A trusted analysis method for medical blockchain data based on multi-granularity sequential three-branch decision making

Abstract

The present invention discloses a medical blockchain data trust analysis method based on multi-granularity sequential three-branch decision-making, which is characterized by comprising the following five modules: 1) an on-chain medical data module, 2) a medical data category classification module, 3) an on-chain medical data attribute simplification module, 4) a medical data module to be on-chain, and 5) a multi-granularity medical data trust partitioning module; the present invention abstracts medical data into a hypergraph model, obtains the medical data core attribute set in the form of solving its vertex cover, and reduces or eliminates redundant attributes in the medical data; combines the vertex cover core attribute set with the rough set digital features to propose a multi-granularity hierarchical strategy suitable for medical data, thereby improving the overall trust partitioning efficiency of the model.

Description

Technical Field

[0001] The present invention relates to a medical blockchain data trustworthy analysis method based on multi-granularity sequential three-branch decision-making. Background Art

[0002] The process of medical informatization has entered the era of big data. The generation and circulation of massive data has become the norm. Many hospitals have entered the stage of data utilization from the stage of information system construction. However, when many medical institutions use medical data for clinical and decision-making research, they are faced with the problems of poor data quality and low data credibility, which makes most medical institutions unable to play the value of their medical data. The main reasons for the low credibility of medical data are: technical limitations or malicious attacks cause the data collected by the system to be inconsistent with the actual situation, the data before and after data transmission are inconsistent, and the data is illegally tampered with and deleted during storage. Therefore, many scholars have conducted extensive research on the problem of poor quality of medical data from different angles such as medical data collection, transmission and storage.

[0003] As an important granular computing and knowledge discovery research method, the three-branch decision theory is very suitable for screening credible medical data to be uploaded to the chain, ensuring the credibility of medical blockchain data from the source. Summary of the invention

[0004] In order to ensure the credibility of the source data to be put on the chain of the medical blockchain, the present invention proposes a medical blockchain data trust analysis method based on multi-granularity sequential three-branch decision. In order to eliminate the redundant attributes in the original medical data, a medical data attribute simplification method based on a hypergraph model is first studied, and the on-chain medical data (trusted by default) is abstracted into a hypergraph model. Then, from the perspective of the hypergraph, a vertex cover-based reduction algorithm ARHVC is proposed to extract its trusted core attribute set from a certain type of medical data on the chain, and improve the efficiency of the trusted division of the three-branch decision of the medical data to be put on the chain from the attribute level of the data; then, the present invention aims at the granular decision conflict problem in the traditional multi-granularity three-branch decision, and studies a medical data trust division method based on multi-granularity sequential three-branch decision. This method proposes a multi-granularity hierarchical strategy for medical data, which further improves the accuracy and efficiency of the trusted decision of the medical data to be put on the chain at each granular layer of the sequential three-branch decision from the perspective of data granularity.

[0005] The method described in the present invention is divided into two parts, namely: 1. Attribute simplification of medical data based on a hypergraph model. There are many redundant attributes in the original medical data, which greatly increases the cost of credible partitioning of medical data. In order to improve the classification ability of attribute equivalence relations in medical data, the present invention abstracts the medical data into a hypergraph model, obtains the medical data core attribute set in the form of vertex cover, and reduces or eliminates redundant attributes in the medical data; 2. The traditional multi-granularity three-branch decision lacks an effective granularity selection strategy, resulting in frequent decision conflicts between different granular layers. In order to reduce decision conflicts between granular layers, the present invention combines the vertex cover core attribute set with the rough set digital features to propose a multi-granularity hierarchical strategy suitable for medical data, which improves the overall credible partitioning efficiency of the model.

[0006] 1. This paper proposes a medical blockchain data trust analysis method based on multi-granularity sequential three-branch decision-making, which includes five modules:

[0007] 1) The on-chain medical data module is a medical blockchain data storage system;

[0008] 2) The medical data classification module is a decision-making information system that classifies medical data into categories according to the ICD International Classification of Diseases.

[0009] 3) The attribute simplification module of on-chain medical data extracts trusted attributes from the trusted medical data on the chain, simplifies the attributes of each type of decision information system in the medical data classification module, obtains the simplified attributes of each type of medical data, provides a core attribute set for the decision-making division of the data to be uploaded to the chain, and divides the subsequent medical data into granularities until the decision cost is greater than the division value of the remaining data;

[0010] 4) The medical data module to be put on the chain refers to the medical data to be put on the chain, which is the local data of the hospital to be put on the chain for trusted division. The trusted core attribute set simplified by the attribute simplification module of the on-chain medical data is used to perform multi-granularity stratification of the data to be put on the chain, so as to reduce the conflicts in the trusted division of the subsequent three decisions;

[0011] 5) The trusted partitioning module for multi-granular medical data divides the data to be uploaded to the chain at each granular layer into three data domains: Pos_data, Bnd_data, and Neg_data. The Pos_data domain is considered trusted data and can be uploaded to the chain directly. The Bnd_data domain needs to be further divided at a finer granular layer. The Neg_data domain is returned to the hospital as untrusted data and is refused to be uploaded to the chain.

[0012] Preferably, the attribute reduction module of the on-chain medical data is based on a hypergraph model.

