A clustering method for incomplete multi-views in medical examinations

Abstract

The application provides a clustering method for incomplete multi-views in medical examinations, which comprises the following steps: firstly, acquiring incomplete multi-views in medical examinations, constructing an extended affinity matrix, then obtaining a structure matrix, a complete sample matrix and a reconstruction mapping matrix, and obtaining an inference loss function according to the three matrices; applying a constraint to the structure matrix to obtain a confidence block diagonal regularization term; obtaining a completed structure matrix according to the structure matrix, and obtaining a completion loss function according to the two matrices; obtaining a relationship loss function according to the completed structure matrix; obtaining a target loss function according to the inference loss function, the confidence block diagonal regularization term, the completion loss function and the relationship loss function, iteratively optimizing part of the matrix in the target loss function until convergence, and finally obtaining the final clustering labels of all incomplete multi-views in medical examinations by using a clustering algorithm to complete clustering. The application has superior structure recovery capability and more robust clustering performance on multi-views with a higher missing rate in medical examinations.

Description

Technical Field

[0001] This invention relates to the field of biomedical and machine learning integration, specifically to a clustering method for incomplete multi-view medical examinations. Background Technology

[0002] In the field of medical examination and clinical diagnosis, to achieve accurate disease assessment, lesion localization, and pathological classification, the same subject (or sample) is typically represented by different dimensions of features through multiple medical examination methods, forming typical multi-view medical data. For example, for lung lesion samples, computed tomography (CT) scans can obtain cross-sectional anatomical views of the lesions, magnetic resonance imaging (MRI) can obtain soft tissue metabolic characteristics views, and ultrasound examinations can obtain real-time morphodynamic views, while simultaneously combining blood biochemical tests to obtain molecular characteristic views such as tumor markers. Similarly, for the diagnosis of brain diseases, electroencephalography (EEG) can be used to simultaneously acquire electrophysiological views, functional magnetic resonance imaging (fMRI) can acquire brain functional connectivity views, and pathological slide tissue morphology views. Multi-view medical data can complementarily characterize the health status of the subject from multiple levels, including anatomical, physiological, molecular, and functional perspectives. Cluster analysis of this data is an important technical means for disease subtyping, lesion benign / malignant differentiation, and clinical prognosis grouping, and has irreplaceable application value in precision medicine, assisted diagnosis, and large-scale clinical sample screening.

[0003] However, in actual clinical examination scenarios, due to various constraints, data for some samples in specific views is often completely missing, resulting in incomplete multi-view medical data. Specifically, the reasons for data loss mainly include: first, individual tolerance limitations of the examinee, such as some patients being unable to undergo MRI due to metal implants, or elderly or critically ill patients being unable to tolerate prolonged CT scans, leading to missing data for corresponding views; second, constraints of examination costs and medical resources, such as primary healthcare institutions lacking high-end imaging equipment, or some patients being unable to afford full-dimensional examinations due to economic reasons, only able to complete basic procedures; third, losses during sample collection and storage, such as failed staining of pathological sections, hemolysis of biochemical test samples, and loss during image data transmission; and fourth, privacy compliance and clinical ethics restrictions, as some sensitive examinations (such as gene testing and imaging of special sites) require explicit authorization from the examinee, and data for corresponding views cannot be collected without authorization. Traditional multi-view clustering methods typically rely on the completeness of all view data, making them difficult to directly apply to such incomplete multi-view medical data, severely restricting their practical application in clinical medical examinations.

