lagrangian function such as the following formula (4) based on formula (3):

[0225]

[0226] Among them, λ and μ represent auxiliary variables, namely Lagrange multipliers;

[0227] Design a distributed algorithm program based on the Lagerrange function; the distributed algorithm program includes:

[0228] Input: h uniformly selected weight vectors w = {w 1 ,w 2 ,…,w h} T ,

[0229] in and the initialized decision variables and auxiliary variables x i [1],λ i [1],μ i [1],i=1,2,…,n;

[0230]

[0231]

[0232] Output: x * ={x 1 [K+1],…,x h [K+1]} T and f(x * )={f(x 1 [K+1]),…,f(x h [K+1])} T ;

[0233] where x o [K+1]={x o,1 [K+1],…,x o,n [K+1]},o∈{1,2,…,h};P Ω represents the Euclidean projection of the feasible region of the decision variables; Represents the feasible region for the auxiliary variable λ (and R + ), α is a constant iteration step, and K is the maximum number of iteration steps;

[0234] The designed distributed algorithm program is applied to the binary classification multi-objective support vector machine model to obtain a support vector machine model based on the distributed algorithm.

[0235] Optionally, the model training module 305 is configured to preset the positive weight factors to be equal weights, ie, ω1=ω2=…=ω n , step size α = 0.002, and initial vector = 0;

[0236] Select z data from the case sample data set to form a training set, and input the training set into the support vector machine model based on the distributed algorithm; run the distributed algorithm to obtain the iterative trajectory diagram of the decision variable;

[0237] The optimal hyperplane parameters are solved through the iteration of the distributed algorithm, the training set is classified, and the trained classifier is obtained, thus completing the model training.

[0238] Optionally, the model training module is used to select mz data from the case sample data set to form a test set, and perform classification tests on the test set of ADHD data distributed and stored in multiple service hosts through the trained classifier, and complete the classification of ADHD cases through the tested support vector machine model based on the distributed algorithm.

[0239] In the embodiments of the present invention, 1. the proposed distributed multi-objective optimization method based on a weighted sum strategy can be used to solve the multi-objective support vector machine problem in a multi-agent system. When the classification data is distributed and stored in different agents, the multi-objective problem is converted into a single-objective problem by selecting uniformly distributed weight factors that are all greater than 0, and a distributed solution is performed to obtain a Pareto optimal solution, thereby obtaining an approximately discrete and representative Pareto frontier of the problem, and this frontier has good uniformity and coverage indicators. This is something that traditional methods cannot achieve;

[0240] 2. This paper applies the proposed method to a binary support vector machine for ADHD classification in a multi-agent system, i.e., a binary support vector machine where data is distributed and stored in different agents and the objectives conflict with each other. By selecting different weight vectors, multiple classifiers can be trained. Decision makers can select the most appropriate classifier based on their needs or the classification performance indicators of the classifier.

[0241] 3. The convergence rate of the method proposed in the present invention reaches linear convergence, that is, when the data sets are distributed and stored in different intelligent agents, the classification optimization efficiency of the multi-objective support vector machine is also high enough. At the same time, the error level also proves that the classification accuracy of the method can be guaranteed.

[0242] Figure 16 1 is a schematic diagram of the structure of an electronic device 400 provided in an embodiment of the present invention. The electronic device 400 may vary significantly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 401 and one or more memories 402. The memories 402 store at least one instruction, which is loaded and executed by the processor 401 to implement the following steps of the method for ADHD case classification based on a distributed algorithm:

[0243] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions. The instructions are executable by a processor in a terminal to implement the above-described method for ADHD case classification based on a distributed algorithm. For example, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, or the like.