[0013] Preferably, the multi-granularity stratification process of the data to be uploaded to the chain includes:

[0014] 4-1) Coarse-grained layer division based on the vertex cover set:

[0015] Let the number of attributes of the minimal vertex cover set mi2Cs be denoted as 丨mi2Cs丨. Since the attributes in the minimal vertex cover set cannot be further reduced, the minimal vertex cover set with fewer attributes is used as the attribute set of the coarse-grained layer. As the number of attributes in the cover set increases, the corresponding grain layer becomes finer. Its formal expression is

[0016]

[0017] where GL i is the i-th grain layer of the medical data, Snmi2Csi is the serial number of the minimal vertex cover set mi2Csi in the sorting of the number of attributes in its cluster, ASi is the attribute set of the i-th grain layer corresponding to the minimal cover set mi2Csi, i represents the grain layer serial number, and its value range is: 1 ≤ i ≤ |vc|, and |vc| is the size of the minimal vertex cover set cluster.

[0018] 4-2) Fine-grained division based on the digital characteristics of rough sets:

[0019] Describe a group or a single system parameter of the system through the importance degree of the system attribute parameters in the decision information table, obtain the importance degree of each system parameter for the relevant patient universe,剔除 the parameters with an importance degree of 0, and denote them as set A sig ;剔除 the attributes with a dependence degree of 0, and denote them as set B γ , taking A sig , B γ as the basic attribute set for fine-grained division; by calculating the intersection Inset, union Unset and difference set Diffset of A sig and B γ , and using the uncertainty method to simplify the attributes of the difference set, obtain the attribute set for fine-grained division based on the digital characteristics of rough sets.

[0020] The beneficial effects of the present invention are:

[0021] (1) Generalize the classical spectral graph theory to the hypergraph model, and study the spectral characteristics of the hypergraph Laplacian and its bipartite graph. Discover the frequency relationship between the independence and dependence of hypergraph vertices and their non-trivial Laplacian eigenvectors, and the consistency relationship between the vertex degrees after the hypergraph is converted into a bipartite graph and the element sizes of the trivial eigenvectors of the hypergraph normalized Laplacian matrix.

[0022] (2) Using the two hypergraph feature information in (1), we propose an attribute reduction algorithm based on hypergraph vertex cover (ARHVC). This algorithm transforms the hypergraph vertex cover problem into a combinatorial problem of traditional bipartite graphs, which can eliminate redundant attributes in medical data and improve the average classification ability of attributes.

[0023] (3) A medical data granularity stratification strategy is proposed by combining vertex cover set attributes with common attributes selected by rough set digital features. This strategy first uses the minimum vertex cover set attribute set and the extremely small vertex cover set attribute set as the stratification attributes of coarse-grained data, and then proposes a granularity partitioning algorithm based on rough set digital features (GPRSDF) to measure the granularity partitioning ability of common attributes outside the cover attribute set, and obtains the candidate set for fine-grained data stratification. This strategy can effectively reduce the conflict problem of sequential three-branch decision making and improve its credible partitioning accuracy for medical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a diagram of the data flow of the medical blockchain source data trust analysis model;

[0025] Figure 2 It is: medical data flow diagram;

[0026] Figure 3 For: A simple hypergraph example of community health check-up;

[0027] Figure 4 For: Blood routine decision information table induced hypergraph;

[0028] Figure 5 To: Trustworthy partitioning of medical data based on multi-granularity sequential three-branch decision; DETAILED DESCRIPTION

[0029] The present invention is further described in detail below in conjunction with specific implementation modes.

[0030] The medical blockchain data trust analysis method based on multi-granularity sequential three-branch decision-making described in the present invention includes the following five modules:

[0031] ①On-chain medical data module

[0032] Module 1 is the current medical blockchain data storage system. We assume that the on-chain medical data is trustworthy. This module is mainly used to obtain the data source of the trusted core attribute, which is used to perform trustworthy division of the on-chain medical data. First, the attributes of each type of trusted medical data on the chain are simplified, and then the trusted decision division of the on-chain data is performed based on the simplified trusted core attribute set, reducing the cost of three decisions from the data attribute level.

[0033] ②Medical data classification module

[0034] Module 2 is the process of classifying medical data. It is a decision-making information system that classifies the trusted data on the chain into different types of disease data according to the ICD (International Classification of Diseases) code.

[0035] ③Attribute simplification module of on-chain medical data

[0036] Module 3 extracts trusted attributes from the trusted medical data on the chain. This part simplifies the attributes of each type of decision-making information system in module 2 to obtain the simplified attributes of each type of medical data, providing a core attribute set for the decision-making division of the data to be uploaded to the chain, and then performs granular division on subsequent medical data until the decision cost is greater than the division value of the remaining data.

[0037] ④ Medical data module to be uploaded to the chain

[0038] Module 4 refers to the medical data to be uploaded to the chain (that is, the local hospital data to be uploaded to the chain for trusted division). The trusted core attribute set simplified in module 3 is used to perform multi-granular stratification on the data to be uploaded to the chain, reducing the trusted division conflicts in the subsequent three decisions.

[0039] ⑤Trustworthy partitioning module for multi-granularity medical data

[0040] The data to be uploaded to the chain at each granular layer is divided into three data domains: Pos_data, Bnd_data, and Neg_data. The Pos_data domain is considered to be credible data and can be uploaded directly to the chain. The Bnd_data domain needs to be further divided at a finer granular layer. The Neg_data domain is returned to the hospital as untrusted data and refused to be uploaded to the chain. Figure 2 As shown in the figure, we agree that the medical data on the blockchain is trusted data, and the on-chain data is taken out and classified according to the ICD code, and the on-chain data is divided into n categories, namely data on = {IS 1 , I.S. 2 , ...IS n}. i The decision information table for each type of disease is

[0041] ISi =(U i , AT i {V a |a∈AT i},{I a :→V a |a∈AT i})(i=1,2,……n)

[0042] Among them U i IS is the domain of the i-th disease (a finite non-empty set of objects), i Representation domain U i ={x 1 x 2 , ...x m}, the domain object in this project is each patient, AT i =C∪D is an attribute set, C and D are called conditional attribute set and decision attribute set respectively, and they satisfy V a It is the attribute a∈AT i The attribute value range of I a is an information function that specifies U i The attribute value of each object in is recorded as x∈U i Under the attribute a∈AT i The attribute value on is a(x).

