Cluster weighted clustering ensemble medical text processing method and system based on member selection

Abstract

The present invention provides a method and system for processing medical data based on member selection by cluster weighted clustering integration, wherein the method comprises: constructing a cluster member set and inputting it into a pre-trained decision tree model; filtering out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generating a target cluster group with the filtered cluster members; determining a target CA matrix of the target cluster group according to a cluster layer weighting coefficient; and executing a hierarchical clustering algorithm based on the target CA matrix to obtain a final clustering result. The method and system for processing medical data based on member selection by cluster weighted clustering integration of the present invention can obtain a more optimized clustering result, can better display the characteristics and similarities of a patient's disease, facilitate doctors to select medical services or customize personalized treatment plans for different patients, and help improve the patient's medical experience and treatment effect.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for processing medical data by cluster weighted clustering integration based on member selection. Background Art

[0002] Medical diagnosis often faces a large number of complex cases and clinical data. Doctors need to quickly and accurately classify patients in order to develop personalized treatment plans. However, due to the complexity and diversity of diseases, the rapid classification and prediction of large-scale data has become a challenge; in addition, different patients may show different symptoms and characteristics, which are also influencing factors that need to be considered in classification.

[0003] Cluster analysis is one of the hot topics in machine learning research. It is widely used in data compression, information retrieval, image segmentation and text clustering. It has received increasing attention in the fields of biology, geology, geography and abnormal data detection. Cluster analysis is an unsupervised machine learning method. It lacks prior knowledge of the data set. It automatically divides the data set into several groups or clusters based on the similarity measurement between data points, samples and objects, so that the similarity between points belonging to the same cluster is as high as possible, while the similarity between points belonging to different clusters is as low as possible. Cluster ensemble introduces the idea of ​​ensemble learning into cluster analysis, thus opening up the study of cluster ensemble. It is mainly divided into two steps: the first step is to take the data set as input, run the clustering algorithm, and output multiple different clustering results. This step is called cluster member generation; the second step is to take the set of all cluster members, that is, the cluster collective, as input, combine them, and output the final clustering result. This step is called cluster ensemble, also known as consensus function design; for the classification of medical data, cluster ensemble algorithms are usually used.

[0004] However, most of the existing clustering ensemble algorithms treat each cluster member equally, without paying attention to the quality differences among cluster members. Even if some algorithms propose to treat each cluster member differently and evaluate the cluster members, they ignore the local diversity of clusters within the same cluster member and directly regard the cluster members as independent individuals, which may result in low-quality clusters or low-quality base clusters. Summary of the invention

[0005] One of the purposes of the present invention is to provide a cluster weighted clustering integrated medical data processing method and system based on member selection, which can obtain more optimized clustering results, better display the characteristics and similarities of patients' diseases, facilitate doctors to choose medical services or customize personalized treatment plans for different patients, and help improve patients' medical experience and treatment effects.

[0006] The embodiment of the present invention provides a method for processing medical data by cluster weighted clustering integration based on member selection, comprising:

[0007] Construct a set of cluster members and input them into a pre-trained decision tree model;

[0008] Filter out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generate a target cluster group with the filtered cluster members;

[0009] Determine the target CA matrix of the target clustering group according to the cluster layer weight coefficient;

[0010] The hierarchical clustering algorithm is executed based on the target CA matrix to obtain the final clustering result.

[0011] Preferably, the construction of the cluster member set includes: clustering the medical data using a K-Means algorithm to generate a plurality of cluster members.

[0012] Preferably, the training steps of the decision tree model are as follows:

[0013] Calculate the Davies-Bouldin index of each sample cluster member in the sample cluster member set and find the overall average value;

[0014] Compare the Davies-Bouldin index of each sample cluster member with the average value, label the sample cluster members whose Davies-Bouldin index is lower than the average value with the first label, and label the sample cluster members whose Davies-Bouldin index is higher than the average value with the second label;

[0015] The training is performed based on the sample cluster members with labels as the training set to obtain a trained decision tree model.

[0016] Preferably, determining a target CA matrix of a target clustering group according to the cluster layer weight coefficients includes:

[0017] Construct the CA matrix of the target cluster collective;

[0018] The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B;

[0019] Capture high confidence information from the CA matrix to obtain the HC matrix;

[0020] The target CA matrix of the target cluster group is determined based on the processed data B and the HC matrix.

[0021] Preferably, constructing a CA matrix about the target cluster collective includes:

[0022]

[0023]

[0024] Where A is the CA matrix; Indicates cluster members; Represents the total number of cluster members in the cluster collective; Represents sample points The cluster in which it is located; Represents sample points The cluster where it is located.

