A joint medication mining method based on binary high-dimensional data graph clustering

Abstract

The application discloses a kind of based on binary high-dimensional data graph clustering combined medication mining method, comprising: S1, obtaining binary high-dimensional sparse data set;S2, binary high-dimensional data graph clustering algorithm is used to the clustering of binary high-dimensional sparse data set, executes the following operation: S21, binary high-dimensional sparse data set is converted into first matrix, and similarity matrix is calculated;S22, each element in similarity matrix is divided by corresponding row main diagonal element to form second matrix, and adjacency matrix is calculated;S24, adopt graph cutting method and cut into c subgraph without direction weight graph, obtain clustering result A={A1, A2,..., A c};S3, based on association rule algorithm, the object of each clustering cluster in clustering result is associated with rule mining, obtains the combined medication association result of corresponding clustering cluster.The method can more quickly, better excavate combined medication scheme in medical service.

Description

Technical Field

[0001] This invention belongs to the field of data mining technology, specifically relating to a method for mining combined drug use based on binary high-dimensional data graph clustering. Background Technology

[0002] The identification and application of combination drug regimens in clinic prescriptions is an important aspect of improving clinic medical services. Combination drug regimens primarily involve the simultaneous or sequential use of two or more drugs, aiming to enhance efficacy while mitigating side effects. For example, in the treatment of hypertension, long-term use of antihypertensive drugs such as guanethidine, minoxidil, and diazoxide can lead to sodium and water retention, weakening their antihypertensive effect. In such cases, the addition of diuretics can enhance the therapeutic effect of stronger antihypertensive drugs. In the medical field, identifying such combination drug regimens facilitates physicians' diagnosis and treatment of patients, and enables pharmacists to improve the quality of pharmaceutical services, thereby enhancing the overall quality of clinic services.

[0003] However, this task faces significant challenges in the healthcare industry. First, there's the data issue: mining combination drug regimens requires substantial medical data, but real-world medical big data suffers from inconsistent quality and non-standardized data structures, all of which affect the effectiveness and reliability of data mining results. Furthermore, medical data involves patient privacy and must adhere to strict privacy protection regulations. Second, there's the model issue: traditional clustering methods perform well with low-dimensional data but struggle with high-dimensional data. This is because traditional clustering methods typically use Euclidean distance as a similarity metric, which can lead to the "curse of dimensionality" problem in high-dimensional sparse datasets. Association rules are often used to effectively uncover implicit relationships between items in a dataset and to discover their intrinsic connections, but these models require significant computational resources and time. The interpretability of the model also needs to be considered so that doctors can understand and accept the mining results. Therefore, addressing the low representational power and poor robustness of Euclidean distance in calculating the similarity of binary vectors for clinic prescriptions, this paper proposes a combination drug mining method based on binary high-dimensional data graph clustering to better uncover combination drug regimens. Summary of the Invention

[0004] The purpose of this invention is to address the above-mentioned problems by proposing a combination drug discovery method based on binary high-dimensional data graph clustering, which can discover combination drug solutions in medical services more quickly and effectively.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The present invention proposes a method for discovering combined drug use based on binary high-dimensional data graph clustering, comprising the following steps:

[0007] S1. Obtain a binary high-dimensional sparse dataset. in, This represents the nth sample point, i.e., the nth prescription. Each sample point includes m drug names, and n is the total number of sample points.

[0008] S2. The binary high-dimensional data graph clustering algorithm is used to cluster the binary high-dimensional sparse dataset. The binary high-dimensional data graph clustering algorithm performs the following operations:

[0009] S21. Transform the binary high-dimensional sparse dataset into the first matrix U, and calculate the similarity matrix S, as shown in the following formula:

[0010] S=UU T ={s ij} n*n

[0011] in,

[0012]

[0013] In the formula, u nm Let s be the m-th element of the n-th sample point, where T represents the transpose. ij Let i be the element in the i-th row and j-th column of the similarity matrix S, where i,j = 1, 2, ..., n;

[0014] S22. Divide each element in the similarity matrix S by the corresponding row diagonal element to form a second matrix S′, and calculate the adjacency matrix W, as shown in the following formula:

[0015]

[0016] S23. Represent the adjacency matrix W as an undirected weighted graph;

[0017] S24. Use the graph cutting method to cut an undirected weighted graph into c subgraphs, and obtain the clustering result A = {A1, A2, ..., A}. c}, A c Let c represent the cluster of the c-th subgraph, where c << n;

[0018] S3. Based on the association rule algorithm, perform association rule mining on the objects in each cluster in the clustering results to obtain the combined drug association results of the corresponding clusters.

[0019] Preferably, the adjacency matrix W is represented as an undirected weighted graph, as follows:

[0020] When w ij When w > 0, the i-th sample point and the j-th sample point are connected by an edge. ij When w = 0, the i-th sample point and the j-th sample point are not connected by an edge, where wij The value in the i-th row and j-th column of the adjacency matrix W represents the similarity between the i-th and j-th sample points. ij =w ji .