[0043] The data to be uploaded to the chain from the hospital is the medical information of the patient. We take a category related to IS i The same type of medical data is recorded as IS' i , so from IS i The extracted simplified attributes are also more suitable for the data to be uploaded to the chain IS' i Then use the sequential three-way decision to IS' i Conduct credible division and evaluate IS' i In terms of data credibility, the acceptance domain data Pos_data is directly uploaded to the chain, and the rejection domain Neg_data is returned to the local hospital. Considering the actual situation of multi-granularity division, we believe that when the three-branch decision boundary domain is reduced to a certain extent, the cost of returning to the hospital for verification is lower than continuing to divide the boundary domain, so we will return the boundary domain data Bnd_data that has been reduced to a certain extent to the hospital His system.

[0044] Attribute reduction of medical data based on hypergraph model

[0045] (1) Hypergraph Theory

[0046] Hypergraph theory solves many practical problems related to feature tracking, scene registration, association calculation, and data mining. We will now take the elderly community public welfare physical examination as an example to introduce the connection between hypergraph and attribute reduction.

[0047] Exp1: A hospital organizes a public welfare physical examination activity for the elderly community, which includes k ≥ 1 physical examination project departments: e 1 , e 2 ,…,e k , let E be a set of k departments, with n residents undergoing physical examinations: v 1 , v 2 , …, v n , let V be the set of all residents who have taken physical examinations. Assuming that there is at least one resident who has taken physical examinations in each department, the following hypergraph can be constructed:

[0048] (1) Each resident who participates in the physical examination is regarded as a vertex of the hypergraph.

[0049] (2) Each physical examination department is regarded as a hyperedge.

[0050] Then V={v 1 , v 2 , …, v n} is a vertex set of the hypergraph, E = {e 1 , e 2 ,…,e k} is the hyperedge set of the hypergraph, where the hyperedge e i The meaning of the connected vertices is all the vertices participating in e i The collection of residents of the department.

[0051] Figure 3 This is a simple hypergraph example of community health checkup activities. We will use it to introduce the relationship between hypergraphs and attribute reduction. Figure 3 As shown, we can see that the hypergraph is H = (V, E), where the hypergraph vertex set V = {v 1 , v 2 , v 3 , v 4 , v 5 , v 6}, indicating that each v i (1≤i≤6) represents the community residents who participate in the physical examination, and the hyperedge set is E={e 1 , e 2 , e 3 , e 4 ,}, indicating that each e j (1≤j≤4) represents different physical examination departments, and the vertices included can be expressed as: E = {{v 2 , v 3},{v 1 , v 2 , v 4},{v 2 , v 3 , v5 , v 6},{v 1 v 2}}, indicating participation in e j The specific number of community residents in the department.

[0052] Then based on the above example we can define the hypergraph as follows:

[0053] Def1: Hypergraph, suppose there is a finite set V = {v 1 , v 2 , …, v n}, the subset cluster E of set V = {e 1 , e 2 ,…,e k}, where for any e i satisfy If a subset cluster E is a hypergraph on a set V, then E satisfies the following two conditions:

[0054] (1) where i = 1, 2, ..., k

[0055] (2)

[0056] If the set V and its subset cluster E satisfy the two conditions in the traditional attribute reduction based on positive domain:

[0057]

[0058] Where C' is an attribute reduction based on the positive domain, then we can use a binary pair H = (V, E) to represent a hypergraph, where V = {v 1 , v 2 ,…v n} represents the vertex set, E={e 1 , e 2 ,…,e k} represents a hyperedge set.

[0059] Def2: Induced sub-hypergraph, in a hypergraph H = (V, E), where E = ∪{e i |i∈I}, I is the label set of the hyperedge, then for a non-empty vertex subset An induced sub-hypergraph of a hypergraph H is H'=(V', E'), where

[0060] The vertex cover problem is an NPC (NP Complete) problem in classical graph theory. Similarly, the vertex cover problem also exists in hypergraph theory. Generally speaking, the vertex cover of a hypergraph can be defined as: in the hypergraph H = (V, E), a non-empty subset of vertices In the sub-hypergraph formed, if |E|=|E'|, then the vertex subset K is called a vertex cover of the hypergraph H; further, if any one or more vertices are deleted from a vertex cover K, it is no longer a vertex cover, and the vertex subset K is called a minimal vertex cover of the hypergraph H; in the minimal vertex cover, the set with the smallest number of vertices or the smallest scale is called the minimum vertex cover.

[0061] For example, in Exp1, in a public welfare physical examination activity for the elderly community held by a certain hospital, various departments are distributed in different areas. Due to time and distance reasons, some elderly residents do not want to undergo physical examinations in all departments. They just want to understand the health problems that they need to pay attention to in their age group by communicating with their peers about the physical examination situation. So how can they communicate with only a minimum of elderly people of the same age about the physical examination situation, understand the physical examination situation of elderly people of the same age in the most comprehensive department with the least time and distance cost, and obtain reference health information suitable for individuals?