[0025] The present invention also provides a cluster weighted clustering integrated medical data processing system based on member selection, comprising: a construction unit, an input unit, a screening unit, a matrix determination unit and an execution unit;

[0026] Among them, the construction unit constructs a cluster member set; the input unit inputs the cluster member set into a pre-trained decision tree model; the screening unit screens out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generates a target cluster group with the screened cluster members; the matrix determination unit determines the target CA matrix of the target cluster group according to the cluster layer weighting coefficient; the execution unit executes the hierarchical clustering algorithm based on the target CA matrix to obtain the final clustering result.

[0027] Preferably, the construction unit uses a K-Means algorithm to cluster the medical data to generate multiple cluster members.

[0028] Preferably, the training steps of the decision tree model are as follows:

[0029] Calculate the Davies-Bouldin index of each sample cluster member in the sample cluster member set and find the overall average value;

[0030] Compare the Davies-Bouldin index of each sample cluster member with the average value, label the sample cluster members whose Davies-Bouldin index is lower than the average value with the first label, and label the sample cluster members whose Davies-Bouldin index is higher than the average value with the second label;

[0031] The training is performed based on the sample cluster members with labels as the training set to obtain a trained decision tree model.

[0032] Preferably, the matrix determination unit performs the following operations:

[0033] Construct the CA matrix of the target cluster collective;

[0034] The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B;

[0035] Capture high confidence information from the CA matrix to obtain the HC matrix;

[0036] The target CA matrix of the target cluster group is determined based on the processed data B and the HC matrix.

[0037] Preferably, the matrix determination unit constructs a CA matrix about the target cluster group and performs the following operations:

[0038]

[0039]

[0040] Where A is the CA matrix; Indicates cluster members; Represents the total number of cluster members in the cluster collective; Represents sample points The cluster in which it is located; Represents sample points The cluster where it is located.

[0041] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0042] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0044] Figure 1 A schematic diagram of a cluster weighted clustering integrated medical data processing method based on member selection in an embodiment of the present invention;

[0045] Figure 2 is a flow chart of constructing a cluster member set according to an embodiment of the present invention;

[0046] Figure 3 is a flow chart of a cluster weighted clustering integration method based on member selection according to yet another embodiment of the present invention;

[0047] Figure 4 is a flowchart of training a decision tree according to one embodiment of the present invention;

[0048] Figure 5 is a schematic diagram of a trained decision tree model according to an embodiment of the present invention;

[0049] Figure 6 It is a schematic diagram of a cluster weighted clustering integrated medical data processing system based on member selection in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0051] The embodiment of the present invention provides a method for processing medical data by cluster weighted clustering integration based on member selection, such as Figure 1 As shown, including:

[0052] Construct a set of cluster members and input them into a pre-trained decision tree model;

[0053] Filter out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generate a target cluster group with the filtered cluster members;

[0054] Determine the target CA matrix of the target clustering group according to the cluster layer weight coefficient;

[0055] The hierarchical clustering algorithm is executed based on the target CA matrix to obtain the final clustering result.

[0056] Among them, the hierarchical clustering algorithm includes: average link method; the preset label is the label marked as high;

[0057] The construction of cluster member set includes: clustering medical data using K-Means algorithm to generate multiple cluster members. Set the number of cluster members r and the number of clusters k. The number of clusters k is set to [2, ], where r∈N+, N is the number of data sample points, and the K-Means algorithm is used to randomly generate r cluster members as the cluster collective P, and then generate a cluster member set, specifically:

[0058] Get the number of cluster members r and the number of clusters k;

[0059] Initialize i=1;

[0060] Determine whether i is less than or equal to r;

[0061] When it is determined that i is less than or equal to r, the K-Means algorithm is used for clustering to generate cluster members and obtain clustering results;

[0062] Assign i=i+1, and continue to judge until i is not less than or equal to r, and then build a cluster member set.

[0063] In addition, cluster member sets can also be generated using clustering methods such as hierarchical clustering and spectral clustering.

