A method and device for preprocessing complex multi-label medical data containing mutual relationships
By preprocessing medical data, including deleting missing values, discrete features and clustering tags, and selecting them based on the relationship between features and tags, the problem of inaccurate feature selection in the prior art is solved, and the prediction accuracy and calculation efficiency of deep learning are improved.
Patent Information
- Application Number
- CN202310603032.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2043-05-25
AI Technical Summary
When processing medical data, existing deep learning methods cannot effectively process the relationship between labels and labels and features, resulting in inaccurate feature selection and affect prediction accuracy and computing efficiency.
By deleting features and irrelevant terms with missing values greater than 99% in the data, using the minimum entropy method to discrete numerical features, clustering the tags and generating representative tags, combining the synergy, correlation, redundancy and dynamic changes between the features and tags, and using interactive information to select features.
It reduces the redundant information of labels and features, improves the computing efficiency of feature selection, and improves the prediction accuracy and training time of deep learning.
Smart Images
Figure CN116842374B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of data processing, and in particular relates to a preprocessing method and device for complex multi-label medical data containing mutual relationships. Background Art
[0002] There are many deep learning methods that can be used to train medical data. By discovering the potential connections between diseases and symptoms from the data, the causes of diseases, the possibility of onset, prevention methods and treatment measures can be obtained.
[0003] However, there is currently little research on the preprocessing part before deep learning. A better preprocessing method can discover features that are irrelevant to the predicted content, have little correlation, or are redundant with each other, and delete these features before deep learning. This can not only improve computing efficiency, but also improve the accuracy of prediction to a certain extent.
[0004] The Relief method determines the relevance of features and labels based on the differences between the nearest neighboring samples. If the features of neighboring samples of the same class differ, the feature weight is reduced. If the features of neighboring samples of different classes differ, the feature weight is increased. Reyes et al. constructed a multi-label ReliefF feature selection method by modifying prior probability estimates. They considered that the nearest instances belonging to different label sets have different feature values when constructing a multi-label ReliefF feature selection method. et al. used a Hamming distance-based dissimilarity function to construct a multi-label ReliefF feature selection method. Lee et al. introduced mutual information into multi-label feature selection and proposed a multi-label feature selection method called PMU. This algorithm uses multivariate mutual information to calculate the correlation between features and labels and selects features incrementally. Lin et al. introduced the maximum minimum redundancy (MAXM) method from single-label feature selection into multi-label feature selection and proposed a multi-label MAXM version of ML-MDMR. Lee et al. proposed a fast multi-label feature selection method called FIFM, which modifies the original feature selection process to accelerate feature selection. Li et al. proposed a multi-label feature selection method called IGMF. Lee et al. proposed a multi-label feature selection method called D2F. Lee et al. proposed a multi-label feature selection method called SCLS. Miao proposed a granular multi-label feature selection method called GMLF to address the problem of label correlation. This method uses information granules to cluster related labels into one granule and uses averaging to reduce the influence of label correlation, resulting in a more appropriate feature subset and improving classification performance. In 2019, Zhang proposed a multi-label feature selection algorithm called LRFS. LRFS introduces the concept of label redundancy and provides a solution. This method uses mutual information to calculate the redundancy between labels, thereby reducing the impact of label redundancy on feature selection and selecting a more appropriate feature subset.
[0005] However, existing methods cannot handle the relationship between features and labels well when processing data where there is a relationship between them, resulting in the inability to select a more appropriate feature subset, making it impossible for deep learning methods to achieve higher accuracy. Summary of the Invention
[0006] In order to solve the interrelationships between labels and features in medical multi-label data, the present invention proposes a preprocessing method for complex multi-label medical data containing interrelationships, which specifically includes the following steps:
[0007] Delete features with missing values greater than 99% in the data, delete the collection items in the data set that are not related to the prediction, and fill the vacancies in the data set with 0 values;
[0008] The minimum entropy method is used to discretize the numerical features in the data set;
[0009] Cluster the labels of the data set, take the label of each cluster close to the cluster center as the representative label of the cluster, and update the labels of all data in each cluster to the corresponding representative label;
[0010] Select data features based on the synergy, correlation, redundancy, and dynamic changes between data features and labels;
[0011] The labels of the processed data and the selected features are used as the preprocessed data to complete the preprocessing.
[0012] Furthermore, the process of obtaining representative labels of data through clustering includes the following steps:
[0013] The label space L of the dataset is L={l1,l2,...,l m} Granulation is performed to obtain k information granules;
[0014] Each label l i are assigned to their nearest representative label c in turn i Thus, a new label information particle is formed. For each new label information particle G i Regenerate representative label c i .