[0021] Preferably, the image cutting method is either the RatioCut method or the Ncut method.

[0022] Preferably, the association rule algorithm is the Apriori algorithm.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This invention enables data clustering of unlabeled binary high-dimensional sparse datasets. First, based on the matrix transformed from the binary high-dimensional sparse dataset, a similarity matrix is ​​obtained using matrix-vector product. Compared to the Euclidean distance used in existing spectral clustering methods, matrix-vector product better expresses the intra-class similarity and inter-class separability of binary vectors, solving the suboptimal clustering problem caused by obtaining the adjacency matrix through the similarity matrix measured by sample point distance in traditional spectral clustering. Second, a graph cutting method is used to cut the undirected weighted graph into multiple subgraphs to obtain clustering results. An association rule algorithm is then used to obtain the combined drug association results for the current cluster, significantly reducing computational load and memory consumption. This method can more quickly and effectively uncover combined drug regimens in medical services, providing a reference for rational drug use by regulatory authorities and clinic physicians. It can also be applied to other fields, such as supermarket bundled marketing and cross-selling analysis of bank customers. Attached Figure Description

[0025] Figure 1 This is a flowchart of the combined drug discovery method based on binary high-dimensional data graph clustering of the present invention;

[0026] Figure 2 This is an undirected weighted graph according to an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram illustrating the classification of antihypertensive drugs according to an embodiment of the present invention;

[0028] Figure 4 This is a visualization of the clustering results in an embodiment of the present invention;

[0029] Figure 5 This is a visualization of the combined drug association results for cluster A5 in an embodiment of the present invention;

[0030] Figure 6 This is a visualization comparing the time overhead (a) and GE index (b) of the prior art and the present application in this embodiment of the invention. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the meaning commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application.

[0033] The combined drug discovery method based on binary high-dimensional data graph clustering proposed in this invention addresses the following practical business needs and current problems in healthcare services:

[0034] First, when diagnosing and treating patients, physicians can maximize the effectiveness of drugs and improve the quality of clinic services by using combination therapy.

[0035] Second, regulatory authorities can promptly grasp medication information for current diseases by observing the frequency and correlation of drugs in combined drug regimens, and can also effectively supervise the medication patterns commonly used in most clinics.

[0036] Third, patients can better understand the treatment of their disease by using combination therapy regimens.

[0037] Fourth, it can be extended to solve clustering problems of binary high-dimensional sparse datasets in other fields.

[0038] like Figure 1-6 As shown, a method for discovering combined drug use based on binary high-dimensional data graph clustering includes the following steps:

[0039] S1. Obtain a binary high-dimensional sparse dataset. in, This represents the nth sample point, i.e., the nth prescription. Each sample point includes m drug names, and n is the total number of sample points.

[0040] S2. The binary high-dimensional data graph clustering algorithm is used to cluster the binary high-dimensional sparse dataset. The binary high-dimensional data graph clustering algorithm performs the following operations:

[0041] S21. Transform the binary high-dimensional sparse dataset into the first matrix U, and calculate the similarity matrix S, as shown in the following formula:

[0042] S=UU T ={s ij} n*n

[0043] in,

[0044]

[0045] In the formula, u nm Let s be the m-th element of the n-th sample point, where T represents the transpose. ij Let i be the element in the i-th row and j-th column of the similarity matrix S, where i,j = 1, 2, ..., n;

[0046] S22. Divide each element in the similarity matrix S by the corresponding row diagonal element to form a second matrix S′, and calculate the adjacency matrix W, as shown in the following formula:

[0047]

[0048] S23. Represent the adjacency matrix W as an undirected weighted graph.

[0049] In one embodiment, the adjacency matrix W is represented as an undirected weighted graph, as follows:

[0050] When w ij When w > 0, the i-th sample point and the j-th sample point are connected by an edge. ij When w = 0, the i-th sample point and the j-th sample point are not connected by an edge, where w ij This represents the value in the i-th row and j-th column of the adjacency matrix W, i.e., the weights of the i-th and j-th sample points. ij =w ji .

[0051] S24. Use the graph cutting method to cut an undirected weighted graph into c subgraphs, and obtain the clustering result A = {A1, A2, ..., A}. c}, A c Let c represent the cluster of the c-th subgraph, where c << n.

[0052] In one embodiment, the graph cutting method is either the RatioCut or Ncut method. Alternatively, other existing graph cutting methods may also be used.

[0053] S3. Based on the association rule algorithm, perform association rule mining on the objects in each cluster in the clustering results to obtain the combined drug association results of the corresponding clusters.

[0054] In one embodiment, the association rule algorithm is the Apriori algorithm.