[0064] like Figures 2 to 5 As shown, the training steps of the decision tree model are as follows:

[0065] Calculate the Davies-Bouldin index of each sample cluster member in the sample cluster member set and find the overall average value;

[0066] The Davies-Bouldin index of each sample cluster member is compared with the average value, and the sample cluster members with Davies-Bouldin index lower than the average value are given a first label, and the sample cluster members with Davies-Bouldin index higher than the average value are given a second label; the first label is "high" and the second label is "low";

[0067] Based on the labeled sample cluster members as the training set, the trained decision tree model is obtained. The ARI (Adjusted Rand Index), NMI (Normalized Mutual Information) and F (Fowlkes-Mallows Index) of each cluster member are used as the feature attribute set. The closer the three feature attributes are to 1, the better. The value range of ARI is [-1, 1], and the value range of NMI and F is [0, 1]. The Gini coefficient is used to calculate the importance of each feature attribute set to determine the root node and internal nodes of the decision tree. The method is as follows:

[0068] (1.1)

[0069] (1.2)

[0070] in, represents the probability that the sample point belongs to class i, The tables are the feature attribute set and the state of the feature attribute. The larger the Gini coefficient, the greater the uncertainty of the feature attribute.

[0071] First, use formulas (1.1) and (1.2) to calculate the Gini coefficients of the labeled cluster members in terms of ARI, NMI, and F, compare the Gini coefficients of the three feature attributes, select "Feature Attribute 1" with the smallest Gini coefficient as the root node, and mark "Feature Attribute 1" with values ​​close to 1 as "high". Then, use the labeled cluster members whose "Feature Attribute 1" values ​​are not close to 1 as the new round of label sets, continue to calculate the Gini coefficients of the remaining two feature attributes, and select the smallest "Feature Attribute 2" as the internal node. Finally, the remaining feature attributes are used as "Feature Attribute 3" to become the final internal node.

[0072] The closer the three feature attributes are to 1, the better. The value range of ARI is [-1, 1], which is divided into two categories, one greater than 0 and one less than or equal to 0; the value range of NMI and F-measure index is [0, 1], which is evenly divided into two categories, one greater than 0.5 and one less than or equal to 0.5. The Gini coefficient is used to calculate the importance of each feature attribute set to determine the root node and internal nodes of the decision tree. The larger the Gini coefficient, the greater the uncertainty of the feature attribute.

[0073] The attributes of the root node and internal nodes of the decision tree are measured using the Gini coefficient;

[0074] For example, to find the Gini coefficient of the ARI attribute, the details are as follows:

[0075] According to the following formula:

[0076] First, calculate the Gini coefficient when ARI is greater than 0:

[0077] at this time, Representing a subset of data Gini impurity of ; Represents the data subset obtained when the feature attribute ARI is greater than 0; It means the number of samples with the class label "high" divided by the number of samples with ARI greater than 0; It means the number of samples with the class label "low" divided by the number of samples with ARI greater than 0;

[0078] Secondly, the Gini coefficient when ARI is less than or equal to 0 is calculated as:

[0079] at this time, Representing a subset of data Gini impurity of ; Represents the data subset obtained when the feature attribute ARI is less than or equal to 0; It means the number of samples with the class label "high" divided by the number of samples with ARI less than or equal to 0; It represents the number of samples with the class label "low" divided by the number of samples with ARI less than or equal to 0;

[0080] Finally, according to the following formula:

[0081] The Gini coefficient of ARI is calculated as follows:

[0082] at this time, Represents the Gini coefficient of the dataset D when the feature attribute is ARI; Indicates the number of samples when ARI is greater than 0, Represents the number of samples in the dataset D.

[0083] When building a decision tree, the value of the feature attribute is equally divided into two different value ranges, using two branches.

[0084] The decision tree selects the best feature attribute as the root node based on the Gini coefficient, and continues to select the feature attributes with higher importance in the subsequent internal node division. After training, the decision tree model can be used to classify and predict new cluster members.

[0085] When constructing a decision tree, the value of the feature attribute can be further subdivided into multiple segments with different value ranges, that is, "multi-branching", which increases the branches of each decision tree, increases the number of decision classifications, and thus improves the accuracy of classification.

[0086] Based on the above method, the clustering algorithm and the decision tree are effectively combined to obtain a decision tree model. The decision tree selects the best feature attribute as the root node according to the Gini coefficient, and continues to select feature attributes with higher importance in the subsequent internal node division. The decision tree model can be used for classification prediction of new cluster members.

[0087] The target clustering collective is a collective constructed by high-quality clustering members. When processing and identifying the clustering member set based on the decision tree model, the ARI, NMI and F-measure index in the characteristic attribute set of each clustering member are output, and the decision tree model will predict its clustering quality label ("high" or "low") according to the learned node division rule. The clustering members labeled "high" are formed into a new cluster collective P'* to participate in subsequent processing, that is, the target clustering collective is generated.

[0088] Among them, the target CA matrix of the target clustering group is determined according to the cluster layer weight coefficient, including:

[0089] Construct the CA matrix of the target cluster collective;

[0090] The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B;

[0091] Capture high confidence information from the CA matrix to obtain the HC matrix;

[0092] The target CA matrix of the target cluster group is determined based on the processed data B and the HC matrix.