[0244] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0245] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for ADHD case classification based on a weighted sum strategy distributed algorithm, characterized in that: include: S1: Acquire a brain image of a test subject, preprocess the brain image, and obtain a brain region image dataset; S2: performing dimensionality reduction processing on the brain region image dataset to obtain a case sample dataset, wherein the case sample dataset includes m data, where m is the number of training samples; S3: constructing a binary classification multi-objective support vector machine model; converting the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model into a single-objective support vector machine problem by using a weighted sum method; The step S3, constructing a binary classification multi-objective support vector machine model, includes: S31: Based on the typical soft margin of the 2-norm, a preliminary model of a binary multi-objective support vector machine is obtained. The preliminary model is described as follows: Among them, ω and b are decision variables and also the parameters of the classifier, ξ j Represents the training sample point x j The slack variable of x j Does not satisfy its corresponding constraint y j (ω T x j +b)≥1, y j is the training sample point x j The label used to determine x j The category to which it belongs; C represents the penalty parameter that determines the regularization cost of the two categories; S32: Based on the characteristics that the data in the case sample dataset is distributed and stored in different intelligent agents, and the data volume is large, the formula (1) is modeled as a binary classification multi-objective support vector machine model, such as the following formula (2): Where i∈{1,2,…,n} represents the i-th agent, n is a positive integer value, and m is the number of training samples; is the slack variable; In step S3, the multi-objective support vector machine problem in the binary multi-objective support vector machine model is converted into a single-objective support vector machine problem by using a weighted sum method, including: By using the weighted sum method, the formula (2) of the multi-objective optimization problem is transformed into the formula (3) of the single-objective optimization problem. where w i is the positive weight factor corresponding to the i-th objective function, i.e. w i >0 and S4: constructing a distributed algorithm for the single-objective support vector machine problem, and obtaining a support vector machine model based on the distributed algorithm; In step S4, a distributed algorithm for the single-objective support vector machine problem is constructed to obtain a support vector machine model based on the distributed algorithm, including: S41: Based on formula (3), construct the Lagrangian function as shown in formula (4): Among them, λ and μ represent auxiliary variables, namely Lagrange multipliers; S42: Designing a distributed algorithm program based on the Lagerrangian function; the distributed algorithm program includes: Input: h uniformly selected weight vectors w = {w 1 ,w 2 ,…,w h } T , in and the initialized decision variables and auxiliary variables x i [1],λ i [1],μ i [1],i=1,2,…,n; Output: x * ={x 1 [K+1],…,x h [K+1]} T and f(x * )={f(x 1 [K+1]),…,f(x h [K+1])} T ; where x o [K+1]={x o,1 [K+1],…,x o,n [K+1]},o∈{1,2,…,h};P Ω represents the Euclidean projection of the feasible region of the decision variables; Represents the feasible region for the auxiliary variable λ (ie R + ), α is a constant iteration step, and K is the maximum number of iteration steps; S43: applying the designed distributed algorithm program to the binary classification multi-objective support vector machine model to obtain a support vector machine model based on the distributed algorithm; S5: Select z training data from the case sample data set to form a training set, and train the support vector machine model based on the distributed algorithm; select mz data from the case sample data set to form a test set, and test the trained support vector machine model based on the distributed algorithm to obtain a tested support vector machine model based on the distributed algorithm; and complete the classification of attention deficit hyperactivity disorder (ADHD) cases through the tested support vector machine model based on the distributed algorithm.

2. The ADHD case classification method based on the weighted sum strategy distributed algorithm according to claim 1, characterized in that: In step S1, a brain image of a test subject is obtained, and the brain image is preprocessed to obtain a brain region image dataset, including: S11: collecting multiple brain images of the test subject, discarding the first 10 brain images of the multiple brain images; and rearranging the remaining brain images; S12: spatially normalizing the rearranged brain image to a standard echo planar imaging template and resampling the image to a functional image; S13: performing processing on the functional image to reduce spatial noise, eliminate low-frequency drift, and eliminate low-frequency filtering; S14: The brain image was divided into 116 regions according to the anatomical automatic labeling (AAL) template. The functional magnetic resonance imaging (fMRI) time series of all voxels in the 116 regions were averaged using the functional connectivity (FC) method to obtain the average time series of each of the 116 regions. S15: Calculate the Pearson correlation coefficient between the AAL template and the average time series of each region to obtain a functional connectivity matrix; the lower triangular data of the functional connectivity matrix is ​​used as the initial features of a single sample, and the initial features of the samples in all functional connectivity matrices constitute a brain region image dataset.

3. The ADHD case classification method based on the weighted sum strategy distributed algorithm according to claim 1, characterized in that: In step S2, dimensionality reduction processing is performed on the brain region image dataset, including: The principal component analysis (PCA) method was used to reduce the original feature dimension of the obtained brain region image dataset.

4. The ADHD case classification method based on the weighted sum strategy distributed algorithm according to claim 3, characterized in that: In step S5, z data in the case sample data set are selected to form a training set, and the support vector machine model based on the distributed algorithm is trained, including: S51: Preset the positive weight factors to equal weights, i.e. ω1=ω2=....=ω n , step size α = 0.002, and initial vector = 0; S52: selecting z data from the case sample data set to form a training set, inputting the training set into the support vector machine model based on the distributed algorithm; running the distributed algorithm to obtain an iterative trajectory diagram of the decision variables; S53: Optimal hyperplane parameters are solved through iteration of the distributed algorithm, the training set is classified, a trained classifier is obtained, and model training is ended.