[0062] The above problem can be formally described as: 1 The participating medical examination departments are How to get from v 2 , v 3 …v n How to select the least number of residents among the residents to obtain the physical examination results of the elderly residents of the same age in all departments, so as to understand the health information of the residents of the same age? We can abstract the above problem into a hypergraph model to obtain its minimum vertex cover. Part of the model hypergraph H hypergraph H'=(V',E'), E'={e 1 , e 2 ,...e k}-E 1 , V'=∪ e'∈E' V(e'). The minimum vertex cover set of the partial hypergraph H' = (V', E') is the solution to the above problem, that is, the core attribute set of the above problem.

[0063] (2) Distinctive Matrix Representation of Medical Data

[0064] Since there is no attribute importance parameter assigned to the original medical data, we need to convert the medical data into a difference matrix form, and use the difference elements as hyperedges to construct a hypergraph model of the medical data to be uploaded to the chain. Abstracting commonly used medical data into a decision information table form not only conforms to the information form of actual medical decision-making, but also facilitates credible research on medical data. Take the blood routine decision information table of patients in Table 1 as an example. The form of the decision information table is defined as: there is a decision information table IS = {U, At = CUD, V, I}, where the domain U = {x 1 , x 2 , …x n}, At is a non-empty finite attribute set, where C is the conditional attribute and D is the decision attribute. V is the range of the finite attribute set At, V = ∪V a (a∈At); I: U×At→V is an information mapping function, I a (x) = V. The equivalence relation E is defined as follows:

[0065]

[0066] E A The set of all the elements of all the different equivalence classes is called U in E A The quotient set under the equivalence relation E A It constitutes a division of the domain U, usually using U / E A express.

[0067] We take the blood routine test data of each patient in the blood routine decision information table in Table 1 as the information decision table DS and make the following statement: domain U = {x 1 , x 2 , ...x n}, conditional attribute set C = {C 1 , C 2 , ... C s}, the decision attribute is D, and the decision attribute D is the result of the doctor's diagnosis of whether the patient has clinical diseases such as hemolytic anemia, sepsis, bacterial and viral infections, etc. Here we only use T as normal and F as abnormal as a unified mark.

[0068] Table 1 Blood routine decision information table

[0069]

[0070] Reference values: red blood cells: 3.5-9.5*10^9 / L, white blood cells: 3.8-5.1*10^12 / L, hemoglobin: 115-150*g / L,

[0071] Hematocrit: 35-50%, neutrophils: 1.8-6.3*10^9 / L, lymphocytes: 1.1-3.2*10^9 / L…….

[0072] In order to better handle the blood routine decision information table, we convert it into Table 2 Blood routine formalized decision table. The specific conversion method is as follows:

[0073]

[0074] where c ijis the jth attribute of the i-th patient in Table 1, where i∈{1, 2, 3...n}, j{1, 2, 3...s}; Normel interval represents the normal variation interval of each attribute, Normal in t erval represents the left interval of each normal interval, Indicates the right interval of each normal interval.

[0075] Table 1 is converted into Table 2, a formalized decision information table for blood routine tests, through formula (2).

[0076] Table 2. Formalized decision table for blood routine

[0077]

[0078] According to the definition of the discernibility matrix, in a decision information system IS = {U, At = CUD, V, I}, the element m of the discernibility matrix M is ij , (i, j∈U) is expressed as follows:

[0079] And all non-empty elements in the discernibility matrix M form the following set:

[0080]

[0081] According to formula (3), we can get the difference matrix M in Table 2 as:

[0082] If the difference matrix of a decision information table is M, then its difference information function can be expressed as:

[0083]

[0084] Where Ele(M) is the formula (4).

[0085] Difference information function f M After the conversion from the conjunctive normal form to the disjunctive normal form, the attribute subset corresponding to each main implicant is called an attribute reduction. Taking the above matrix M as an example, according to formula (5), the complete conjunctive normal form of the decision information table can be calculated as:

[0086] where f M Each element of can be used as a hyperedge of the induced hypergraph, thereby constructing a hypergraph model of the medical decision information table.

[0087] (3) Medical data core attribute reduction algorithm based on hypergraph vertex cover

[0088] After the above analysis, we can regard a difference element of the difference matrix as a hyperedge of the induced hypergraph. However, generally speaking, the dimension of the initial difference matrix of the decision information table is too large, and the dimension of the hypergraph formed is high. We use the absorption law in logical operations, f M The conjunction normal form can be simplified to:

[0089] After the simplification, f M Constructing a hypergraph model based on the conjunction paradigm can greatly reduce the dimension of the hypergraph of medical data to be uploaded to the chain, improve the efficiency of solving the hypergraph minimum vertex cover algorithm and reduce the trusted computing load of the subsequent three-branch decision of the medical data to be uploaded to the chain.

[0090] From the above disjunctive expression f' of the difference matrix M M , we can construct the induced hypergraph H'={V',E'} of the blood routine decision information table in Table 1, where V'={v 1 , v 2 , v 3 , v 4 , v 5 , v 6}, E'={e 1 , e 2 , e 3 , e 4 , e 5}, the vertex set represents the blood routine attributes in Table 1, and the hyperedge represents the patients in Table 1. Figure 4 The blood routine decision information table is shown in the induced hypermap.