[0015] Further, generate representative label c i The process includes:
[0016]
[0017] Among them, S(G i ,l s ) represents an information particle G i Its representative label l s distance.
[0018] Furthermore, an information particle G i Its representative label l s The distance S(G i ,l s ) is expressed as:
[0019]
[0020]
[0021] Among them, H(l j ) indicates label l j The information entropy of p(a) represents the label l j The probability that the value of is a; MI(l j ;c i ) indicates label l j With center label c i The mutual information of p(a,b) represents the label l j The value is a, the center label ci The joint probability of taking value b; NMI(l j ,c i ) indicates label l j With center label c i The label mutual information of
[0022] Furthermore, the process of selecting data features based on the synergy, correlation, redundancy, and dynamic changes between data features and labels includes:
[0023]
[0024] Among them, J(x i ) is the feature selection function, the candidate feature x i The larger the value of the corresponding feature selection function, the higher the candidate feature x i The higher the possibility of being selected; MI(x i ; C|S) represents feature correlation; MI(S; C|x i ) represents the dynamic feature correlation; MI(x i ; S|C) represents feature redundancy; MI(x i ; S; C) represents the synergy of features; C represents the label set, c k represents the kth label in the label set; S is the selected feature set.
[0025] Furthermore, the feature synergy MI(x i ; S; C) calculation includes:
[0026]
[0027]
[0028] Among them, MI(x i ;S;c k ) is the feature synergy of single-label feature selection; MI(x i ;x j ;c k ) indicates that in the candidate feature x i and selected features x j In label c k The mutual information under i ;x j ;c k )=MI(x i ;x j )-MI(x i ;x j |c k ), p(a) represents the label l j The probability of taking the value a, p(a, b) represents the candidate feature x i The value of is a, the selected feature x j The joint probability of taking the value b; p(a, b, c) represents the candidate feature x i The value of is a, the selected feature x j The value is b, the center label is c k is the joint probability of c; p(a, b|c) represents the probability of the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b, p(a|c) represents the probability of taking the value b at the center label c k The candidate feature x is the premise of c i The probability that the value of is a.
[0029] Furthermore, the feature correlation MI(x i ; C|S) calculation includes:
[0030]
[0031]
[0032] Among them, p(a,b,c) represents the candidate feature x i The value of is a, the selected feature x j The value is b, the center label is c k is the joint probability of c; p(a|b) is the selected feature x j The candidate feature x is taken as b i The probability of taking the value a; p(c|b) is the selected feature x j The central label c is assumed to be b. k is the probability of c; p(a,c|b) represents the selected feature x j The candidate feature x is taken as b i The value of is a, the selected feature x j The probability of taking the value b.
[0033] Furthermore, feature redundancy MI(x i ; The calculation of S|C) includes:
[0034]
[0035]
[0036] Among them, p(a,b,c) represents the candidate feature x i The value of is a, the selected feature xj The value is b, the center label is c k is the joint probability of c; p(a|c) represents the probability of the center label c k The candidate feature x is the premise of c i The probability of taking the value of a; p(b|c) represents the probability of taking the value of a at the center label c k The selected feature x is c j The probability of taking the value b; p(a,b|c) represents the probability of taking the value b at the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b.
[0037] Furthermore, the dynamic correlation of features MI(S;C|x i ) calculation includes:
[0038] Among them, p(a,b,c) represents the candidate feature x i The value of is a, the selected feature x j The value is b, the center label is c k is the joint probability of c; p(b|a) represents the probability of candidate feature x i The value of a is the selected feature x j The probability of taking the value b; p(c|a) represents the probability of taking the value b in the candidate feature x i The value of is a, the central label c k is the probability of taking value c; p(b,c|a) represents the probability of taking value c in candidate feature x i The feature x is selected under the premise that the value of is a j The value is b, the center label is c k is the probability of taking the value c.
[0039] The present invention also proposes a preprocessing device for complex multi-label medical data containing mutual relationships, which is used to implement a preprocessing method for complex multi-label medical data containing mutual relationships, including a data cleaning module, a data discretization module, a representative label generation module, a feature selection module, and a data output module, wherein:
[0040] The data cleaning module is used to delete features with missing values greater than 99% in the data, delete the collection items in the data set that are not related to the prediction, and fill the vacancies in the data set with 0 values;
[0041] The data discretization module is used to discretize the numerical features in the data set using the minimum entropy method;
[0042] The representative label generation module is used to cluster the labels of the data set, take the label of each cluster close to the cluster center as the representative label of the cluster, and update the labels of all data in each cluster to the corresponding representative label;
[0043] Feature selection module, used to select data features based on the synergy, correlation, redundancy and dynamic changes between data features and labels;
[0044] The data output module is used to use the labels of the processed data and the selected features as the preprocessed data to complete the preprocessing.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] 1. The model uses label granulation to identify relationships between labels, resulting in a set of label subsets that contain less redundant information while minimizing the loss of useful information in the labels. Providing a better label subset for feature selection can reduce feature selection computation time and avoid inaccurate feature subset selection due to redundant information in the labels.
