Axis type spondyloarthritis auxiliary diagnosis and analysis system based on PSO-TabNet

Abstract

The invention relates to the field of middle-axis type spinal arthritis diagnosis, in particular to a middle-axis type spinal arthritis auxiliary diagnosis and analysis system based on PSO-TabNet, and the system comprises a data collection unit which is used for collecting clinical feature data and iconography feature data; the data preprocessing unit is used for preprocessing the feature data; the axSpA prediction model construction unit is used for optimizing a TabNet model hyper-parameter by using a particle swarm algorithm to obtain a PSO-TabNet model; and the prediction analysis unit is used for performing axSpA classification prediction on the clinical data based on the PSO-TabNet model to obtain auxiliary diagnosis suggestions. According to the invention, the PSO algorithm is adopted to optimize the hyper-parameters in the TabNet model, the PSO algorithm can automatically adjust the hyper-parameters, manual intervention is not needed, the complexity of hyper-parameter adjustment is reduced, the stability and effect of the model are improved, and doctors can be more reliably assisted in diagnosing diseases.

Description

Technical Field

[0001] The present invention relates to the field of axial spondyloarthritis diagnosis, and in particular to an axial spondyloarthritis auxiliary diagnosis and analysis system based on PSO-TabNet. Background Art

[0002] Axial spondyloarthritis (axSpA) is a chronic inflammatory arthritis that primarily affects the spine and sacroiliac joints. The disease causes inflammation of the joints in the spine and pelvis, which may fuse over time, completely losing the flexibility of the spine, severely limiting the range of motion of the spine, and affecting breathing and other body functions. Therefore, timely diagnosis is crucial to the effective management of axial spondyloarthritis, helping to slow the progression of the disease, improve quality of life, and prevent long-term complications.

[0003] Because the early symptoms of axSpA are difficult to accurately judge with a single symptom, for example, although the presence of the HLA-B27 antigen is associated with the disease, not all patients carrying the antigen will develop the disease. Therefore, the accurate diagnosis of axSpA usually requires the combination of other clinical data and imaging data, among which clinical data include gender, age, course of disease, HLA-B27, erythrocyte sedimentation rate and C-reactive protein. The imaging data needs to be extracted by manually interpreting the magnetic resonance imaging images of both sides of the sacroiliac joints, mainly including erosion, sclerosis, joint space enhancement, ankylosis, fatty lesions and bone marrow edema.

[0004] Currently, there are the following challenges in analyzing heterogeneous data such as clinical data: data sparsity, mixed attribute types, lack of locality and prior knowledge about data structure, and lack of interpretability. As a standardized deep table learning architecture, the TabNet model can handle the above problems well and provide higher prediction accuracy and better model interpretation capabilities in the diagnosis of various diseases. The TabNet model can effectively process medical data containing multiple features. Through its internal feature selection mechanism, it automatically identifies the most important features for prediction, reducing reliance on feature engineering. It can also provide model interpretability to help medical staff understand the decision-making process of the model.

[0005] The TabNet model is very sensitive to hyperparameters and requires tuning of multiple hyperparameters, such as learning rate, depth of decision tree, parameters of attention mechanism, etc. Manual or random selection of parameters is very cumbersome. Grid search (GS) is the most commonly used hyperparameter optimization method in artificial intelligence applications, but it is very time-consuming. The random search (RS) method is an improvement on GS in terms of efficiency. This method proposes to randomly sample points from the hyperparameter search space. The search continues until the predetermined allocation is exhausted. Although more efficient than GS, it is still time-consuming compared to other methods, so effective parameter adjustment methods are urgently needed to improve model training results.

[0006] The particle swarm optimization (PSO) algorithm is an optimization algorithm that simulates the foraging behavior of bird flocks in nature. The PSO algorithm searches for the optimal solution by keeping the current position, speed, and historical best position of each particle. Particles update their speed and position based on their own best experience and the best experience of their neighbors, thereby finding an approximate optimal solution in the solution space. The PSO algorithm has the following advantages: (1) PSO can handle these complex hyperparameter spaces, and is not prone to falling into local optimality like some traditional optimization methods; (2) PSO accelerates convergence by updating particles in parallel, and usually finds a better parameter configuration faster than other optimization algorithms (such as GS or RS); (3) The implementation of the PSO algorithm is relatively simple and easy to adjust. It can help the TabNet model achieve optimal performance and improve prediction accuracy and generalization ability, which is particularly important for the axSpA prediction task. Summary of the invention

[0007] Based on the advantages of the PSO algorithm, the present invention proposes an axSpA auxiliary diagnosis and analysis system based on PSO-TabNet, which uses the PSO algorithm to perform multiple iterative optimizations on the hyperparameters in the TabNet model, thereby improving the accuracy of model predictions, so as to more effectively assist doctors in diagnosing diseases.