[0091] According to the adjacency matrix formula of the hypergraph A=HωH T -D v , where H is the incidence matrix of the hypergraph, ω is the edge weight diagonal matrix of each hyperedge in the hypergraph, and since each patient has equal importance, we do not consider the weight, ω is recorded as the unit matrix, D v Then it represents the diagonal matrix of each vertex degree. v =diag(d(v 1 ), d(v 2 )...d(v s )), So Figure 4 The adjacency matrix A(H) of the induced hypergraph is

[0092]

[0093] According to the Laplace matrix formula L(H) = DA(H), we can obtain the Laplace matrix L(H) of the above hypergraph, where the hypergraph degree matrix is ​​D = diag(D(x 1 ), D(x 2), ..., D(x s )), D(x) = ∑ y∈V a xy , a xy =|{e∈E:x, y∈e}|.

[0094]

[0095] The eigenvector matrix of the hypergraph Laplacian matrix is

[0096]

[0097] From the classical spectral graph theory, we know that the graph Laplacian matrix is ​​semi-positive definite, and the size of its non-trivial eigenvalues ​​can reflect the frequency law of its corresponding eigenvectors. Usually, low-frequency eigenvalues ​​can be used to construct the graph structure. Therefore, we analyze the frequency change law of the non-trivial eigenvectors of the above hypergraph Laplacian matrix. We find that the size of the maximum or minimum frequency eigenvector element f(v) is related to the dependence and independence of the hypergraph vertices. That is, the larger the f(v) value, the more it can reflect the independence or dependence of the corresponding v vertex. The above characteristics can be formally defined as:

[0098]

[0099] in represents the independence or dependence of vertex v, f(v) is the value of the current non-trivial eigenvector element, and f(pv) is the value of the eigenvector element corresponding to the previous vertex of v. This has important implications for solving hypergraph vertex covers.

[0100] On the other hand, in the spectral features of the normalized graph Laplacian on the tree structure, eigenvalue 1 and eigenvalues ​​near eigenvalue 1 are strongly correlated with the minimum vertex cover. Specifically, for eigenvalue 1, its multiplicity is related to the size of the minimum vertex cover, and the 0 entry of its eigenvector corresponds to the vertex in the minimum vertex cover. The tree structure can be generalized as a bipartite graph, Figure 4 The induced hypergraph of the blood routine decision information table is converted into a bipartite graph, denoted as HyBip, and its adjacency matrix can be defined as the following block matrix:

[0101]

[0102] Where H is the incidence matrix of the hypergraph. Then we can get the adjacency matrix of the hypergraph converted to a bipartite graph:

[0103]

[0104] The eigenvector matrix of the normalized Laplacian matrix of the HyBip bipartite graph is further obtained as:

[0105]

[0106] Its normalized Laplace matrix eigenvalues ​​are

[0107] HyBip eig =[-0.0000 0.1235 0.3510 0.5856 0.5918 1.0000 1.4082 1.41441.6490 1.8765 2.0000] In the tree structure, let n be the number of vertices of the tree, |C| be a minimum vertex cover of the tree, when n=2|C|, f=0 is the only solution, because VC is also a minimum vertex cover. In the case of n>2|C|, the unknown quantity is greater than the number of equations, and the dimension of the solution space is n-2|C|, which is the same as the multiplicity of the eigenvalue 1 of the normalized Laplacian matrix.

[0108] In the HyBip bipartite graph, the multiplicity of the eigenvalue 1 of the normalized Laplacian matrix is ​​1, so it must satisfy n-2|C|=mulEigVal(1). After the hypergraph H' is converted to the bipartite graph HyBip, its number of vertices n=11, and the multiplicity of the eigenvalue 1 of its bipartite normalized Laplacian matrix is ​​1, and the corresponding vertex corresponding to the 0 item in the eigenvector is {v 7 , v 8 , v 9 , v 10 , v 11}, which is the minimum vertex cover of this bipartite graph. At this time, |C|=5, that is, the bipartite graph HyBip satisfies n-2|C|=mulEigVal(1).

[0109] In a bipartite graph, the number of vertices in its minimum vertex cover is related to the multiplicity of the eigenvalue 1 of its normalized Laplacian matrix. The number of vertices n, the minimum vertex cover number |C| and the multiplicity of the eigenvalue 1 mulEigVal(1) satisfy n-2|C|=mulEigVal(1); the vertex corresponding to the 0th item of the eigenvector corresponding to the eigenvalue 1 of the normalized Laplacian matrix is ​​the vertex of the minimum vertex cover.

[0110] Although the relationship between the minimum vertex cover of a bipartite graph and its Laplace spectrum has been analyzed, the minimum vertex cover obtained when converting a hypergraph to a bipartite graph is not a vertex cover of a hypergraph, because when the hypergraph is converted to a bipartite graph, the hyperedge is already regarded as a vertex set. The minimum vertex cover problem of a hypergraph is converted into a combinatorial optimization problem of an ordinary bipartite graph, that is, how to solve the problem of how to solve the minimum hypergraph vertex subset that can connect all hyperedge vertex subsets in the bipartite graph corresponding to the hypergraph.

[0111]

[0112]

[0113]

[0114] After the hypergraph is converted to a bipartite graph, solving the hypergraph minimum vertex cover problem becomes solving the problem of solving the minimum hypergraph vertex subset that can connect all hyperedge vertex subsets. To solve this problem, a bipartite vertex priority strategy for solving hypergraph vertex cover is proposed. This strategy divides the bipartite graph vertex set into a hypergraph vertex set V HBi , and the hyperedge set E Hbi , according to the hypergraph vertex set V of the bipartite graph HBi The size of the vertex degree in is used as the selection criterion for the candidate vertex cover set. The larger the vertex degree, the higher the candidate priority. However, considering that as the hypergraph vertex set covers more and more hyperedge sets in the bipartite graph, especially when the hypergraph vertex set V HBi The number of covered hyperedges exceeds |E Hbi | / 2, the influence of vertex degree on vertex candidate priority will be reduced, because the probability of covering repeated hyperedges will continue to increase as the vertex degree increases, which will lead to a decrease in its priority. The candidate priority of hypergraph vertices can be expressed by the following formula:

[0115]

[0116] Where d(V Hbi ) represents the degree of a bipartite graph vertex. Its size is strongly related to the size of the corresponding element of the trivial eigenvector of its normalized Laplacian matrix. The size of the trivial eigenvector element can be used to represent the degree of the corresponding vertex. As shown in formula (6), represents the independent value or dependent value of vertex v, and |E(v)| represents the set of hyperedges that have been covered. According to the vertex priority, the minimal vertex cover set of the hypergraph can be quickly obtained.