[0004] Existing clustering methods for incomplete multi-view data still have significant limitations in medical applications, and can be mainly divided into two categories: The first category is matrix completion-based methods. These methods first fill in the missing medical view data using interpolation, fitting, or statistical inference, and then perform cluster analysis on the completed data. However, this type of method completely separates data completion from clustering tasks. During the completion process, it often ignores the physiological correlations and clinical rationality of medical data, easily generating false features that do not conform to clinical reality, leading to biased clustering results and failing to meet the accuracy requirements of disease diagnosis. The second category is subspace learning-based methods. These methods directly learn common subspaces from incomplete multi-view medical data, achieving cross-view information fusion and clustering through subspace mapping. However, this type of method mostly focuses on global feature alignment, often ignoring the local geometric structure within the medical view, and failing to effectively maintain structural consistency between different views, resulting in clustering results that cannot accurately reflect the clinical feature correlations of the disease.

[0005] Considering the unique nature of medical examinations, the core shortcomings of existing technologies can be further refined as follows: 1. The missing data inference process fails to fully utilize the local geometric structure and physiological symmetry of known medical samples. For example, lesion features in symmetrical organs such as the lungs and breasts have significant symmetrical correlations. Existing methods do not effectively leverage this symmetry to improve the inference accuracy of missing views, resulting in discrepancies between the completed lesion features and the actual physiological structure. 2. During data completion and clustering, the methods fail to explicitly enforce a consistent clustering structure across all medical views, i.e., a block diagonal structure. For instance, the same lesion should be classified into the same category in CT and MRI views. However, existing methods are prone to conflicting classification results for the same lesion in different views, failing to meet the consistency requirements of clinical diagnosis. 3. The information complementarity mechanism between views is not flexible enough. It fails to adaptively integrate information based on the confidence level of medical examinations. Different medical examinations have different diagnostic values. For example, MRI has a higher confidence level for identifying soft tissue lesions than CT, and biochemical indicators have a higher confidence level for characterizing inflammation than imaging. Existing methods often use fixed-weight fusion, which cannot adapt to these confidence level differences and is easily affected by low-confidence views. Summary of the Invention

[0006] To overcome the problems of the prior art, such as the inability to infer and complete a full multi-view, the inconsistency of the multi-view structure after completion, and the lack of flexibility in the multi-view complementation mechanism, this invention provides a clustering method for incomplete multi-views in medical examinations.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0008] Step 1: Obtain incomplete multi-views from medical examinations. The multi-views are matrices composed of multi-dimensional data from medical examinations. Construct an extended affinity matrix for the incomplete multi-views. Step 2: Obtain the complete sample matrix and the reconstructed mapping matrix based on the extended affinity matrix. Normalize the extended affinity matrix to obtain the structure matrix. Obtain the inference loss function based on the structure matrix, the complete sample matrix, and the reconstructed mapping matrix. Step 3: Stack the structure matrices of all views into a three-dimensional tensor and apply nuclear norm constraints to obtain a low-rank constraint matrix. Apply norm constraints to each structure matrix to obtain a sparse constraint matrix. Obtain the confidence block diagonal regularization term based on the low-rank constraint matrix and the sparse constraint matrix. Step 4: Obtain the completed structure matrix based on the structure matrix and the preset weight matrix, and obtain the completed loss function based on the completed structure matrix and the structure matrix; Step 5: Construct a Laplacian matrix for each view based on the completed structure matrix, obtain a consensus representation matrix based on the Laplacian matrix, and obtain a relation loss function based on the consensus representation matrix and the Laplacian matrix; Step 6: Obtain the target loss function based on the inference loss function, the confidence block diagonal regularization term, the completion loss function, and the relational loss function. Using the target loss function, iteratively optimize the complete sample matrix, reconstruct the mapping matrix, complete the structure matrix, structure matrix, weight matrix, and consensus representation matrix using the alternating direction multiplier method until the target loss function converges. Step 7: Normalize the consensus representation matrix obtained after the objective loss function converges, and use a clustering algorithm to obtain the final cluster labels for all incomplete multi-views in the medical examination, thus completing the clustering of incomplete multi-views in the medical examination.