[0055] Specifically, the flowchart of the binary high-dimensional data graph clustering algorithm in this embodiment is shown in Table 1-1:

[0056] Table 1-1 Clustering Algorithm for Binary High-Dimensional Data

[0057]

[0058]

[0059] Binary high-dimensional data graph clustering algorithm: First, obtain the similarity matrix S by performing matrix-vector product on the binary vectors of the sample data. Then, construct an undirected weighted graph with n vertices based on the adjacency matrix W. Each edge in the undirected weighted graph is represented by the similarity between two vertices. Finally, use the Ncut method to cut the undirected weighted graph into c subgraphs to minimize the sum of edge weights for each vertex in each subgraph. The detailed steps of the algorithm are as follows:

[0060] (1.1) As a preferred method, the construction method of the undirected weighted graph is as follows:

[0061] (1.1.1) The traditional method of generating similarity matrices based on Euclidean distance is not suitable for binary data. Therefore, this invention proposes to generate similarity matrices quickly through matrix-vector product, as shown in the following equation:

[0062]

[0063] By transforming the input binary high-dimensional sparse dataset into a first matrix U, and then combining it with U... T Multiplying the transposes of the first matrix U yields the similarity matrix S = UU T ,in:

[0064] S={s ij} n*n

[0065] In the formula, i,j=1,2,…,n,l=1,2,…,m.

[0066] (1.1.2) The adjacency matrix W reflects the similarity between vertices in an undirected weighted graph to some extent, and can be obtained through the similarity matrix S. Since the similarity matrix S already contains the sample similarity, there is no need to use the ∈-nearest neighbor method, K-nearest neighbor method, or fully connected method to obtain the adjacency matrix. We only need to divide each element of the existing similarity matrix S by the corresponding row's main diagonal element to form a second matrix S′. Based on the symmetric nature of matrices in undirected graphs, we can then derive the following:

[0067]

[0068] We obtain an n×n adjacency matrix W, where the element in the i-th row and j-th element corresponds to the weight w. ij .

[0069] (1.1.3) The sample points are abstracted into an undirected weighted graph using the adjacency matrix W, such as... Figure 2 As shown in the figure, the set of n m-dimensional vertices in the figure is a binary high-dimensional sparse dataset U. nm ∈R n×m u i Let w represent the i-th vertex. ij Represents vertex u i u j Weights between vertices. For any two vertices in the graph, there may or may not be an edge connecting them. Since it is an undirected graph, w... ij =w ji For two vertices u connected by an edge... i and u j w ij >0, for two points u that are not connected by an edge. i and u j w ij =0.

[0070] (1.2) Preferably, the graph cutting method involves cutting the undirected weighted graph after obtaining it, resulting in different subgraphs, and then classifying each independent subgraph into a category. Generally, there are many ways to segment a graph. This embodiment uses the Ncut graph cutting method, or other existing graph cutting methods such as the RatioCut method can also be used. Different graph cutting methods require different costs. Therefore, a cost function can be defined as the initial objective function. The Ncut graph cutting method is as follows:

[0071] (1.2.1) Formation of the initial objective function

[0072] Suppose we want to divide all vertices into c subgraphs, then let A k This represents the set of vertices contained in the k-th subgraph (i.e., the cluster corresponding to the k-th subgraph). Let U represent the complement of the k-th subgraph, where k = 1, 2, ..., c. nm =A1∪A2∪...∪A c Let Cut(U) nm The cost of slicing the image is represented by ). Subgraph A k And its supplementary image The cost of the cut can be represented by the sum of the weights of all edges connecting the subgraph, thus yielding the following formula (considering symmetry w). ij =w ji ), where i,j=1,2,...,n:

[0073]

[0074] Therefore, we further obtain the preliminary discrete optimization problem, namely the minimum cut objective function:

[0075]

[0076] (1.2.2) Refining the initial objective function by introducing an indicator vector

[0077] After obtaining the minimum cut objective function, the function constraint u i ∈A k , The approach is too general and vague, making it difficult to calculate and solve. Therefore, it is necessary to qualitatively introduce an indicator vector to optimize the minimum cut objective function.

[0078] ① Expansion of the minimum cut objective function

[0079] For two constraints u i ∈A k , Define the indicator vector h k =(r (1,k) ,r (2,k) ,...,r (i,k) ,...,r (n,k) ) T ∈R n×1 And satisfy the following conditions:

[0080]

[0081] Use c indicator vectors h k Combined into an indicator matrix H∈R n×c Assume that each sample point belongs to a certain class with a probability of 1, and there is no case where a sample belongs to multiple classes. Therefore, the minimum cut objective function can be refined using the indicator vector as follows (considering symmetry w). ij =w ji ):

[0082]

[0083] Expanding the quadratic term as follows:

[0084]

[0085] ② Introduction of the non-normalized Laplacian matrix

[0086] The degree d of the i-th vertex in an undirected weighted graph i It is expressed as follows:

[0087]

[0088] Therefore, the above equation can be further expanded as follows:

[0089]

[0090] Where D is the degree matrix, W is the adjacency matrix, and L = DW is the Laplace matrix. Therefore, the original minimum cut objective function is again equivalent to:

[0091]

[0092] Here, Tr represents the trace of the corresponding matrix, and the trace operation returns the sum of the diagonal elements of the matrix. Finally, this is combined with the extended index matrix H∈R. n×c A similar minimum cut objective function can be obtained. However, due to the limitations of the minimum cut objective function, other constraints need to be added to improve the segmentation effect of the objective function.