[0117] Through the above analysis, the core attribute reduction algorithm ARHVC based on hypergraph vertex cover is obtained, as shown in Algorithm 1.

[0118] Trustworthy Partitioning of Medical Data Based on Multi-granularity Sequential Three-branch Decision

[0119] After obtaining the vertex cover core attribute set of medical data from the on-chain trusted medical data, a granularity partitioning method based on vertex cover sets is proposed. This method divides the vertex cover sets into three levels as the attribute set for the trusted three-branch decision partitioning of medical data at different granularity levels, which provides good classification capabilities for the subsequent sequential three-branch decision partitioning of medical data.

[0120] (1) Multi-granularity layering strategy for medical data

[0121] The above algorithm can be used to obtain the minimum vertex cover and minimal vertex cover of the hypergraph model of the medical data decision information table. Since these two types of vertex covers are sets of simplified attributes, the average attribute importance in the cover set decreases as the number of attributes increases. Therefore, the minimum vertex cover is used as the first granular layer, and the minimal vertex cover is divided into other granular layers according to the number of attributes in its cover set. Finally, the ability to divide data granular layers is measured in the common attributes according to the two numerical characteristics of the rough set, and combined with the simplified attribute set in the minimal vertex cover as a deeper granular layer attribute set. The granular layering process of the medical data set is as follows:

[0122] ① Coarse-grained layer partitioning based on vertex cover sets

[0123] According to Algorithm 1, the simplified attribute set is obtained from the trusted medical data on the chain, which can be divided into the minimum vertex cover set miCs, the minimal vertex cover set mi2Cs, and the ordinary vertex cover set ordCs. According to the degree of attribute coverage domain, the set miCs is used as the first granular layer GL of medical data 1 The attribute set of other granular layers GL i Make the following division.

[0124] The number of attributes of the minimal vertex cover set mi2Cs is |mi2Cs|. Since the attributes in the minimal vertex cover set cannot be reduced any further, the minimal vertex cover set with the fewer attributes is taken as the attribute set of the coarse-grained layer. As the number of attributes in the cover set increases, the corresponding granular layer becomes finer. Its formal expression is:

[0125]

[0126] Among them, GL i is the i-th granular layer of medical data, is the minimal vertex cover set mi2Cs i The number of attributes in its cluster, AS i is the attribute set of the i-th granular layer and the minimum covering set mi2Cs i Corresponding.

[0127] ② Fine-grained division based on rough set digital features

[0128] The medical data was initially divided into granular layers using the core attribute sets in the minimum vertex cover and the extreme vertex cover. However, it is still necessary to add common attributes outside the core attribute set to study the medical data from a deeper granular perspective. Therefore, the two rough set digital characteristics of common attributes (attributes not in the vertex cover set) are comprehensively considered, and the system attribute importance and its knowledge dependency are used to describe its granularity division capability.

[0129] Def3: Importance of system attribute parameters in the decision information table, IS = (U, At = C∪D, {V a |a∈At}), R∈IND(IS) represents a group or a single system parameter (essentially one or a group of attributes) that describes the system. and the partition of the universe U independent of the system parameter Rπ(U)={X 1 , X 2 , ..., X n},(X i The importance of set X on the system parameter R and the importance of partition π(U) on the system parameter R are defined as follows:

[0130]

[0131]

[0132] where bn R (X) is X in and R (X) are the lower approximation set and upper approximation set of X with respect to the knowledge (parameters, attributes) R.

[0133]

[0134]

[0135] From Def3, we can see that the greater the importance of the system parameter R, the greater the domain subset X i The smaller the boundary domain is, the more accurate the R granularity division is.

[0136] Def4: Attribute dependency, given a decision information table definition

[0137]

[0138] is the degree to which knowledge Q (generally an attribute in the decision attribute set D) depends on knowledge P (generally an attribute in the condition attribute set C), denoted as Among them, Pos P (Q) = U X∈U / Q P (X) is the P-positive domain of knowledge Q.

[0139] Through Def3, we can get the importance of each system parameter to the relevant patient domain, remove the parameters with an importance of 0, and record them as set A sig From Def4, we can see that if the decision attribute D 1 For conditional attribute C 1and C 2 Dependence Then we know that D 1 Knowledge C 1 has a higher degree of dependence. Further, we can get the decision attribute D of each patient i For each patient condition attribute C i The dependency of the attributes with a dependency of 0 is removed and recorded as set B γ , with A sig , B γ As a basic attribute set for fine-grained partitioning.