[0009] Furthermore, the construction of the extended affinity matrix for the incomplete multi-view includes constructing the affinity matrix and projection matrix for each view based on the incomplete multi-view, and extending the affinity matrix and projection matrix to the complete sample space to obtain the extended affinity matrix, including: Affinity Matrix : ,

[0010] In the formula, For bandwidth parameters, The first incomplete multi-view k row element, The first incomplete multi-view j Column elements, and set the diagonal elements to 0 or 1. The total number of samples, For the first The number of observable samples in each view; Projection matrix :

[0011] In the formula, It is a row selection matrix, if the global first row is selected... k Each sample corresponds to a view The first in j If there are 100 available samples, then Otherwise, it is 0; Extended Affinity Matrix :

[0012] In the formula, T This indicates transpose.

[0013] Further, the step of obtaining the complete sample matrix and the reconstructed mapping matrix based on the extended affinity matrix, normalizing the extended affinity matrix to obtain the structure matrix, and obtaining the inference loss function based on the structure matrix, the complete sample matrix, and the reconstructed mapping matrix includes: Structure matrix :

[0014] In the formula, This is represented as normalizing the extended affinity matrix; Inference loss function:

[0015]

[0016] In the formula, To infer the complete sample matrix, To reconstruct the mapping matrix, For projection operators, This is a matrix representing incomplete multi-view data, where V is the total number of views. It represents the feature dimension, and F represents the norm.

[0017] Further, the step of stacking the structure matrices of all views into a three-dimensional tensor and applying nuclear norm constraints to obtain a low-rank constraint matrix, and applying norm constraints to each structure matrix to obtain a sparse constraint matrix, includes: 3D tensor : =

[0018] Low-rank constraint matrix :

[0019] In the formula, The matrix nuclear norm; Sparse constraint matrix :

[0020] In the formula, The first equilibrium parameter, It represents the matrix norm.

[0021] Furthermore, the confidence block diagonal regularization term is obtained based on the low-rank constraint matrix and the sparse constraint matrix. The formula is: .

[0022] Further, obtaining the completed structure matrix based on the structure matrix and the preset weight matrix includes: Complete the structure matrix :

[0023] In the formula, Represents the weight matrix The Middle The view for the first The contribution weight of each view. Indicates the first A structure matrix, a weight matrix The following constraints must be met:

[0024] In the formula, Indicates the first The view for the first The contribution weight of each view.

[0025] Further, obtaining the completion loss function based on the completion structure matrix and the structure matrix includes: The formula for the completion loss function is:

[0026] In the formula, This indicates element-wise multiplication. The fusion strength coefficient, This is represented as a missing indicator mask. If and only if the sample k and samples j In the v Observational data exists in each view.

[0027] Further, the step of constructing a Laplacian matrix for each view based on the completed structure matrix, obtaining a consensus representation matrix based on the Laplacian matrix, and obtaining a relation loss function based on the consensus representation matrix and the Laplacian matrix includes: Laplace matrix :

[0028] In the formula, For a degree matrix, its diagonal elements , Represented as the th in the complete structure matrix k OK j Column elements, Represented as the th in the complete structure matrix j OK k Column elements, k , j ; Consensus Representation Matrix for The matrix formed by the eigenvectors corresponding to the first c smallest eigenvalues, where c is the preset number of clusters; The formula for the relationship loss function is:

[0029] In the formula, The second equilibrium parameter, It is the identity matrix. ( ) represents the trace of a matrix.