[0093] (1.2.3) Improvement of the objective function (Ncut slicing method)

[0094] As the minimum cut objective function shows, the segmentation result obtained by minimum cut tends to separate isolated points with the fewest connected edges and lower edge weights. To achieve a balanced result, the original minimum cut objective function needs to be divided by the size of each subgraph after dicing. Since it is an undirected weighted graph, the Ncut dicing method uses the sum of edge weights of the subgraph as the size of the subgraph, as follows:

[0095] Subgraph A k volume

[0096] Where k = 1, 2, ..., c. Based on the definition of scale, the improved objective function can be obtained:

[0097]

[0098] (1.2.4) Improved solution of objective function

[0099] Indicator vector h k =(r (1,k) ,r (2 ,k),...,r (i ,k),...,r (n,k) ) T ∈R n×1 The conditions are updated in the improved objective function as follows:

[0100]

[0101] In the formula, vol represents the volume of the subgraph, which is used to represent the sum of the subgraph weights.

[0102] So for have:

[0103]

[0104] The improved objective function remains:

[0105]

[0106] But at this time H T H≠E, but H T DH = E, where E is the identity matrix. The improved objective function is now:

[0107] Ncut(A1,A2,...,A c )=Tr(H T LH)

[0108] s.tH T DH = E

[0109] Introducing the Lagrange multiplier, we obtain the L(H,λ) function:

[0110] L(H,λ)=Tr(H T LH)-λ(H T HE)

[0111] This yields an unconstrained and continuous objective function, and the derivative is obtained by differentiating the function. Let... have to:

[0112] LH=λH

[0113] The above equation shows that the function has extrema when the Lagrange multiplier λ is an eigenvalue of the Laplacian matrix L and the indicator vector H is an eigenmatrix of L. Multiplying both sides of the above equation by H on the left... T The objective function can be reconstructed into the following form:

[0114]

[0115] Since the number of clusters c is a constant, the element that clearly depends on the magnitude of the objective function value is λ, that is:

[0116]

[0117] In other words, the larger the eigenvalue λ of L, the larger the objective function value; the smaller the eigenvalue λ, the smaller the objective function value. Therefore, the optimal graph partitioning problem is relaxed to a continuous space of Laplacian matrices, transforming the continuous optimization problem into an eigenvalue decomposition problem.

[0118] (1.2.5) Use the K-Means algorithm to solve F n×c Real-valued discretization in

[0119] Normalization of the Laplace matrix is ​​usually required. This invention obtains the normalized Laplace matrix L based on random walk.rw ,Right now:

[0120] L rw =D -1 L=D -1 (DW)

[0121] Where D is the degree matrix, a diagonal matrix whose diagonal elements are equal to the sum of the row vectors of the adjacency matrix W. Since the eigenvalues ​​here physically represent the cost of tiling the graph, the c eigenvectors (i.e., indicator vectors h) corresponding to the first c eigenvalues ​​(all eigenvalues ​​are sorted in ascending order) can be used. k ) concatenate them into a feature matrix F n×c ∈R n×c Then, using the traditional K-Means algorithm, the continuous real values ​​are discretized, ultimately achieving clustering A1, A2, ..., A c .

[0122] As a preferred approach, association rules are used to reflect the interdependence and correlation between one thing and other things. They are one of the main methods for constructing data relationships in data mining. This involves identifying frequently occurring samples, sequences, or substructures in a large dataset as frequent itemsets by pre-setting support and confidence levels, and then constructing association rules from these frequent itemsets. These association rules can intuitively reflect the relationships between data and provide effective information, forming rules such as "the occurrence of certain events causes the occurrence of other events."

[0123] (2.1) Formalization of Association Rules

[0124] Let I={I1,I2,...,I e} is an itemset, where I e Let represent the e-th item, corresponding to a drug. If an itemset contains k distinct items, it is called a k-itemset. For example, {drug 1, drug 2} constitutes a binary itemset. Transaction (t) i Let I represent a subset of prescriptions. The set of transactions is denoted as D = {t1, t2, ..., t...} i ,...,t n}, also commonly referred to as a transactional database (i.e., a binary high-dimensional sparse dataset U). nm In a typical description, each transaction has a unique number, denoted as TID, which corresponds to the prescription ID. Each transaction (prescription) contains several items (medications).