[0093] Construct the CA matrix of the target cluster collective, including:

[0094]

[0095]

[0096] Where A is the CA matrix; Indicates cluster members; Represents the total number of cluster members in the cluster collective; Represents sample points The cluster in which it is located; Represents sample points The cluster where it is located.

[0097] The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B, including:

[0098]

[0099]

[0100]

[0101]

[0102] Capture high-confidence information from the CA matrix, and use the captured high-confidence information to supplement and improve the ideal CA matrix. The high-confidence information HC (High-Confidence Matrix) matrix is ​​denoted as H. The method is as follows:

[0103]

[0104]

[0105] in, Record the location of high-reliability elements, is an element operator. When the ratio of the number of times two sample points are classified into the same cluster to the total number of cluster members exceeds a predefined threshold , then the corresponding position of the A matrix is ​​regarded as a highly reliable information (i.e., an element in the H matrix).

[0106] The target CA matrix of the target clustering group is determined according to the processed data B and the HC matrix, including:

[0107] The final CA matrix, denoted as C, is obtained from the H matrix and the B matrix as follows:

[0108]

[0109]

[0110] in, is the Laplacian matrix, , Represents the Lagrange term, the default value is 1, Used to balance the error loss term, and represents the Lagrange multiplier, represents the Frobenius norm of the matrix, E represents the error term, and F is an intermediate matrix used to alleviate the value range constraints and symmetry constraints of the ideal CA matrix.

[0111] The information entropy is introduced to the target cluster collectively, and the uncertainty index IEI of each cluster is calculated as the cluster layer weight coefficient. The calculation method of IEI is as follows:

[0112]

[0113]

[0114]

[0115] in, Indicates the total number of clusters; To measure clusters With cluster The IEI index reflects the similarity between clusters. The possibility that the points in the cluster are still classified in the same cluster in other base clusters. The larger the IEI, the smaller the cluster. The more likely the points in the cluster are to be classified into the same cluster in other base clusters; capture high-confidence information from the CA matrix, use the captured high-confidence information to supplement and improve the ideal CA matrix, and obtain the HC (High-Confidence Matrix) matrix with high-confidence information, denoted as H; is the maximum value in the H matrix; is the minimum value in the H matrix; is the cluster in the H matrix The corresponding value;

[0116] Or introduce KL divergence to determine the cluster layer weight coefficient;

[0117] Or introduce multiple evaluation indicators and combine them to form a new evaluation indicator to determine the cluster layer weight coefficient.

[0118] This application introduces a decision tree to assist in selecting high-quality cluster members. By training a decision tree model to assist in selecting cluster members, the quality and diversity of cluster members are considered from multiple perspectives, while considering their internal diversity, and more comprehensively achieving differentiated treatment of cluster members. The diversity of the cluster layer of cluster members is then considered, the uncertainty of the cluster is measured, and the weights are assigned to the cluster layer; finally, the CA matrix is ​​fine-tuned according to the weights and high-confidence information to improve the accuracy and robustness of clustering.

[0119] The cluster weighted clustering integrated medical data processing method based on member selection of the present application uses a clustering algorithm to cluster patient data and generate multiple cluster members. Each cluster member divides the patient into different clusters, and each cluster represents a patient group with similar symptoms and characteristics. Then, high-quality cluster members are selected with the assistance of a decision tree model. In order to further improve the accuracy and interpretability of the clustering results, attention is paid to the internal diversity of the cluster members, the diversity of the internal clusters is measured, and the CA matrix is ​​fine-tuned. Finally, the final clustering results are obtained using hierarchical cluster analysis. Through this method, we obtain more optimized clustering results, which can better show the characteristics and similarities of patient diseases, facilitate doctors to choose medical services or customize personalized treatment plans for different patients, and help improve patients' medical experience and treatment effects.

[0120] In one embodiment, the processing method also includes: evaluating the final clustering result based on external indicators, determining an evaluation value, and judging the effectiveness of the final clustering result based on the evaluation value. Clustering result effectiveness evaluation indicators are generally divided into internal indicators and external indicators. In most cases, the category labels of the data set are known (not used in the clustering process). At this time, external indicators can be used to evaluate the effectiveness of clustering, among which the F value (F-measure) is a relatively commonly used comprehensive indicator for evaluating the quality of text clustering. The larger the F value, the higher the clustering quality. When the clustering result is completely consistent with the true category, the F value reaches a maximum value of 1. In addition, the NMI value is also a relatively popular clustering result effectiveness evaluation indicator, which can quantify the degree of match between the clustering result and the true text category label.