5. The ADHD case classification method based on the weighted sum strategy distributed algorithm according to claim 4, characterized in that: In step S5, mz data in the case sample data set are selected to form a test set, and the trained support vector machine model based on the distributed algorithm is tested to obtain a tested support vector machine model based on the distributed algorithm; The tested support vector machine model based on the distributed algorithm is used to optimize the classification of attention deficit hyperactivity disorder (ADHD) cases, including: mz data from the case sample data set are selected to form a test set. The test set of ADHD data distributed and stored in multiple service hosts is classified and tested by the trained classifier. The classification of ADHD cases is completed by the tested support vector machine model based on the distributed algorithm.

6. An ADHD case classification device based on a weighted sum strategy distributed algorithm, characterized in that: The device comprises: A data acquisition module is used to obtain a brain image of a test subject, pre-process the brain image, and obtain a brain region image dataset; A data dimensionality reduction module is used to perform dimensionality reduction processing on the brain region image dataset to obtain a case sample dataset, wherein the case sample dataset includes m data, where m is the number of training samples; A model building module is used to build a binary classification multi-objective support vector machine model; by using a weighted sum method, the multi-objective support vector machine problem in the binary classification multi-objective support vector machine model is converted into a single-objective support vector machine problem; The model building module is used to obtain a preliminary model of a binary multi-objective support vector machine based on the typical soft margin of the 2-norm. The preliminary model is described as follows: Among them, ω and b are decision variables and also the parameters of the classifier, ξ j Represents the training sample point x j The slack variable of x j Does not satisfy its corresponding constraint y j (ω T x j +b)≥1, y j is the training sample point x j The label used to determine x j The category to which it belongs; c represents the penalty parameter that determines the regularization cost of the two categories; According to the characteristics that the data in the case sample dataset is distributed and stored in different intelligent agents, and the data volume is large, formula (1) is modeled as a binary classification multi-objective support vector machine model, as shown in the following formula (2): Where i∈{1,2,…,n} represents the i-th agent, n is a positive integer value, and m is the number of training samples; is the slack variable; The model building module is used to transform the formula (2) of the multi-objective optimization problem into the formula (3) of the single-objective optimization problem through the weighted sum method. where w i is the positive weight factor corresponding to the i-th objective function, i.e. w i >0 and A distributed algorithm design module is used to construct a distributed algorithm for the single-objective support vector machine problem and obtain a support vector machine model based on the distributed algorithm; The distributed algorithm design module is used to construct the Lagrangian function as shown in the following formula (4) based on formula (3): Among them, λ and μ represent auxiliary variables, namely Lagrange multipliers; Design a distributed algorithm program based on the Lagerrange function; the distributed algorithm program includes: Input: h uniformly selected weight vectors w = {w 1 ,w 2 ,…,w h } T , in and the initialized decision variables and auxiliary variables x i [1],λ i [1],μ i [1],i=1,2,…,n; Output: x * ={x 1 [K+1],…,x h [K+1]} T and f(x * )={f(x 1 [K+1]),…,f(x h [K+1])} T ; where x o [K+1]={x o,1 [K+1],…,x o,n [K+1]},o∈{1,2,…,h};P Ω represents the Euclidean projection of the feasible region of the decision variables; Represents the feasible region for the auxiliary variable λ (ie R + ), α is a constant iteration step, and K is the maximum number of iteration steps; Applying the obtained designed distributed algorithm program to the binary classification multi-objective support vector machine model to obtain a support vector machine model based on the distributed algorithm; The model training module is used to select z training data from the case sample data set to form a training set, and train the support vector machine model based on the distributed algorithm; select mz data from the case sample data set to form a test set, and test the trained model to complete the classification of ADHD cases.

7. The ADHD case classification device based on the weighted sum strategy distributed algorithm according to claim 6, characterized in that: The data acquisition module is further configured to acquire brain images of the test subject, discard the first ten brain images, and rearrange the remaining brain images; normalizing the rearranged image spatially to a standard echo planar imaging template and resampling the rearranged image to a functional image; performing spatial noise reduction, low-frequency drift elimination, and low-frequency filtering operations on the functional image; The brain images were divided into 116 regions according to the AAL template. The functional magnetic resonance imaging time series of all voxels in the regions were averaged using the functional connectivity FC method to obtain the average time series of each of the 116 regions. The Pearson correlation coefficient between the AAL template and each of the average time series is calculated to obtain a functional connectivity matrix; the lower triangular data of the functional connectivity matrix is ​​used as the initial features of a single sample, namely, a brain region image dataset.

Images