[0125] Association rules are generally denoted in the form X→Y, where X and Y are disjoint itemsets. The itemset X on the left side of the association rule is called the prerequisite, and the itemset Y on the right side is the corresponding association result, used to represent the implicit association within the data. For example, if Drug 1→Drug 2 holds true, it means that there is a certain association between these two drugs, indicating that Drug 1 is often used in conjunction with Drug 2. The degree of association is controlled and evaluated by the three core concepts in association analysis—support, confidence, and lift.

[0126] 1) Support is understood as the probability of occurrence in all itemsets {X, Y}, i.e., the probability that an itemset contains both X and Y simultaneously: support(X→Y) = P(X,Y), where P(X,Y) represents the probability that X and Y occur simultaneously. This metric serves as the first threshold for establishing strong association rules. Its significance lies in setting a minimum support threshold to eliminate meaningless rules with low "occurrence rates," while retaining the rules implied by more frequently occurring itemsets. Itemets with support not less than minsup are considered frequent; otherwise, they are considered infrequent.

[0127] 2) Confidence is understood as the probability that the association result Y occurs given that the prerequisite X of the association rule has occurred. That is, the probability that Y is also present in an itemset containing X is: confidence(X→Y) = P(Y|X) = P(X,Y) / P(X), where P(Y|X) represents the probability that X occurs given Y, and P(X) is the probability that X occurs. This is the second threshold for generating strong association rules. Typically, a minimum confidence threshold (minconf) needs to be set to further filter the data, ultimately generating strong association rules that meet the requirements. Therefore, after filtering out frequent itemsets, rules satisfying confidence(X→Y) ≥ minconf need to be selected, thus completing the generation of the required association rules.

[0128] 3) Lift is understood as the ratio of the probability that Y is also present given X to the probability that Y is present in the term set without X. In other words, it's the degree to which the presence of X increases the "occurrence rate" P(Y|X) of Y, based on the probability P(Y) of Y itself. This indicator, like confidence, is used to measure the reliability of a rule and can be seen as a complementary indicator to confidence.

[0129] Based on the above concepts, the basic algorithm steps of association analysis can be summarized as follows: (1) Select all itemsets that satisfy the minimum support threshold, i.e., frequent itemsets. (2) Find all rules that satisfy the minimum confidence from the frequent itemsets.

[0130] (2.2) Joint scheme based on Apriori algorithm

[0131] The most well-known algorithm for association rule learning is the Apriori algorithm. It is an algorithm used for frequent itemset mining and association rule learning in transactional databases.

[0132] The Apriori algorithm consists of the following steps:

[0133] a) First, perform a scan of the transaction database to count the number of occurrences of each item and form a candidate 1-itemset;

[0134] b) Filter out frequent 1-itemsets based on the minimum support threshold minsup;

[0135] c) Combine the selected frequent 1-itemsets to form candidate 2-itemsets;

[0136] d) Perform a second scan of the transaction database, count each candidate 2-itemset, and filter out frequent 2-itemsets;

[0137] e) Repeat the above steps until the candidate set is empty;

[0138] f) Generate management rules by calculating the corresponding confidence levels based on the generated frequent itemsets.

[0139] The clustering result of the Apriori algorithm on binary high-dimensional data graph clustering algorithm is A={A1,A2,...,A... c Each cluster A of} k The method performs association rule mining on objects in the dataset, which retains the advantages of the Apriori algorithm being suitable for sparse datasets and having a simple principle. It also adapts the Apriori algorithm to perform better on small datasets by dividing the massive dataset, which greatly reduces the amount of computation and memory consumption, thereby obtaining the combination drug regimen corresponding to the current cluster.

[0140] It is easy to understand that both the Ncut slicing method and the Apriori algorithm are existing technologies well known to those skilled in the art, and will not be elaborated upon here.

[0141] This invention integrates a binary high-dimensional data graph clustering algorithm with a joint scheme based on association rules, and combines real medical prescription data to obtain combined medication plans for diseases. The clustering results of the binary high-dimensional data graph clustering algorithm are compared with those of the spectral clustering algorithm in the prior art, further demonstrating the advantages of this invention.

[0142] (1) Data source

[0143] The medical prescription dataset used in this embodiment is prescription drug data collected and integrated into an Oracle database from the clinic information system by the Yiwu Municipal Health and Family Planning Bureau. The data volume is as high as millions of records. For example, the latest year's prescription drug data for hypertension was extracted using "2022.01-2023.01" and "hypertension" as keywords. The prescription drug data includes clinic information tables, consultation information tables, patient information tables, doctor's order information tables, case diagnosis tables, electronic medical records, standard drug tables, etc.