[0121] In one embodiment, the Gini coefficients of the feature attribute set regarding the ARI, NMI and F-measure index are calculated, including:

[0122]

[0123]

[0124] in, Represents the Gini impurity of the data set D; Indicates that the category label is The proportion in data set D, that is, belonging to category The number of samples divided by the total number of samples; Indicated in the feature attribute Under the condition of , the Gini impurity of data set D; Indicated in the feature attribute The value of The number of samples in the data subset obtained when ; Representing a subset of data Based on the above formula, the Gini coefficients of the labeled cluster members in terms of ARI, NMI, and F-measure index are calculated to accurately calculate the corresponding Gini coefficients, improve the accuracy of judging the size of the Gini coefficients of the three, and then achieve accurate division of the root node and internal nodes.

[0125] In one embodiment, the method for processing medical data by cluster weighted clustering integration based on member selection further processes the medical data as follows before constructing the cluster member set to mark out data that interferes with the clustering result. The specific processing steps are as follows:

[0126] Analyze the feedback data corresponding to the clustering results and determine the marked nodes;

[0127] When the number of marked nodes is greater than or equal to a preset number (e.g., any value from 2 to 100), the marked nodes and related data are output, as well as the training data of the decision tree model; after receiving the re-marking from professionals, the decision tree model is retrained using the marked data;

[0128] When the number of marked nodes is less than a preset number, the medical data is screened and marked based on each marked node;

[0129] When the clustering results are output, the marked medical data synchronously outputs the corresponding marking information.

[0130] The steps for obtaining feedback data are as follows:

[0131] After the clustering results of the patients are sent to the preset physician terminal, the annotation information of the physician terminal is received;

[0132] and / or,

[0133] After the patient's clustering results and the patient's treatment plan are sent to the patient terminal, when the patient receives a correction instruction for doubts about the received data, the clustering results, treatment plan and doubtful information are sent to the preset professional physician terminal, and the annotation information of the professional physician terminal is received;

[0134] Annotation information, clustering results, treatment plans and / or questionable information are used as feedback data.

[0135] The analysis steps for feedback data are as follows:

[0136] Screening the feedback data; specifically: based on the preset keyword library, extracting keywords from the annotation information and questionable information in the feedback data, obtaining the keyword set corresponding to each feedback data, and deleting the feedback data without the extracted keywords; because some feedback data contain both annotation information and questionable information, the keywords extracted from the two may interfere with each other, so according to the preset keyword interference association library, the keywords in the keyword set may be monitored for interference, and the feedback data corresponding to the monitored interference keyword set may be deleted;

[0137] Obtaining the original medical data corresponding to the clustering results corresponding to the filtered feedback data and the characteristic data during classification;

[0138] Associating the original medical data, the classification feature data and the feedback data to form the data to be analyzed;

[0139] The data to be analyzed are grouped according to the original medical data and the classification feature data; each group corresponds to a marked node; the specific grouping can be performed by calculating the data similarity between the original medical data and the classification feature data, and when the sum of the data similarities is greater than a preset threshold, it can be determined that the two data to be analyzed are the same group;

[0140] According to the sum of similarities between each data to be analyzed in the group and the original medical data and classification feature data of other data to be analyzed, a feature set constructed by the original medical data and classification feature data corresponding to the data to be analyzed having the largest sum is determined as the feature set of the marking node;

[0141] The composition and content of the data to be analyzed are scored based on the preset scoring model, and the key data extracted from the labeled data in the data to be analyzed with the largest score value is used as the labeled data; the more components of the data to be analyzed, the higher the score; the larger the total amount of data in the data to be analyzed, the higher the score; the larger the total amount of key data extracted, the higher the score; the specific scoring rules are:

[0142] The data to be analyzed are intercepted according to their sources;

[0143] According to the feature extraction rules corresponding to the source of each data, feature extraction is performed on the intercepted data; the features include parameter feature values ​​corresponding to the total amount of data, parameter feature values ​​corresponding to the keywords contained in the preset feature library in the intercepted data, etc.;

[0144] Determine the aspect score value of each intercepted data according to the scoring rules corresponding to each data source and the data features corresponding to each intercepted data;

[0145] The weights corresponding to the scoring values ​​of each aspect are determined based on the source of each data; the weight corresponding to the classification feature data is a fixed value and is preset; the weight corresponding to the annotation data corresponds to the authority value of the physician terminal or professional physician terminal that provides the data; the weight corresponding to the medical data corresponds to the authority value of the terminal that uploaded the medical data;

[0146] The score is calculated based on the weights and the score values ​​of each aspect. The score is the sum of the product of the aspect score and the corresponding weight.