[0140] Next, by calculating A sig and B γ The intersection Inset, union Unset and difference Diffset of the i and A j , two attributes

[0141]

[0142]

[0143] The uncertainty between sets is:

[0144]

[0145] Gain(A i |A j ) represents the attribute set A i Relative to A j The information gain, H(A i ) represents the attribute set A i If SU(a, Core)≤SU(a, D), then a is taken as the difference attribute, otherwise it is deleted. The difference after uncertainty analysis is recorded as R 2 . The above steps can be used to obtain R 1 =A sig ∩B γ , R 2 , R 3 =A sig ∪B γ As an alternative attribute set for granular partitioning. In summary, we obtain an attribute set for fine-grained partitioning based on rough set digital features.

[0146] The above common attributes are based on the rough set digital feature granularity partitioning algorithm (GPRSDF) as shown in Algorithm 2:.

[0147] (2) Trustworthy Partitioning of Multi-granularity Medical Data

[0148] According to the multi-granularity stratification strategy of medical data, the minimum vertex cover attribute set and the extremely small vertex cover attribute set are combined with the GPRSDF algorithm to perform credible division of medical data into three decisions from different granularity levels.

[0149] The classic three-branch decision method was proposed by Yao and can be defined on medical data sets as:

[0150] Def5: There is a medical decision information table IS = {U, At = CUD, V, I}, where the domain U = {x 1 , x 2 , …x n}, At is a non-empty finite attribute set, where C is the conditional attribute and D is the decision attribute. V is the range of the finite attribute set At, I: U×At→V is an information mapping function,

[0151] I a (x) = V. The equivalence relation E is defined as follows:

[0152]

[0153] E A The set of all the elements of all the different equivalence classes is called U in E A The quotient set under the equivalence relation E A It constitutes a division of the domain U, usually using U / E A express.

[0154] For the domain U there is a stateful set Indicates whether the target object belongs to X. X can be regarded as a concept, and the objects it contains are regarded as the extension of the concept. In the problem of credible partitioning of medical data, X is represented by

[0155] Table 3. Loss functions of three decision branches

[0156]

[0157] The elementary credible category consisting of credible equivalence classes, set Λ={a p , a B , a N} represents the decision action of dividing the domain object into three regions, namely, acceptance, delay and rejection. Table 3 shows the loss function λ of taking different decision actions under different state sets ij (i=P, B, N, j=P, N) where λ pp ,λ BP ,λ NP Indicates that when an object belongs to X, the decision action ap , a B , a N The loss of pN ,λ BN ,λ NN It means that when an object does not belong to X, the decision action is a p , a B , a N In medical decision making, losses can be explained in terms of the harms of treating the unaffected patient and the benefits of treating the affected patient.

[0158] Let [x] E Represents a credible equivalence class divided by the equivalence relation E. In the process of processing medical data, we define the equivalence relation E as:

[0159] E={(v(x i ), d i ])|v(x i ), d i ∈D∧v(x i )≈d}#(15)

[0160] in Represents object x i The property c i The value of w i Its attribute weight.

[0161] [x] E The probability of belonging to set X is expressed as:

[0162]

[0163] The expected loss for each object taking different actions in different states is expressed as:

[0164]

[0165] According to the Bayesian decision criterion, the action set with the smallest expected loss is selected as the best solution to the problem, and the following three decision criteria can be obtained:

[0166] P: and At the same time, then x∈POS(X)

[0167] B: and At the same time, then x∈BND(X)

[0168] N: and At the same time, then x∈NEG(X)

[0169] Because of the conditional probability relationship:

[0170]

[0171] Then the above rules P, B, N are only related to the classification conditional probability p(X|[x] E ) and the loss function λ ij (i=P,B,N,j=P,B,N). And there exists a loss of accepting the correct object is not greater than the loss of delaying the acceptance of the correct object, and both losses are less than the loss of accepting the wrong object; similarly, the loss of rejecting the wrong object is not greater than the loss of delaying the rejection of the wrong object, and both losses are less than the loss of accepting the wrong object. Therefore, the size relationship between the loss functions satisfies the following relationship:

[0172] 0≤λ PP ≤λ BP ≤λ NP and 0≤λ NN ≤λ BN ≤λ PN #(19)

[0173] Then, substituting the above conditional probability relationship (18) and loss function relationship (19) into the expected loss (17), the decision rule can be rewritten as:

[0174]

[0175] The two thresholds α and β are expressed as follows

[0176]

[0177] From formula (21), we can see that the values ​​of α and β are related to the loss function λ ij The final result of the classic three-way decision in the medical data set is determined by both the conditional probability and the loss function. Decision rough set gives a precise mathematical definition of the three-way decision from the perspective of Bayesian minimum risk.

[0178] Sequential three-branch decision-making, as the name implies, is a multi-step and multi-level three-branch decision-making. Sequential three-branch decision-making is a theory that integrates three-branch decision-making with granular computing. In a large amount of medical data, we need to make credible divisions from different knowledge granularities. The core idea of ​​sequential three-branch decision-making is to first obtain preliminary knowledge decisions at a coarse granularity, and then make three-branch decisions at each granular level by gradually adding attributes or other information, and gradually obtain the final decision domain. Sequential three-branch decision-making analyzes medical data at different granularity levels, reduces decision-making costs, and improves decision-making efficiency.