[0030] Furthermore, the target loss function is obtained based on the inference loss function, the confidence block diagonal regularization term, the completion loss function, and the relational loss function. The formula for the target loss function is:

[0031] , , , , , , ; In the formula, To infer the complete sample matrix, To reconstruct the mapping matrix, For structure matrix, For projection operators, This is a matrix representing incomplete multi-view data, where V is the total number of views. It represents the feature dimension, and F represents the norm. For the confidence block diagonal regularization term, This indicates element-wise multiplication. The fusion strength coefficient, This is represented as a missing indicator mask. If and only if the sample k and samples j In the v Observational data exists in each view. To complete the structure matrix, Represented as the diagonal elements of the structure matrix. For Laplace matrix, The consensus representation matrix, It is the identity matrix. The second equilibrium parameter, ( ) represents the trace of the matrix. For the set of all optimization variables, For view to adapt weights, The sharpening factor is used to enhance the weight of important views. The regularization coefficient is . Represents the weight matrix The Middle The view for the first The contribution weight of each view. Indicates the first A structure matrix, This is the weight matrix. Indicates the first The view for the first The contribution weight of each view, where T is the transpose.

[0032] Furthermore, the step of iteratively optimizing the complete sample matrix, reconstructed mapping matrix, completed structure matrix, structure matrix, weight matrix, and consensus representation matrix using the alternating direction multiplier method based on the target loss function until the target loss function converges, wherein the convergence condition includes: convergence occurs when any of the following conditions are met: The relative change in the target loss function is less than the threshold. , ; Reaching the preset maximum number of iterations T ; The changes in all optimization variables are less than the threshold. .

[0033] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This method precisely addresses the core pain points of incomplete multi-view clustering in medical examinations, effectively compensating for the shortcomings of existing technologies in structure preservation, information fusion, and missing data processing. It significantly improves clustering performance and the application value in medical examinations. Under simulated clinical scenarios with 30% and 50% sample missing rates, it demonstrates excellent clustering accuracy for incomplete multi-view data in medical examinations through inference and completion loss functions, exhibiting stronger feature capture capabilities and accurately adapting to complex scenarios with partially missing multi-view data in medical examinations. Simultaneously, it achieves intelligent adaptive information fusion of multiple views. By constructing a dynamic fusion mechanism through confidence block diagonal regularization and relational loss functions, it automatically fuses the confidence levels of each multi-view without tedious manual parameter tuning, flexibly integrating features from various dimensions of the multi-view, maintaining structural consistency between the multi-views, and solving the problems of information redundancy or weakening of key features caused by fixed-weight fusion in traditional methods. Attached Figure Description

[0034] Figure 1 This is a flowchart of a clustering method for incomplete multi-view medical examinations.

[0035] Figure 2 This is a structural diagram of the algorithm model for a clustering method for incomplete multi-view medical examinations. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0037] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0038] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] In a first embodiment of the present invention, a clustering method for incomplete multi-view medical examinations is provided, combined with... Figure 1 and Figure 2 Explanation: Step 1: Obtain incomplete multi-views from medical examinations. The multi-views are matrices composed of multi-dimensional data from medical examinations. Construct an extended affinity matrix for the incomplete multi-views. Furthermore, the construction of the extended affinity matrix for the incomplete multi-view includes constructing the affinity matrix and projection matrix for each view based on the incomplete multi-view, and extending the affinity matrix and projection matrix to the complete sample space to obtain the extended affinity matrix, including: Affinity Matrix : ,

[0040] In the formula, For bandwidth parameters, The first incomplete multi-view k row element, The first incomplete multi-view j Column elements, and set the diagonal elements to 0 or 1. The total number of samples, For the first The number of observable samples in each view; Projection matrix :

[0041] In the formula, It is a row selection matrix, if the global first row is selected... k Each sample corresponds to a view The first in j If there are 100 available samples, then Otherwise, it is 0; Extended Affinity Matrix :

[0042] In the formula, T This indicates transpose.

[0043] Step 2: Obtain the complete sample matrix and the reconstructed mapping matrix based on the extended affinity matrix. Normalize the extended affinity matrix to obtain the structure matrix. Obtain the inference loss function based on the structure matrix, the complete sample matrix, and the reconstructed mapping matrix. Further, the step of obtaining the complete sample matrix and the reconstructed mapping matrix based on the extended affinity matrix, normalizing the extended affinity matrix to obtain the structure matrix, and obtaining the inference loss function based on the structure matrix, the complete sample matrix, and the reconstructed mapping matrix includes: Structure matrix :

[0044] In the formula, This is represented as normalizing the extended affinity matrix; Inference loss function:

[0045]

[0046] In the formula, To infer the complete sample matrix, To reconstruct the mapping matrix, For projection operators, This is a matrix representing incomplete multi-view data, where V is the total number of views. It represents the feature dimension, and F represents the norm.