[0147] This embodiment pre-marks the data by marking nodes to assist in the clustering and classification of medical data, which can better display the characteristics and similarities of patients' diseases, facilitate doctors to select medical services or customize personalized treatment plans for different patients, and help improve patients' medical experience and treatment effects.

[0148] In view of the on-site application scenario of medical data processing, in one embodiment, the construction method of medical data in the cluster weighted clustering integrated medical data processing method based on member selection is as follows:

[0149] Acquire the first data uploaded by each associated first terminal within a preset legal upload time interval; the legal upload time interval can be manually configured, for example: when the first terminal is a computer used in outpatient clinics of each hospital, the legal upload time interval can be configured as half an hour before the doctor's work to half an hour after get off work;

[0150] Determining whether the data volume of the first data meets the preset first data requirement;

[0151] If it is not reached, the standard data stored in the system will be used to complete it;

[0152] The first data requirement includes: data of a preset number of patients included in the medical data, for example: 100, 200, 300, ..., 1000.

[0153] Among them, the association rule of the first terminal is that the first terminals of different hospitals in the same department are associated as one group. This association rule is used to facilitate clustering;

[0154] In addition, the legal upload time can be set in segments, with each half hour as a segment, and one legal time corresponds to one analysis, to facilitate the outpatient doctors' needs for data processing.

[0155] In the scenario of post-analysis and secondary analysis of medical data, in one embodiment, the construction method of medical data in the cluster weighted clustering integrated medical data processing method based on member selection is as follows:

[0156] Acquire the second data uploaded by the second terminal; the second terminal is a terminal device registered in the system by each hospital (for example, a computer set in an analysis room, etc.);

[0157] Determine whether other second terminals upload data; when uploading, integrate all second data to determine whether the preset second data requirements are met, otherwise directly determine whether the second data meets the preset second requirements;

[0158] When the preset second data requirement is not met, a data list is output according to the historical uploaded data of the second terminal, and a selection of data on the data list by the second terminal is received, so as to complete the second data to meet the preset second data requirement, wherein the second data requirement includes: data of a preset number of patients included in the medical data, for example: 100, 200, 300..., 1000.

[0159] By selecting historical uploaded data, on the one hand, the data can meet the requirements of clustering processing, and on the other hand, cluster analysis can be performed in combination with historical data, which can be used to verify and analyze historical data.

[0160] The present invention also provides a cluster weighted clustering integrated medical data processing system based on member selection, such as Figure 6 As shown, it includes: a construction unit 1, an input unit 2, a screening unit 3, a matrix determination unit 4 and an execution unit 5;

[0161] Among them, the construction unit 1 constructs a cluster member set; the input unit 2 inputs the cluster member set into a pre-trained decision tree model; the screening unit 3 screens out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generates a target cluster group with the screened cluster members; the matrix determination unit 4 determines the target CA matrix of the target cluster group according to the cluster layer weighting coefficient; the execution unit 5 executes the hierarchical clustering algorithm based on the target CA matrix to obtain the final clustering result.

[0162] Construction unit 1 uses K-Means algorithm to cluster medical data and generate multiple cluster members.

[0163] The training steps of the decision tree model are as follows:

[0164] Calculate the Davies-Bouldin index of each sample cluster member in the sample cluster member set and find the overall average value;

[0165] Compare the Davies-Bouldin index of each sample cluster member with the average value, label the sample cluster members whose Davies-Bouldin index is lower than the average value with the first label, and label the sample cluster members whose Davies-Bouldin index is higher than the average value with the second label;

[0166] The training is performed based on the sample cluster members with labels as the training set to obtain a trained decision tree model.

[0167] The matrix determination unit 4 performs the following operations:

[0168] Construct the CA matrix of the target cluster collective;

[0169] The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B;

[0170] Capture high confidence information from the CA matrix to obtain the HC matrix;

[0171] The target CA matrix of the target cluster group is determined based on the processed data B and the HC matrix.

[0172] The matrix determination unit 4 constructs a CA matrix about the target clustering group and performs the following operations:

[0173]

[0174]

[0175] Where A is the CA matrix; Indicates cluster members; Represents the total number of cluster members in the cluster collective; Represents sample points The cluster in which it is located; Represents sample points The cluster where it is located.