[0008] In order to achieve the above object, the technical solution specifically adopted by the present invention is as follows:

[0009] An auxiliary diagnosis and analysis system for axial spondyloarthritis based on PSO-TabNet, comprising:

[0010] A data collection unit, used for collecting clinical characteristic data and imaging characteristic data;

[0011] A data preprocessing unit, used to preprocess the collected clinical characteristic data and imaging characteristic data;

[0012] axSpA prediction model building unit, which is used to optimize TabNet model hyperparameters using particle swarm (PSO) algorithm to obtain PSO-TabNet model;

[0013] A prediction and analysis unit is used to perform axSpA classification prediction on clinical data based on the PSO-TabNet model and obtain auxiliary diagnosis suggestions through the system.

[0014] Furthermore, the workflow of the axSpA prediction model building unit is as follows:

[0015] Step 1, obtain the axSpA dataset, which includes clinical feature data and imaging feature data;

[0016] Step 2: Preprocess the data, including data cleaning, processing of missing values ​​and outliers;

[0017] Step 3. Use the PSO algorithm to optimize the TabNet model hyperparameters and obtain the optimal hyperparameter set.

[0018] Furthermore, the clinical characteristic data include gender, age, disease course, HLA-B27, erythrocyte sedimentation rate and C-reactive protein; the imaging characteristic data are extracted by manually interpreting the magnetic resonance images of both sides of the sacroiliac joints, including erosion, sclerosis, joint space enhancement, ankylosis, fatty lesions and bone marrow edema.

[0019] Furthermore, the step Step 2 specifically includes the following contents:

[0020] Missing value cleaning: determine the missing value range, calculate the missing value ratio for each field, and formulate strategies based on the missing value ratio and field importance;

[0021] Missing value filling: Different filling methods are used for missing values ​​of different features;

[0022] Re-data collection: When certain features are very important but the missing ratio is too high, it is necessary to contact the outpatient department and re-data collection;

[0023] Logical error cleaning: Delete some data that can be directly found by simple logical reasoning to prevent the analysis results from being biased;

[0024] Non-demand data cleaning: that is, deleting unnecessary fields;

[0025] Outlier processing: Z-score standardization method is used for normalization to identify outliers, and then the following processing methods are used: outliers are removed, replaced with other values ​​(such as mean or median), and filled using interpolation.

[0026] Furthermore, the step Step 3 specifically includes the following steps:

[0027] 1) Initialize the particle swarm and initialize the position and velocity of each particle;

[0028] 2) Initialize the parameters in the PSO algorithm: acceleration factors c1, c2; inertia weight ω; maximum number of iterations; maximum and minimum speed values ​​V max ,V min ;

[0029] 3) Determine the hyperparameters that need to be optimized in the TabNet model;

[0030] 4) Assign the initial positions of all randomly generated particles to the hyperparameters to be optimized of the TabNet model;

[0031] 5) Input the data set into the TabNet model for training and record the prediction results;

[0032] 6) Use the prediction results to calculate the fitness value of each particle through formula (1): i , get the particle with the best fitness, that is, the particle with the best fitness;

[0033]

[0034] Where N is the number of samples in the test set, y i represents the actual value in the test set, represents the predicted value of the corresponding sample;

[0035] 7) Obtain the individual extreme value pbest of the particle with the best fitness value among the particles with the best fitness value i and the group extreme value gbest i , and update the speed and position of the particle with the optimal fitness value according to formulas (2) and (3) to obtain the updated position;

[0036] v i (t+1)=ω×v i (t)+c1×r1×(pbest i -x i (t+c2×r2)×(gbest i -x i (t)) (2)

[0037] x i (t+1)=x i (t)+v i (t) (3)

[0038] Among them, v i (t+1) and v i(t) are the speeds of particle i at iterations t+1 and t respectively; ω is the inertia factor, which is used to control the speed of the particle; r1 and r2 are random numbers between (0,1); c1 and c2 are learning factors, usually c1=c2=2; learning factors c1 and c2 control how particles update their positions. Specifically, c1 is the individual learning factor, which determines the force with which particles move to their own optimal position; and c2 is the social learning factor, which controls the force with which particles move to the optimal position of all particles; the balance of these two factors helps particles find the optimal solution; x i (t+1) and x i (t) are the positions of particle i at iterations t+1 and t, respectively;

[0039] 8) Assign the updated position to the selected hyperparameter to be optimized in the TabNet model;

[0040] 9) Repeat steps 6)-9) to determine whether the maximum number of iterations has been reached. If so, count the optimal particles in all iterations and proceed to the next step. If not, return to step 6);

[0041] 10) Apply the optimal solution of particles to the hyperparameter setting of the TabNet model.