[0047] Step 3: Stack the structure matrices of all views into a three-dimensional tensor and apply nuclear norm constraints to obtain a low-rank constraint matrix. Apply norm constraints to each structure matrix to obtain a sparse constraint matrix. Obtain the confidence block diagonal regularization term based on the low-rank constraint matrix and the sparse constraint matrix. Further, the step of stacking the structure matrices of all views into a three-dimensional tensor and applying nuclear norm constraints to obtain a low-rank constraint matrix, and applying norm constraints to each structure matrix to obtain a sparse constraint matrix, includes: 3D tensor : =

[0048] Low-rank constraint matrix :

[0049] In the formula, The matrix nuclear norm; Sparse constraint matrix :

[0050] In the formula, The first equilibrium parameter, It represents the matrix norm.

[0051] Furthermore, the confidence block diagonal regularization term is obtained based on the low-rank constraint matrix and the sparse constraint matrix. The formula is: .

[0052] Step 4: Obtain the completed structure matrix based on the structure matrix and the preset weight matrix, and obtain the completed loss function based on the completed structure matrix and the structure matrix; Further, obtaining the completed structure matrix based on the structure matrix and the preset weight matrix includes: Complete the structure matrix :

[0053] In the formula, Represents the weight matrix The Middle The view for the first The contribution weight of each view. Indicates the first A structure matrix, a weight matrix The following constraints must be met:

[0054] In the formula, Indicates the first The view for the first The contribution weight of each view.

[0055] Further, obtaining the completion loss function based on the completion structure matrix and the structure matrix includes: The formula for the completion loss function is:

[0056] In the formula, This indicates element-wise multiplication. The fusion strength coefficient, This is represented as a missing indicator mask. If and only if the sample k and samples j In the v Observational data exists in each view.

[0057] Step 5: Construct a Laplacian matrix for each view based on the completed structure matrix, obtain a consensus representation matrix based on the Laplacian matrix, and obtain a relation loss function based on the consensus representation matrix and the Laplacian matrix; Further, the step of constructing a Laplacian matrix for each view based on the completed structure matrix, obtaining a consensus representation matrix based on the Laplacian matrix, and obtaining a relation loss function based on the consensus representation matrix and the Laplacian matrix includes: Laplace matrix :

[0058] In the formula, For a degree matrix, its diagonal elements , Represented as the th in the complete structure matrix k OK j Column elements, Represented as the th in the complete structure matrix j OK k Column elements, k , j ; Consensus Representation Matrix for The matrix formed by the eigenvectors corresponding to the first c smallest eigenvalues, where c is the preset number of clusters; The formula for the relationship loss function is:

[0059] In the formula, The second equilibrium parameter, It is the identity matrix. ( ) represents the trace of a matrix.

[0060] Step 6: Obtain the target loss function based on the inference loss function, the confidence block diagonal regularization term, the completion loss function, and the relational loss function. Using the target loss function, iteratively optimize the complete sample matrix, reconstruct the mapping matrix, complete the structure matrix, structure matrix, weight matrix, and consensus representation matrix using the alternating direction multiplier method until the target loss function converges. Furthermore, the target loss function is obtained based on the inference loss function, the confidence block diagonal regularization term, the completion loss function, and the relational loss function. The formula for the target loss function is:

[0061] , , , , , , ; In the formula, To infer the complete sample matrix, To reconstruct the mapping matrix, For structure matrix, For projection operators, This is a matrix representing incomplete multi-view data, where V is the total number of views. It represents the feature dimension, and F represents the norm. For the confidence block diagonal regularization term, This indicates element-wise multiplication. The fusion strength coefficient, This is represented as a missing indicator mask. If and only if the sample k and samples j In the v Observational data exists in each view. To complete the structure matrix, Represented as the diagonal elements of the structure matrix. For Laplace matrix, The consensus representation matrix, It is the identity matrix. The second equilibrium parameter, ( ) represents the trace of the matrix. For the set of all optimization variables, For view to adapt weights, The sharpening factor is used to enhance the weight of important views. The regularization coefficient is . Represents the weight matrix The Middle The view for the first The contribution weight of each view. Indicates the first A structure matrix, This is the weight matrix. Indicates the first The view for the first The contribution weight of each view, where T is the transpose.

[0062] Furthermore, the step of iteratively optimizing the complete sample matrix, reconstructed mapping matrix, completed structure matrix, structure matrix, weight matrix, and consensus representation matrix using the alternating direction multiplier method based on the target loss function until the target loss function converges, wherein the convergence condition includes: convergence occurs when any of the following conditions are met: The relative change in the target loss function is less than the threshold. , ; Reaching the preset maximum number of iterations T ; The changes in all optimization variables are less than the threshold. .

[0063] Preferably, a projection matrix is ​​also constructed for each incomplete multi-view. This is used to map observed samples within a view to the complete sample space. For the first... The view, if the first view in the complete sample space The sample corresponds to the first one in this view. For each observation sample, then Otherwise, it is 0. This projection matrix accurately records the location information of the observed sample in the complete sample set in each view.

[0064] The iterative optimization based on the alternating direction multiplier method is as follows: We introduce auxiliary variables, Lagrange multipliers, and penalty parameters, and fix other variables in each iteration while updating each variable in turn.

[0065] First, the completed structure matrix is ​​solved element-wise using a closed-form iterative solution; then, the weight matrix is ​​solved using the accelerated projection gradient method to solve the constrained quadratic programming problem, as shown in the following formula:

[0066]

[0067] In the formula, For Lagrange multipliers, This is the penalty parameter.

[0068] The consensus representation matrix is ​​calculated using eigenvalue decomposition, with the following formula:

[0069] In the formula, Indicates before calculation The eigenvectors corresponding to the smallest eigenvalues.

[0070] The complete sample matrix is ​​updated using the least squares solution, as shown in the formula:

[0071] In the formula, For the first v Each view loss, .

[0072] The reconstructed mapping matrix is ​​solved using a closed-form solution, as shown in the formula:

[0073] The formula for updating the structure matrix is:

[0074] After the update, constraints must be applied, the diagonal must be set to zero, and non-negativity and symmetry must be maintained.

[0075] Finally, update the view adaptive weights. The formula is:

[0076] In the formula, For the first k Each view loss.

[0077] Step 7: Normalize the consensus representation matrix obtained after the objective loss function converges, and use the k-means clustering algorithm to obtain the final cluster labels for all incomplete multi-views in the medical examination, thus completing the clustering of incomplete multi-views in the medical examination.

[0078] A second embodiment of the present invention provides a clustering system for incomplete multi-view clustering in medical examinations, comprising a construction module, an inference module, a structure module, a completion module, a consensus module, and an optimized clustering module, wherein: The building module is used to construct extended affinity matrices and structure matrices; The inference module is used to calculate the inference loss function; The structural module is used to apply constraints to the structural matrix to obtain the confidence block diagonal regularization term; The completion module is used to construct the completion structure matrix and obtain the completion loss function; The consensus module is used to obtain the consensus representation matrix and the Laplacian matrix and construct the relation loss function; The optimized clustering module is used to combine the loss function and confidence block diagonal regularization term obtained from other modules. The target loss function is obtained by integration and iterative optimization of the target loss function is performed. After iterative optimization, a clustering algorithm is used to obtain the clustering labels of the incomplete multi-view. The other modules include an inference module, a completion module, and a consensus module.