[0176] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A cluster weighted clustering ensemble medical text processing method based on member selection, characterized in that include: Construct a set of cluster members and input them into a pre-trained decision tree model; Filter out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generate a target cluster group with the filtered cluster members; Determine the target CA matrix of the target clustering group according to the cluster layer weight coefficient; Execute the hierarchical clustering algorithm based on the target CA matrix to obtain the final clustering result; Evaluate the final clustering results based on external indicators, determine the evaluation value, and judge the effectiveness of the final clustering results based on the evaluation value; Before constructing the cluster member set, the medical text data is processed as follows to mark out the data that interferes with the clustering results. The specific processing steps are as follows: Analyze the feedback data corresponding to the clustering results and determine the marked nodes; When the number of marked nodes is greater than or equal to the preset number, the marked nodes and related data are output, and the training data of the decision tree model is output; After receiving the re-labeling from professionals, the decision tree model is retrained using the labeled data; When the number of marked nodes is less than a preset number, the medical text data is screened and marked based on each marked node; When outputting the clustering results, the marked medical text data synchronously outputs the corresponding marking information; The steps for obtaining feedback data are as follows: After the clustering results of the patients are sent to the preset physician terminal, the annotation information of the physician terminal is received; and / or, After the patient's clustering results and the patient's treatment plan are sent to the patient terminal, when the patient receives a correction instruction for doubts about the received data, the clustering results, treatment plan and doubtful information are sent to the preset professional physician terminal, and the annotation information of the professional physician terminal is received; Using annotation information, clustering results, treatment plans and / or questionable information as feedback data; The analysis steps for feedback data are as follows: Screening feedback data; Obtaining the original medical text data corresponding to the clustering results corresponding to the filtered feedback data and the feature data during classification; Associating the original medical text data, classification feature data and feedback data to form data to be analyzed; The data to be analyzed are grouped according to the original medical text data and classification feature data; According to the sum of similarities between each data to be analyzed in the group and the original medical text data and classification feature data of other data to be analyzed, a feature set constructed by the original medical text data and classification feature data corresponding to the data to be analyzed having the largest sum is determined as the feature set of the marking node; Scoring the composition and content of the data to be analyzed based on a preset scoring model, using the key data extracted from the labeled data of the data to be analyzed with the largest scoring value as the labeled data; The data to be analyzed are intercepted according to their sources; According to the feature extraction rules corresponding to each data source, feature extraction is performed on the intercepted data; Determine the aspect score value of each intercepted data according to the scoring rules corresponding to each data source and the data features corresponding to each intercepted data; Determine the weights corresponding to the scores of each aspect based on the data sources; Calculate the score based on the weights and scores of each aspect; In most cases, the category labels of medical text data are known. In this case, external indicators are used to evaluate the effectiveness of clustering. The F value is a comprehensive indicator for evaluating the quality of clustering of medical text data. The larger the F value, the higher the clustering quality. When the clustering result is completely consistent with the category of the medical text data, the F value reaches its maximum value, which is 1. The NMI value is an evaluation indicator of the effectiveness of the clustering result, which quantifies the degree of match between the clustering result and the category label of the real medical text data.

2. The method for processing medical texts by cluster weighted clustering integration based on member selection according to claim 1, characterized in that The construction of the cluster member set includes: clustering the medical text data using the K-Means algorithm to generate multiple cluster members.

3. The method for processing medical texts by cluster weighted clustering integration based on member selection according to claim 1, characterized in that The training steps of the decision tree model are as follows: Calculate the Davies-Bouldin index of each sample cluster member in the sample cluster member set and find the overall average value; Compare the Davies-Bouldin index of each sample cluster member with the average value, label the sample cluster members whose Davies-Bouldin index is lower than the average value with the first label, and label the sample cluster members whose Davies-Bouldin index is higher than the average value with the second label; The training is performed based on the sample cluster members with labels as the training set to obtain a trained decision tree model.

4. The method for processing medical texts by cluster weighted clustering integration based on member selection according to claim 1, characterized in that The target CA matrix of the target clustering group is determined according to the cluster layer weight coefficient, including: Construct the CA matrix of the target cluster collective; The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B; Capture high confidence information from the CA matrix to obtain the HC matrix; The target CA matrix of the target cluster group is determined based on the processed data B and the HC matrix.

5. The method for processing medical texts by cluster weighted clustering integration based on member selection according to claim 4, characterized in that Construct the CA matrix of the target cluster collective, including: Where A is the CA matrix; Indicates cluster members; Represents the total number of cluster members in the cluster collective; Represents sample points The cluster in which it is located; Represents sample points The cluster where it is located.