[0042] Furthermore, the hyperparameters that need to be optimized in step 3) include: the width of the decision prediction layer and the width of the attention embedding of each mask; the number of steps in the architecture; the feature reuse coefficient in the mask; the number of shared GLUs in each step; the sparsity loss coefficient; the maximum training; the number of instances per batch; and the mini-batch for BN.

[0043] Furthermore, when applied, the following steps are included:

[0044] Step 1, data collection;

[0045] Step 2, data preprocessing: including data cleaning, processing of missing values ​​and outliers;

[0046] Step 3, input the processed data into the PSO-TabNet model;

[0047] Step 4: Get the prediction results for doctors to assist in diagnosis.

[0048] Compared with the prior art, the present invention has the following characteristics and beneficial effects:

[0049] 1) In the present invention, the PSO algorithm is used to optimize the hyperparameters in the TabNet model, helping the TabNet model to achieve optimal performance and improve prediction accuracy and generalization ability, which is particularly important for axSpA analysis and prediction.

[0050] 2) In the present invention, the optimized TabNet model is used for classification prediction, and the features that are more important for prediction can be identified more accurately and quickly through an interpretable feature selection mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0052] Figure 1 This is a system block diagram of an axial spondyloarthritis auxiliary diagnosis and analysis system based on PSO-TabNet according to an embodiment of the present invention.

[0053] Figure 2 The figure is a flowchart of the PSO algorithm in the present invention for obtaining the optimal hyperparameters of the TabNet model.

[0054] Figure 3 It is a structural diagram of the TabNet model in the present invention. DETAILED DESCRIPTION

[0055] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0056] like Figure 1 As shown, a PSO-TabNet-based axial spondyloarthritis (axSpA) auxiliary diagnosis and analysis system includes:

[0057] A data collection unit, used for collecting clinical characteristic data and imaging characteristic data;

[0058] A data preprocessing unit, used to preprocess the collected clinical characteristic data and imaging characteristic data;

[0059] axSpA prediction model building unit, which is used to optimize TabNet model hyperparameters using particle swarm (PSO) algorithm to obtain PSO-TabNet model;

[0060] A prediction and analysis unit is used to perform axSpA classification prediction on clinical data based on the PSO-TabNet model and obtain auxiliary diagnosis suggestions through the system.

[0061] like Figure 2 As shown, the workflow of the axSpA prediction model building unit is as follows:

[0062] The axSpA dataset is acquired based on the data acquisition module, including clinical feature data and imaging feature data, wherein the clinical feature data includes: gender, age, course of disease, HLA-B27, erythrocyte sedimentation rate and C-reactive protein. The imaging feature data is extracted by manually interpreting the nuclear magnetic resonance images of both sides of the sacroiliac joint, including: erosion, sclerosis, joint space enhancement, ankylosis, fatty lesions and bone marrow edema.

[0063] The acquired data is preprocessed based on the data preprocessing module, which specifically includes the following contents:

[0064] Missing value cleaning: determine the missing value range, calculate the missing value ratio for each field, and formulate strategies according to the missing value ratio and field importance. The following cleaning methods can be used: remove unnecessary fields and delete some meaningless fields, such as the patient's physical examination number, etc.;

[0065] Missing value filling: Different filling methods are used for missing values ​​of different features, such as using doctors’ experience to fill missing values, or using special values, medians, hot cards, etc.

[0066] Re-collection: Because some features are very important but the missing ratio is too high, it is necessary to contact the outpatient department and re-collect the data.

[0067] Logical error cleaning: Delete some data that can be directly found by simple logical reasoning to prevent the analysis results from being biased. It mainly includes deduplication, removal of unreasonable values, correction of contradictory content, etc.

[0068] Non-required data cleaning: that is, deleting unnecessary fields.

[0069] Outlier processing: Z-score standardization method is used for normalization to identify outliers, and then the following processing methods are used: outliers are removed, replaced with other values ​​(such as mean or median), and filled using interpolation.