[0144] Inclusion and exclusion criteria for prescription drug data:

[0145] 1) Inclusion criteria:

[0146] ① The included study subjects were mainly diagnosed with hypertension (diagnostic criteria refer to ICD-10);

[0147] ② There are no restrictions on basic information such as age and gender;

[0148] ③ The patient's medical records, including the diagnosis and treatment process and doctor's orders, are complete.

[0149] 2) Exclusion criteria:

[0150] ①The primary diagnosis of the study subjects was not hypertension;

[0151] ② Medical record data is severely lacking.

[0152] The prescription drug data for data mining was exported according to the above standards, as shown in Table 1-2. Since combination drug regimens are being considered, prescriptions containing two or more drugs need to be selected, as shown in Table 1-3.

[0153] Table 1-2 Data Mining Prescription Data Information

[0154]

[0155] Table 1-3 Prescription data for two or more drugs

[0156]

[0157] (2) Data preprocessing

[0158] Data preprocessing is the process of converting prescription drug data into a understandable format. This is also a crucial step in data mining because raw data cannot be used directly; therefore, data quality should be checked before applying machine learning or data mining algorithms. In reality, data is often incomplete, inconsistent, and "dirty" data, making direct data mining impossible or yielding unsatisfactory results. To improve the quality of data mining, data preprocessing is essential, and the main tasks are as follows:

[0159] 1) Data Acquisition

[0160] Based on the data source, prescription medication data from the Oracle database was exported to a CSV file according to the export criteria. The CSV file was read using the `read_csv()` method in Pandas, merging two tables with the same column index from different tables, such as merging columns with matching attributes from the patient information table and the medical record diagnosis table. Therefore, the data integration phase can proceed directly. To protect patient privacy, the prescription data needed to be anonymized to remove important patient information and ensure patient privacy.

[0161] 2) Data Cleaning

[0162] There are many cases of "synonyms" or "different drugs with the same name." If drug names are not standardized, data discrepancies will occur. To address this, a combination of keyword replacement and manual verification is used to standardize drug names in prescription data. Referencing the Chinese Drug Generic Name (CADN), clinic drug lists, and prescription management regulations, terms such as "Amoxicillin Tablets," "Amoxicillin Capsules," and "Amoxicillin Granules" are uniformly changed to "Amoxicillin," as shown in Tables 1-4.

[0163] Table 1-4 Prescription Drug Standard Table

[0164]

[0165] Using the aforementioned prescription drug standardization table, the drug names in 2232 prescriptions were standardized, and the frequencies of the standardized drugs were sorted. Since the goal is to mine prescription combination therapy regimens, a dataset consisting of prescriptions containing more than one drug is required; therefore, the current prescription dataset has a dimension of 2232×163.

[0166] 3) Data Conversion

[0167] Transform the prescription dataset into a binary high-dimensional sparse dataset (e.g., denoted as U). 2232×163 This is used as input data for a binary high-dimensional data graph clustering algorithm. A 1 indicates that the drug is present in a prescription, and a 0 indicates that it is not present. As shown in Table 1-5, C_REG_ID represents the prescription ID.

[0168] Table 1-5 Prescription Drug Fact Sheet

[0169]

[0170] (3) Data Mining and Analysis

[0171] 1) Binary high-dimensional data graph clustering algorithm

[0172] The binary high-dimensional data graph clustering algorithm proposed in this invention is applied to a cleaned actual prescription dataset (binary high-dimensional sparse dataset) to mine and analyze actual combination drug regimens, thereby helping regulators and medical staff to provide better medical services to patients.

[0173] To obtain the similarity matrix S between prescriptions through matrix-vector product, the input binary high-dimensional sparse dataset matrix needs to be transposed. As shown in Table 1-6, the similarity relationship between prescriptions can be obtained.

[0174] Table 1-6 Drug Prescription Fact Sheet

[0175] In this similarity matrix, the elements on the diagonal represent prescription medication information; larger values ​​indicate a wider variety of prescription medications, while smaller values ​​indicate a simpler variety. The similarity matrix S = {s...} ij} 2232×2232 As shown in Table 1-7, where s ij Indicates prescription With prescription The higher the similarity value, the more similar the two values ​​are, which is the opposite of the similarity value measured by distance in existing technologies.

[0176] Table 1-7 Prescription Similarity Matrix

[0177]

[0178] To efficiently obtain the adjacency matrix W, we need to divide the similarity of each prescription (row) in the similarity matrix S by the similarity of the current prescription (main diagonal) to form a second matrix S'. Furthermore, due to the symmetry of the adjacency matrix in an undirected graph, the prescriptions... With prescription The weights between them can be calculated by taking the average, that is:

[0179]

[0180] Thus, the symmetric adjacency matrix between prescriptions can be obtained efficiently and accurately, as shown in Table 1-8:

[0181] Table 1-8 Prescription Adjacency Matrix

[0182]

[0183] The adjacency matrix is ​​used to obtain the sum of the weights of each object (row), thus yielding the degree matrix D of the prescription. In this way, the adjacency matrix of the current prescription dataset is obtained through simple matrix operations. Compared to the method in spectral clustering that first obtains the similarity matrix through sample point distance metrics (Euclidean distance) and then derives the adjacency matrix through common edge weight calculations, this invention proposes a similarity method more suitable for actual binary prescription datasets.