[0079] A third embodiment of the present invention provides a clustering device for incomplete multi-view medical examinations, comprising: Memory, used to store computer programs; A processor, used to execute the computer program to implement any one of steps 1-7 of the clustering method for incomplete multiviews in medical examinations.

[0080] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A clustering method for incomplete multi-view medical examinations, characterized in that, include: Step 1: Obtain incomplete multi-views from medical examinations. The multi-views are matrices composed of multi-dimensional data from medical examinations. Construct an extended affinity matrix for the incomplete multi-views. Step 2: Obtain the complete sample matrix and the reconstructed mapping matrix based on the extended affinity matrix. Normalize the extended affinity matrix to obtain the structure matrix. Obtain the inference loss function based on the structure matrix, the complete sample matrix, and the reconstructed mapping matrix. Step 3: Stack the structure matrices of all views into a three-dimensional tensor and apply nuclear norm constraints to obtain a low-rank constraint matrix. Apply norm constraints to each structure matrix to obtain a sparse constraint matrix. Obtain the confidence block diagonal regularization term based on the low-rank constraint matrix and the sparse constraint matrix. Step 4: Obtain the completed structure matrix based on the structure matrix and the preset weight matrix, and obtain the completed loss function based on the completed structure matrix and the structure matrix; Step 5: Construct a Laplacian matrix for each view based on the completed structure matrix, obtain a consensus representation matrix based on the Laplacian matrix, and obtain a relation loss function based on the consensus representation matrix and the Laplacian matrix; Step 6: Obtain the target loss function based on the inference loss function, the confidence block diagonal regularization term, the completion loss function, and the relational loss function. Using the target loss function, iteratively optimize the complete sample matrix, reconstruct the mapping matrix, complete the structure matrix, structure matrix, weight matrix, and consensus representation matrix using the alternating direction multiplier method until the target loss function converges. Step 7: Normalize the consensus representation matrix obtained after the objective loss function converges, and use a clustering algorithm to obtain the final cluster labels for all incomplete multi-views in the medical examination, thus completing the clustering of incomplete multi-views in the medical examination.

2. The clustering method for incomplete multi-view medical examinations according to claim 1, characterized in that, The construction of the extended affinity matrix for incomplete multi-views includes constructing the affinity matrix and projection matrix for each view based on the incomplete multi-views, and extending the affinity matrix and projection matrix to the complete sample space to obtain the extended affinity matrix, including: Affinity Matrix : , In the formula, For bandwidth parameters, The first incomplete multiview k row element, The first incomplete multiview j Column elements, with diagonal elements set to 0 or 1. The total number of samples, For the first Number of observable samples in each view; Projection matrix : In the formula, It is a row selection matrix, if the global first row is selected... k Each sample corresponds to a view The first in j If there are 100 available samples, then Otherwise, it is 0; Extended Affinity Matrix : In the formula, T This indicates transpose.

3. The clustering method for incomplete multi-view medical examinations according to claim 2, characterized in that, The process of obtaining the complete sample matrix and reconstructed mapping matrix based on the extended affinity matrix, normalizing the extended affinity matrix to obtain the structure matrix, and obtaining the inference loss function based on the structure matrix, the complete sample matrix, and the reconstructed mapping matrix includes: Structure matrix : In the formula, This is represented as normalizing the extended affinity matrix; Inference loss function: In the formula, To infer the complete sample matrix, To reconstruct the mapping matrix, For projection operators, This is a matrix representing incomplete multi-view data, where V is the total number of views. It represents the feature dimension, and F represents the norm.