6. A cluster weighted clustering integrated medical text processing system based on member selection, characterized by include: A construction unit, an input unit, a screening unit, a matrix determination unit and an execution unit; Among them, the construction unit constructs a cluster member set; the input unit inputs the cluster member set into a pre-trained decision tree model; the screening unit screens out cluster members with pre-labeled labels from the cluster member set output by the decision tree model, and generates a target cluster group with the screened cluster members; the matrix determination unit determines the target CA matrix of the target cluster group according to the cluster layer weight coefficient; the execution unit executes the hierarchical clustering algorithm based on the target CA matrix to obtain the final clustering result; Before constructing the cluster member set, the medical text data is processed as follows to mark out the data that interferes with the clustering results. The specific processing steps are as follows: Analyze the feedback data corresponding to the clustering results and determine the marked nodes; When the number of marked nodes is greater than or equal to the preset number, the marked nodes and related data are output, and the training data of the decision tree model is output; after receiving the re-marking from professionals, the decision tree model is retrained using the marked data; When the number of marked nodes is less than a preset number, the medical text data is screened and marked based on each marked node; When outputting the clustering results, the marked medical text data synchronously outputs the corresponding marking information; The steps for obtaining feedback data are as follows: After the clustering results of the patients are sent to the preset physician terminal, the annotation information of the physician terminal is received; and / or, After the patient's clustering results and the patient's treatment plan are sent to the patient terminal, when the patient receives a correction instruction for doubts about the received data, the clustering results, treatment plan and doubtful information are sent to the preset professional physician terminal, and the annotation information of the professional physician terminal is received; Using annotation information, clustering results, treatment plans and / or questionable information as feedback data; The analysis steps for feedback data are as follows: Screening feedback data; Obtaining the original medical text data corresponding to the clustering results corresponding to the filtered feedback data and the feature data during classification; Associating the original medical text data, classification feature data and feedback data to form data to be analyzed; The data to be analyzed are grouped according to the original medical text data and classification feature data; According to the sum of similarities between each data to be analyzed in the group and the original medical text data and classification feature data of other data to be analyzed, a feature set constructed by the original medical text data and classification feature data corresponding to the data to be analyzed having the largest sum is determined as the feature set of the marking node; Scoring the composition and content of the data to be analyzed based on a preset scoring model, using the key data extracted from the labeled data of the data to be analyzed with the largest scoring value as the labeled data; The data to be analyzed are intercepted according to their sources; According to the feature extraction rules corresponding to each data source, feature extraction is performed on the intercepted data; Determine the aspect score value of each intercepted data according to the scoring rules corresponding to each data source and the data features corresponding to each intercepted data; Determine the weights corresponding to the scores of each aspect based on the data sources; Calculate the score based on the weights and scores of each aspect; The execution unit also evaluates the final clustering result based on the external indicator, determines the evaluation value, and judges the validity of the final clustering result based on the evaluation value; In most cases, the category labels of medical text data are known. In this case, external indicators are used to evaluate the effectiveness of clustering. The F value is a comprehensive indicator for evaluating the quality of clustering of medical text data. The larger the F value, the higher the clustering quality. When the clustering result is completely consistent with the category of the medical text data, the F value reaches its maximum value, which is 1. The NMI value is an evaluation indicator of the effectiveness of the clustering result, which quantifies the degree of match between the clustering result and the category label of the real medical text data.

7. The cluster weighted clustering integrated medical text processing system based on member selection according to claim 6 is characterized in that The construction unit uses the K-Means algorithm to cluster the medical text data and generate multiple cluster members.

8. The cluster weighted clustering integrated medical text processing system based on member selection according to claim 6 is characterized in that The training steps of the decision tree model are as follows: Calculate the Davies-Bouldin index of each sample cluster member in the sample cluster member set and find the overall average value; Compare the Davies-Bouldin index of each sample cluster member with the average value, label the sample cluster members whose Davies-Bouldin index is lower than the average value with the first label, and label the sample cluster members whose Davies-Bouldin index is higher than the average value with the second label; The training is performed based on the sample cluster members with labels as the training set to obtain a trained decision tree model.

9. The cluster weighted clustering integrated medical text processing system based on member selection according to claim 6, characterized in that The matrix determination unit performs the following operations: Construct the CA matrix of the target cluster collective; The CA matrix is ​​weighted based on the cluster layer weight coefficient to obtain the processed data B; Capture high confidence information from the CA matrix to obtain the HC matrix; The target CA matrix of the target cluster group is determined based on the processed data B and the HC matrix.

10. The cluster weighted clustering integrated medical text processing system based on member selection according to claim 9, characterized in that The matrix determination unit constructs a CA matrix about the target cluster collective, including: Where A is the CA matrix; Indicates cluster members; Represents the total number of cluster members in the cluster set; Represents sample points The cluster in which it is located; Represents sample points The cluster where it is located.

Images