[0184] The adjacency matrix and degree matrix of the prescriptions were obtained using binary similarity. The Laplacian matrix and K-Means algorithm of the spectral clustering algorithm were then used to divide the prescription dataset into clusters. By adjusting the cluster parameters, it was found that dividing the prescription dataset into 8 clusters yielded the best results. The clustering results are shown in Table 1-9.

[0185] Table 1-9 Clustering Results of the Binary Prescription Dataset

[0186]

[0187] 2) Joint operation based on association rule algorithm

[0188] Based on the clustering results of the binary high-dimensional data graph clustering algorithm, A5 was randomly selected as the research object. Further analysis was conducted on the combined medications prescribed for hypertension to understand the clinical medication characteristics of hypertension and provide a reference for rational drug use in clinics.

[0189] The results of the two associations are shown in Table 1-10. Apart from drugs commonly used to regulate water, salt, electrolyte, and acid-base balance, such as glucose, sodium chloride, and potassium chloride, most clinics choose to use combinations of drugs such as indapamide and telmisartan, nifedipine and indapamide, and nifedipine and telmisartan to treat hypertension.

[0190] Table 1-10 Association Rules for Two Common Drug Combinations

[0191]

[0192] The results of the three associations are shown in Table 1-11. It can be seen that indapamide + telmisartan, gastrodin + mannitol, tanshinone + gastrodin, indapamide + gastrodin, and indapamide + tanshinone are commonly used combination drugs in clinics to improve hypertension.

[0193] Table 1-11 Association Rules for Three Common Drug Combinations

[0194]

[0195]

[0196] Based on the association rules of the Apriori algorithm, each cluster of the clustering results is analyzed to obtain the corresponding combination drug association results. Different combination drug regimens can be obtained for different symptoms. Common combination drug regimens for treating hypertension in clinics are shown in Table 1-12:

[0197] Table 1-12 Combined Medication Regimens for Hypertension

[0198]

[0199] Because there are many types of drugs with similar effects in treating hypertension, common antihypertensive drugs can be classified into five categories: calcium channel blockers (CCBs), angiotensin-converting enzyme inhibitors (ACEIs), angiotensin receptor blockers (ARBs), diuretics, and beta-blockers. These can all be used as initial or long-term maintenance medications for antihypertensive treatment. In addition, alpha-blockers or other types of antihypertensive drugs can sometimes be used in certain hypertensive individuals. Common antihypertensive drug classifications are as follows: Figure 3 As shown.

[0200] Based on drug classification, the preliminary combination drug regimens in Table 1-12 can be converted into general combination drug regimens to facilitate rational drug use by regulatory authorities and clinic physicians, as shown in Table 1-13.

[0201] Table 1-13 General Combination Drug Regimens for Hypertension

[0202]

[0203] (4) Results visualization

[0204] The purpose of presenting data mining and analysis results in the form of charts and graphs is to enable direct, fast, and efficient understanding.

[0205] 1) Visualization of binary high-dimensional data graph clustering results

[0206] t-SNE (t-Distributed Stochastic Neighbor Embedding) is a dimensionality reduction technique used to represent high-dimensional datasets in a low-dimensional space (two or three dimensions), thus enabling visualization. t-SNE works by converting the similarity between data points into joint probabilities and attempts to minimize the KL divergence between the joint probabilities of the low-dimensional embedded data and the high-dimensional data. Visualization of clustering results is shown below. Figure 4As shown in the figures, Figure (a) represents the clustering results of spectral clustering (previous technology), and Figure (b) represents the clustering results of binary graph clustering (this application). The results show that the clustering results of spectral clustering are largely mixed together, with extremely poor boundary characteristics, making them difficult to distinguish. In contrast, the clustering results of binary graph clustering have clearer boundaries and are of better quality.

[0207] 2) Visualization of joint scheme results based on association rules

[0208] A cluster of randomly selected clusters is used to mine association rules in order to discover combination drug regimens for prescriptions. For example... Figure 5 The image shows a visualization of the association results of combined drug use in A5. Among the multiple associations, indapamide, telmisartan, nifedipine, gastrodin, and tanshinone are more likely to be used in combination, providing a reference for rational drug use in clinics.