[0070] The PSO algorithm is used to optimize the TabNet model hyperparameters and obtain the optimal hyperparameter set, which includes the following:

[0071] 1) Initialize the particle swarm and initialize the position and velocity of each particle. Specifically, in the D-dimensional search space, PSO initializes a population of M particles, and the position of the i-th particle can be expressed as The velocity of the i-th particle is Where i = 1, 2, 3, ..., M; D is the total number of hyperparameters that need to be optimized in the TabNet model;

[0072] 2) Initialize the parameters in the PSO algorithm: acceleration factors c1, c2; inertia weight ω; maximum number of iterations; maximum and minimum speed values ​​V max ,V min ;

[0073] 3) Determine the hyperparameters that need to be optimized in the TabNet model, including: the width of the decision prediction layer (n_d) and the width of the attention embedding for each mask (n_a) [4-20]; the number of steps in the architecture (n_steps) [1-3]; the feature reuse coefficient in the mask (gamma) [1.0-1.4]; the number of shared GLUs in each step (n_shared) [1-3]; the sparsity loss coefficient (lambda_sparse) [1E-6–1E-3]; the maximum number of training epochs (max_epochs) [256-512]; the number of instances per batch (batch_size) [8, 16, 32]; the mini-batch for BN (virtual_batch_size) [4, 8, 16].

[0074] 4) Assign the initial positions of all randomly generated particles to the hyperparameters to be optimized of the TabNet model;

[0075] 5) Input the data set into the TabNet model for training and record the prediction results;

[0076] 6) Use the prediction results to calculate the fitness value of each particle through formula (1) i , get the particle with the best fitness, that is, the particle with the best fitness;

[0077]

[0078] Where N is the number of samples in the test set, y i represents the actual value in the test set, Represents the predicted value of the corresponding sample.

[0079] 7) Obtain the individual extreme value pbest of the particle with the best fitness value among the particles with the best fitness value i and the group extreme value gbest i , and update the speed and position of the particle with the optimal fitness value according to formulas (2) and (3) to obtain the updated position;

[0080] v i (t+1)=ω×v i (t)+c1×r1×(pbest i -x i (t+c2×r2)×(gbest i -x i(t)) (2)

[0081] x i (t+1)=x i (t)+v i (t) (3)

[0082] Among them, v i (t+1) and v i (t) are the velocities of particle i at iterations t+1 and t, respectively. ω is the inertia factor, which is used to control the velocity of the particle. r1 and r2 are random numbers between (0,1). c1 and c2 are learning factors, usually c1=c2=2; the learning factors c1 and c2 control how the particle updates its position. Specifically, c1 is the individual learning factor, which determines the force with which the particle moves to its own optimal position; and c2 is the social learning factor, which controls the force with which the particle moves to the optimal position of all particles. The balance of these two factors helps the particle find the optimal solution. x i (t+1) and x i (t) are the positions of particle i at iterations t+1 and t, respectively.

[0083] 8) Assign the updated position to the selected hyperparameter to be optimized in the TabNet model;

[0084] 9) Repeat steps 6)-9) to determine whether the maximum number of iterations has been reached. If so, count the optimal particles in all iterations and proceed to the next step. If not, return to step 6);

[0085] 10) Apply the optimal solution of particles to the hyperparameter setting of the TabNet model.

[0086] The PSO-TabNet-based axial spondyloarthritis (axSpA) auxiliary diagnosis and analysis system of the present invention uses the trained PSO-TabNet model to perform axSpA classification prediction on clinical data, comprising the following steps:

[0087] Step 1, collect data, including clinical characteristic data and imaging characteristic data, among which the clinical characteristic data include: gender, age, course of disease, HLA-B27, erythrocyte sedimentation rate and C-reactive protein; and the imaging characteristic data need to be extracted by manually interpreting the nuclear magnetic resonance images of both sides of the sacroiliac joint, which include: erosion, sclerosis, joint space enhancement, ankylosis, fatty lesions and bone marrow edema;

[0088] Step 2: Data preprocessing, including the following contents:

[0089] Missing value cleaning: determine the missing value range, calculate the missing value ratio for each field, and formulate strategies according to the missing value ratio and field importance. For example, the following cleaning methods can be used: remove unnecessary fields and delete some meaningless fields, such as the patient's physical examination number, etc.

[0090] Missing value filling: Different filling methods are used for missing values ​​of different features, such as using doctors’ experience to fill missing values, or using special values, medians, hot cards, etc.