[0179] Def6 For decision table IS = {U, At = C∪D, V, I}, let C 1 , C 2 ...Ck Represents a set of conditional attributes that satisfy In the case of the sequential conditional attribute set, the multi-granularity space is denoted as GS at the i-th (i=1, 2, ...k) layer GL i , GL i and GS is defined as follows:

[0180] GL i = {U i , C i ∪D,V i , I i}#(twenty two)

[0181] GS = {GL 1 , GL 2 , ...GL j}#(twenty three)

[0182] In the multi-granularity space, the threshold (α i , β i )Then the acceptance domain, boundary domain, and rejection domain of the i-th layer can be expressed as

[0183]

[0184]

[0185]

[0186] A multi-granular three-branch decision partitioning model based on medical data Figure 5 As shown. i represents the domain of the i-th granular layer, Represents the object set of the i-th layer. Dynamic domain update is an important part of the sequential three-branch decision-making process. Through the above introduction, we can understand that the cost of dividing knowledge attributes at the coarse data granularity level of sequential three-branch decision-making is much smaller than that at the fine granularity level, and when the positive or negative domain samples are divided at the coarse granularity level, they are no longer considered at the fine granularity level, which not only reduces the division load of object samples, but also greatly reduces the decision cost. On the other hand, with the granularity GL i The number of samples in the boundary area is getting smaller and smaller. At this time, the three-branch decision becomes a two-branch decision, which also reduces the cost of data partitioning to a certain extent.

[0187] The present invention simplifies the attributes of medical data before reliable division. First, the redundant attributes of the domain data itself are eliminated, the decision division time is improved at the data attribute level, and the cost loss of sequential decision-making is reduced. On the other hand, in order to minimize the decision conflict of each granular layer, the present invention uses a multi-granularity stratification strategy for medical data, and uses the division of the previous layer as a reference for the next granular layer. The next granular layer mainly focuses on different parts of the decision, reduces decision conflicts, and further improves the decision accuracy and efficiency of each granular layer.

[0188] The above descriptions are only some specific implementation methods of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any person familiar with the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A trusted analysis method for medical blockchain data based on multi-granularity sequential three-branch decision-making, characterized by It includes the following five modules: 1) The on-chain medical data module, which is a medical blockchain data storage system; 2) The medical data category classification module, which is a process of classifying medical data. It is a decision-making information system that classifies the on-chain trusted data into different types of disease data according to the ICD international disease classification code; 3) The attribute reduction module of on-chain medical data, which extracts trusted attributes from the on-chain trusted medical data, reduces the attributes of each type of decision-making information system in the medical data category classification module, obtains the reduced attributes of each type of medical data, provides a core attribute set for the decision-making division of the data to be uploaded to the chain, and performs granularity division on the subsequent medical data until the decision-making cost is greater than the division value of the remaining data and then stops; 4) The medical data module to be uploaded to the chain, which refers to the medical data to be uploaded to the chain. It is the data to be uploaded to the chain locally in the hospital for trusted division. The trusted core attribute set reduced by the attribute reduction module of on-chain medical data is used to perform multi-granularity stratification on the data to be uploaded to the chain, reducing the trusted division conflict of the subsequent three-way decision-making; 5) The trusted division module of multi-granularity medical data, which divides the data to be uploaded to the chain at each granularity layer into three data domains: Pos_data, Bnd_data, and Neg_data. Among them, the Pos_data domain is regarded as trusted data and can be directly uploaded to the chain. The Bnd_data domain needs to be further divided at a finer granularity layer. The Neg_data domain is returned to the hospital as untrusted data and is refused to be uploaded to the chain.

2. The medical blockchain data trust analysis method based on multi-granularity sequential three-branch decision-making according to claim 1 is characterized in that: The attribute reduction module of the on-chain medical data is based on a hypergraph model.

3. The medical blockchain data trust analysis method based on multi-granularity sequential three-branch decision-making according to claim 1 is characterized in that: The process of performing multi-granularity stratification on the data to be uploaded to the chain includes: 4-1) Coarse-grained layer division based on the vertex cover set: Denote the number of attributes of the minimum vertex cover set mi2Cs as 丨mi2Cs丨. Since the attributes in the minimum vertex cover set cannot be reduced any further, the minimum vertex cover set with fewer attributes is used as the attribute set of the coarse-grained layer. As the number of attributes in the cover set increases, the corresponding granularity layer becomes finer; its formal expression is As i =mi2Cs i Among them, GL i is the i-th granular layer of medical data, is the minimal vertex cover set mi2Cs i The ordinal number in its cluster sorted by the number of attributes, As i is the attribute set of the i-th granular layer and the minimum covering set mi2Cs i Correspondingly, i represents the granular layer number, and its value range is: 1≤i≤|vc|, |vc| is the size of the minimal vertex cover cluster; 4-2) Fine-grained division based on the digital characteristics of rough sets: The importance of the system attribute parameters in the decision information table is used to describe a group or a single system parameter of the system, and the importance of each system parameter to the relevant patient domain is obtained. The parameters with an importance of 0 are eliminated and recorded as set A. sig ; Eliminate the attributes with a dependency of 0 and record them as set B γ , with A sig , B γ As the basic attribute set for fine-grained division; by calculating A sig and B γ The intersection Inset, union Unset and difference Diffset of the two sets are obtained, and the attributes of the difference set are simplified by using uncertainty method to obtain the attribute set of fine-grained division based on the digital features of the rough set. The decision information table is defined as: IS = (U, At = C∪D, {V a |a∈At}), U is the domain, At is the attribute set, V a is the attribute value range; C and D are called the conditional attribute set and decision attribute set respectively, and they satisfy Define the importance of set X with respect to the system parameter R. The system parameter is calculated by the formula: Calculate; where bn R (X) is X in and R (X) are the lower approximation set and upper approximation set of X with respect to knowledge R; The dependency is given by the formula: calculate; The degree to which knowledge Q depends on knowledge P is denoted as Among them, Pos P (Q) = U X∈U / Q P (X) is the P-positive domain of knowledge Q; knowledge Q is an attribute in the decision attribute set D; knowledge P is an attribute in the conditional attribute set C.

Images