[0209] (5) Algorithm Evaluation

[0210] Because complete labeled data is rarely available in real-world datasets to aid clustering, the evaluation method based on clustering results can refer to the principle of the internal evaluation metric, the Davies-Bouldin Index (DB Index). The core idea of ​​the DB Index is to calculate the similarity between each cluster and its most similar cluster, and then measure the overall clustering result by averaging all similarities. The metric in this invention can utilize the weights w of the generated undirected weighted graph, calculated using the following formula:

[0211]

[0212] Among them, s ij Let be the element in the i-th row and j-th column of the similarity matrix S (i.e., the similarity). The similarity between clusters is defined as the ratio of the inter-cluster weight to the intra-cluster weight between two clusters. This represents the average weight among all sample points in the p-th cluster. Let g represent the average weight among all sample points in the q-th cluster, also known as the intra-cluster weight, where p,q = 1, 2, ..., c; if we select the object with the highest weight in the cluster as the cluster center, then g... pq J(A) represents the similarity between the cluster centers of the p-th and q-th clusters, also known as the inter-cluster weight. p A q ) represents cluster A p and cluster A q The similarity index is used to measure the similarity between two clusters. It is defined as the number of elements in the intersection of the two sets divided by the number of elements in the union, which is suitable for binary datasets.

[0213]

[0214] At this point, the similarity R between the p-th cluster and the q-th cluster is defined. pq for:

[0215]

[0216] Furthermore, the formula for calculating the evaluation metric GE (Graph-Evaluation of Binary) is as follows:

[0217]

[0218] Where c represents the total number of clusters. The GE index ranges from [0, +∞], with smaller values ​​indicating better clustering results and vice versa. The clustering results from the analysis of real prescription datasets are shown in Table 1-14:

[0219] Table 1-14 Clustering results of the real prescription dataset

[0220]

[0221] like Figure 6 As shown in the visualization, the time overhead and GE index of the prior art and this application are compared. In Figure (a), the time overhead of the binary graph clustering algorithm (i.e., the broken line connecting the pentagrams) is less than that of the spectral clustering algorithm (i.e., the broken line connecting the squares), thus achieving optimization in running efficiency. In Figure (b), since the GE index represents the ratio of inter-cluster weight to intra-cluster weight, where the smaller the inter-cluster weight and the larger the intra-cluster weight, the smaller the GE index and the better the clustering result, it can be seen from the figure that the GE index of the binary graph clustering algorithm is less than that of the spectral clustering algorithm, and it is the smallest when the total number of clusters is 8, indicating that the proposed binary high-dimensional graph clustering algorithm is superior to the spectral clustering algorithm in terms of the accuracy of the clustering results.

[0222] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0223] The embodiments described above are merely specific and detailed examples of the embodiments described in this application, and should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the appended claims.

Claims

1. A method for discovering combined drug use based on binary high-dimensional data graph clustering, characterized in that: The combined drug discovery method based on binary high-dimensional data graph clustering includes the following steps: S1. Obtain a binary high-dimensional sparse dataset. in, This represents the nth sample point, i.e., the nth prescription. Each sample point includes m drug names, and n is the total number of sample points. S2. A binary high-dimensional data graph clustering algorithm is used to cluster the binary high-dimensional sparse dataset. The binary high-dimensional data graph clustering algorithm performs the following operations: S21. Transform the binary high-dimensional sparse dataset into the first matrix U, and calculate the similarity matrix S, as shown in the following formula: S=UU T ={s ij } n*n in, In the formula, u nm Let s be the m-th element of the n-th sample point, where T represents the transpose. ij Let i be the element in the i-th row and j-th column of the similarity matrix S, where i,j = 1, 2, ..., n; S22. Divide each element in the similarity matrix S by the corresponding row diagonal element to form a second matrix S′, and calculate the adjacency matrix W, as shown in the following formula: S23. Represent the adjacency matrix W as an undirected weighted graph; S24. Use the graph cutting method to cut an undirected weighted graph into c subgraphs, and obtain the clustering result A = {A1, A2, ..., A}. c }, A c Let c represent the cluster of the c-th subgraph, where c << n; S3. Based on the association rule algorithm, perform association rule mining on the objects in each cluster in the clustering results to obtain the combined drug association results of the corresponding clusters.

2. The method for discovering combined drug use based on binary high-dimensional data graph clustering as described in claim 1, characterized in that: The adjacency matrix W is represented as an undirected weighted graph, as follows: When w ij When w > 0, the i-th sample point and the j-th sample point are connected by an edge. ij When w = 0, the i-th sample point and the j-th sample point are not connected by an edge, where w ij This represents the value in the i-th row and j-th column of the adjacency matrix W, i.e., the weights of the i-th and j-th sample points. ij =w ji .

3. The method for discovering combined drug use based on binary high-dimensional data graph clustering as described in claim 1, characterized in that: The image cutting method is either the RatioCut method or the Ncut method.

4. The method for discovering combined drug use based on binary high-dimensional data graph clustering as described in claim 1, characterized in that: The association rule algorithm is the Apriori algorithm.

Images