[0091] Re-collection: Because some features are very important but the missing ratio is too high, it is necessary to contact the outpatient department and re-collect the data.

[0092] Logical error cleaning: Delete some data that can be directly found by simple logical reasoning to prevent the analysis results from being biased. It mainly includes deduplication, removal of unreasonable values, correction of contradictory content, etc.

[0093] Non-required data cleaning: that is, deleting unnecessary fields.

[0094] Outlier processing: Z-score standardization method is used for normalization to identify outliers, and then the following processing methods are used: outliers are removed, replaced with other values ​​(such as mean or median), and filled using interpolation.

[0095] Step 3: Input the processed data into the final axSpA prediction model, including the following:

[0096] (1) Standardize and initialize the preprocessed data X. The BN layer processes the data matrix X into a matrix f as the input of each decision step, as shown in formula (4):

[0097] f=BN(X) (4)

[0098] in, B represents the batch size and D represents the feature dimension.

[0099] (2) Let i = i + 1, where i represents the decision step currently being executed, and input a[i-1] into the Attention transformer module to learn and obtain the Mask matrix M[i]. The Mask module achieves the purpose of feature selection, so that the model focuses on the data that contributes mainly to the prediction of axial spondyloarthritis during learning, thereby improving the learning efficiency of the model. The Attention transformer module implements feature selection of the current decision step through the Mask matrix in formula (5).

[0100] M[i]=sparse max(P[i-1]*hi (a[i-1])) (5)

[0101] Where a[i-1] is the input information of the current decision step obtained by the Split module in step (i-1), and h i (·) represents the fully connected FC layer and BN layer operation, which is used to realize the linear combination of features, thereby extracting higher-dimensional and abstract features. i (a[i-1]) is multiplied by P[i-1], and then the required M[i-1] is generated through the sparse maximum algorithm. M[i] is multiplied by the feature matrix to achieve feature selection in the current decision step. P[i] is the Prior scales item, which indicates the degree of use of features in the previous step. If a feature has been used many times in the previous step, it should not be selected by the model again. Therefore, the model uses this Prior scales item to reduce the weight of such features. P[i] is updated by formula (6).

[0102]

[0103] Among them, γ is a relaxation parameter. When γ = 1, a feature is forced to be used only in one decision step. In addition, the sparsemax algorithm is used to assign weights to each feature of each sample to ensure that the sum of the weights of all features in the sample is equal to 1, that is, Where D represents the dimension of the feature.

[0104] (3) The feature matrix f and M[i] are passed through the Mask module for sparse selection of the most significant features. This makes the model parameter efficient because the learning capacity of the decision step is not wasted on irrelevant features. The feature selection in the current decision step is achieved by multiplying M[i] with the feature matrix f to obtain the important features M[i]·f.

[0105] (4) The important feature M[i]·f is input into the Feature transformer module for processing to obtain the feature y[i], as shown in formula (7):

[0106] y[i]=f i (M[i]·f) (7)

[0107] Among them, f i (·) is the Feature transformer module operation, and f is the feature matrix.

[0108] (5) The processed feature y[i] is placed in the Split module for division. The processed feature y[i] is divided into two parts by the Split module; one part is used as the output d[i] of the current decision step, and the other part is used as the input information of the next decision step a[i]. N d is the number of features in the total decision output, N a is the number of features input to the Attention transformer module for the next step.

[0109] (6) Determine whether i is less than N steps , if it is less than, return to (2) and continue execution.

[0110] (7) TabNet borrows the idea of ​​tree model aggregation and aggregates the output vectors d[i] of all decision steps into d out ,in Finally, an FC layer is mapped to the final output, that is, the prediction result is obtained.

[0111] Step 4: Analyze the prediction results and obtain auxiliary diagnosis suggestions through the system.

[0112] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A PSO-TabNet-based auxiliary diagnosis and analysis system for axial spondyloarthritis, characterized by: include: A data collection unit, used for collecting clinical characteristic data and imaging characteristic data; A data preprocessing unit, used to preprocess the collected clinical characteristic data and imaging characteristic data; axSpA prediction model building unit, which is used to optimize TabNet model hyperparameters using particle swarm optimization to obtain the PSO-TabNet model; A prediction and analysis unit is used to perform axSpA classification prediction on clinical data based on the PSO-TabNet model and obtain auxiliary diagnosis suggestions through the system.