4. The clustering method for incomplete multi-view medical examinations according to claim 3, characterized in that, The process of stacking the structure matrices of all views into a three-dimensional tensor and applying nuclear norm constraints to obtain a low-rank constraint matrix, and applying norm constraints to each structure matrix to obtain a sparse constraint matrix, includes: 3D tensor : = Low-rank constraint matrix : In the formula, The matrix nuclear norm; Sparse constraint matrix : In the formula, The first equilibrium parameter, It represents the matrix norm.

5. The clustering method for incomplete multi-view medical examinations according to claim 4, characterized in that, The confidence block diagonal regularization term is obtained based on the low-rank constraint matrix and the sparse constraint matrix. The formula is: 。 6. The clustering method for incomplete multi-view medical examinations according to claim 3, characterized in that, The process of obtaining the completed structure matrix based on the structure matrix and the preset weight matrix includes: Complete the structure matrix : In the formula, Represents the weight matrix The Middle The view for the first The contribution weight of each view. Indicates the first A structure matrix, a weight matrix The following constraints must be met: In the formula, Indicates the first The view for the first The contribution weight of each view.

7. The clustering method for incomplete multi-view medical examinations according to claim 6, characterized in that, The step of obtaining the completion loss function based on the completed structure matrix and the structure matrix includes: The formula for the completion loss function is: In the formula, This indicates element-wise multiplication. The fusion strength coefficient, This is represented as a missing indicator mask. If and only if the sample k and samples j In the v Observational data exists in each view.

8. The clustering method for incomplete multi-view medical examinations according to claim 6, characterized in that, The process of constructing a Laplacian matrix for each view based on the completed structure matrix, obtaining a consensus representation matrix based on the Laplacian matrix, and obtaining a relation loss function based on the consensus representation matrix and the Laplacian matrix includes: Laplace matrix : In the formula, For a degree matrix, its diagonal elements , Represented as the th in the complete structure matrix k OK j Column elements, Represented as the th in the complete structure matrix j OK k Column elements, k , j ; Consensus Representation Matrix for The matrix formed by the eigenvectors corresponding to the first c smallest eigenvalues, where c is the preset number of clusters; The formula for the relationship loss function is: In the formula, The second equilibrium parameter, It is the identity matrix. ( ) represents the trace of a matrix.

9. The clustering method for incomplete multi-view medical examinations according to claim 1, characterized in that, The target loss function is obtained by considering the inference loss function, the confidence block diagonal regularization term, the completion loss function, and the relational loss function. The formula for the target loss function is as follows: ， ， ， ， ， ， ； In the formula, To infer the complete sample matrix, To reconstruct the mapping matrix, For structure matrix, For projection operators, This is a matrix representing incomplete multi-view data, where V is the total number of views. It represents the feature dimension, and F represents the norm. For the confidence block diagonal regularization term, This indicates element-wise multiplication. The fusion strength coefficient, This is represented as a missing indicator mask. If and only if the sample k and samples j In the v Observational data exists in each view. To complete the structure matrix, Represented as the diagonal elements of the structure matrix. For Laplace matrix, The consensus representation matrix, It is the identity matrix. The second equilibrium parameter, ( ) represents the trace of the matrix. For the set of all optimization variables, For view to adapt weights, The sharpening factor is used to enhance the weight of important views. The regularization coefficient is . Represents the weight matrix The Middle The view for the first The contribution weight of each view. Indicates the first A structure matrix, This is the weight matrix. Indicates the first The view for the first The contribution weight of each view, where T is the transpose.

10. The clustering method for incomplete multi-view medical examinations according to claim 9, characterized in that, The process involves iteratively optimizing the complete sample matrix, reconstructed mapping matrix, completed structure matrix, structure matrix, weight matrix, and consensus representation matrix using the alternating direction multiplier method, based on the target loss function, until the target loss function converges. The convergence condition includes: convergence occurs when any of the following conditions are met: The relative change in the target loss function is less than the threshold. , ; Reaching the preset maximum number of iterations T ; The changes in all optimization variables are less than the threshold. .

Images