2. The PSO-TabNet-based axial spondyloarthritis auxiliary diagnosis and analysis system according to claim 1, characterized in that: The workflow of the axSpA prediction model building unit is as follows: Step 1, obtain the axSpA dataset, which includes clinical feature data and imaging feature data; Step 2: Preprocess the data, including data cleaning, processing of missing values ​​and outliers; Step 3. Use the PSO algorithm to optimize the TabNet model hyperparameters and obtain the optimal hyperparameter set.

3. The PSO-TabNet-based axial spondyloarthritis auxiliary diagnosis and analysis system according to claim 1, characterized in that: The clinical characteristic data include gender, age, disease course, HLA-B27, erythrocyte sedimentation rate and C-reactive protein; the imaging characteristic data are extracted by manually interpreting the magnetic resonance images of both sides of the sacroiliac joints, including erosion, sclerosis, joint space enhancement, ankylosis, fatty lesions and bone marrow edema.

4. The PSO-TabNet-based axial spondyloarthritis auxiliary diagnosis and analysis system according to claim 1, characterized in that: The step 2 specifically includes the following contents: Missing value cleaning: determine the missing value range, calculate the missing value ratio for each field, and formulate strategies based on the missing value ratio and field importance; Missing value filling: Different filling methods are used for missing values ​​of different features; Re-data collection: When certain features are very important but the missing ratio is too high, it is necessary to contact the outpatient department and re-data collection; Logical error cleaning: Delete some data where problems can be directly found using simple logical reasoning; Non-requirement data cleaning: delete unnecessary fields; Outlier processing: Z-score standardization method is used for normalization to identify outliers, and then the following processing methods are used: outliers are removed, replaced with other values, and filled using interpolation.

5. The PSO-TabNet-based axial spondyloarthritis auxiliary diagnosis and analysis system according to claim 1, characterized in that: The step Step 3 specifically includes the following steps: 1) Initialize the particle swarm and initialize the position and velocity of each particle; 2) Initialize the parameters in the PSO algorithm: acceleration factors c1, c2; inertia weight ω; maximum number of iterations; maximum and minimum speed values ​​V max ,V min ; 3) Determine the hyperparameters that need to be optimized in the TabNet model; 4) Assign the initial positions of all randomly generated particles to the hyperparameters to be optimized of the TabNet model; 5) Input the data set into the TabNet model for training and record the prediction results; 6) Use the prediction results to calculate the fitness value of each particle through formula (1) i , get the particle with the best fitness, that is, the particle with the best fitness; Where N is the number of samples in the test set, y i represents the actual value in the test set, Represents the predicted value of the corresponding sample; 7) Obtain the individual extreme value pbest of the particle with the best fitness value among the particles with the best fitness value i and the group extreme value gbest i , and update the speed and position of the particle with the optimal fitness value according to formulas (2) and (3) to obtain the updated position; v i (t+1)=ω×v i (t)+c1×r1×(pbest i -x i (t+c2×r2)×(gbest i -x i (t)) (2) x i (t+1)=x i (t)+v i (t) (3) Where v i (t+1) and v i (t) are the velocities of particle i at iterations t+1 and t respectively; ω is the inertia factor used to control the velocity of the particle; r1 and r2 are random numbers between (0,1); c1 and c2 are learning factors, x i (t+1) and x i (t) are the positions of particle i at iterations t+1 and t, respectively; 8) Assign the updated position to the selected hyperparameter to be optimized in the TabNet model; 9) Repeat steps 6)-9) to determine whether the maximum number of iterations has been reached. If so, count the optimal particles in all iterations and proceed to the next step. If not, return to step 6); 10) Apply the optimal solution of particles to the hyperparameter setting of the TabNet model.

6. The PSO-TabNet-based axial spondyloarthritis auxiliary diagnosis and analysis system according to claim 5, characterized in that: The hyperparameters that need to be optimized in step 3) include: the width of the decision prediction layer and the width of the attention embedding of each mask; the number of steps in the architecture; the feature reuse coefficient in the mask; the number of shared GLUs in each step; the sparsity loss coefficient; the maximum training; the number of instances per batch; and the mini-batch for BN.

7. The PSO-TabNet-based axial spondyloarthritis auxiliary diagnosis and analysis system according to claim 1, characterized in that: When applied, the following steps are included: Step 1, data collection; Step 2, data preprocessing: including data cleaning, processing of missing values ​​and outliers; Step 3, input the processed data into the PSO-TabNet model; Step 4: Get the prediction results for doctors to assist in diagnosis.

Images