[0047] 2. In feature selection, the model incorporates interactive information into the evaluation function of feature selection, so that the feature selection method can discover the interaction between features. Two features that can generate interactive information can bring more information than two single features, so that the selected feature subset can contain more information.
[0048] 3. The feature subset obtained by this method significantly shortens the training time of neural network training compared to neural network training without feature selection. Compared with feature subsets obtained by other feature selection methods, the prediction accuracy is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 A flowchart of a method for preprocessing complex multi-label medical data containing interrelationships according to the present invention;
[0050] Figure 2 A schematic diagram of the process of obtaining a representative tag in the present invention;
[0051] Figure 3 Schematic diagram of the process of selecting features in the present invention. DETAILED DESCRIPTION
[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0053] The present invention proposes a preprocessing method for complex multi-label medical data containing mutual relationships, which specifically includes the following steps:
[0054] Delete features with missing values greater than 99% in the data, delete the collection items in the data set that are not related to the prediction, and fill the vacancies in the data set with 0 values;
[0055] The minimum entropy method is used to discretize the numerical features in the data set;
[0056] Cluster the labels of the data set, take the label of each cluster close to the cluster center as the representative label of the cluster, and update the labels of all data in each cluster to the corresponding representative label;
[0057] Select data features based on the synergy, correlation, redundancy, and dynamic changes between data features and labels;
[0058] The labels of the processed data and the selected features are used as the preprocessed data to complete the preprocessing.
[0059] In this embodiment, a dynamic multi-label feature selection method combining granular labels with interactive information is used to preprocess multi-label medical data containing mutual relationships. The process is as follows: Figure 1 As shown, the method includes the following steps:
[0060] Step 1: Data filling. Delete features with more than 99% missing values in the data, delete items not related to prediction, and fill the gaps in the data set with 0 values.
[0061] Step 2: Data normalization. Use the minimum entropy method to discretize the numerical features in the dataset. Convert all features into discrete values for easier processing.
[0062] Step 3: Label Granulation. After standardization, the interrelationships between labels are processed. The labels in the dataset are processed according to the process of Algorithm 1. First, the degree of granulation is selected. Then, all labels are clustered into a specified number of information granules. The label closest to the cluster center is used as the label for subsequent feature selection. The central label of each information granule is combined into an information label set. This reduces the impact of interrelationships between labels.
[0063] Step 4: Multi-label feature selection based on mutual information. Use Algorithm 2 to find features related to the processed label set. This algorithm can discover the relationships between features. Each time, select the feature with the highest score as the selected feature until the specified number of features is selected.
[0064] Algorithm 1 uses label granulation to find relationships between labels. It clusters similar labels into a single category, then uses the label closest to the cluster center as the representative label for that category. This replaces the representative labels of all clusters and forms the new label set for subsequent calculations. Label granulation reduces redundant information between labels, resulting in a more concise label set for feature selection.
[0065] Algorithm 2 incorporates interaction information and feature dynamics into feature selection to identify interrelationships between features. By identifying correlations and redundancies between features, it selects features that are most relevant to the label and least redundant with each other. This selected feature set enables subsequent deep learning to achieve better results while using fewer computational resources.
[0066] Algorithm 1. Label granulation
[0067] The label granulation algorithm converts the label space L={l1,l2,...,l m} is granulated into k information particles, each of which is G i , the representative label of the information particle is c i The similarity between an information particle and its representative label is S(G i ,c i ), similarity S(G i ,c i ) is defined as:
[0068]
[0069] Among them, D(l j ,c i ) The distance between the label contained in the information particle and its central label. The labels contained in each information particle do not intersect with each other, that is, The union of all information particles is the label space, that is,
[0070] Each label l i are assigned to their nearest representative label c in turn i Thus, a new label information particle is formed. For each new label information particle G i Regenerate representative label c i , and c i ∈G i , representing label c i The update method is defined as:
[0071]
[0072] Among them, l s ∈G i , calculate the similarity between each label in the information granule and the information granule, and select the label with the smallest similarity as the new representative label c i . Distance (l j ,c i ) is calculated as:
[0073]
[0074] Among them, H(l j ) indicates label l j The information entropy of p(a) represents the label l j The probability that the value of is a; MI(l j ;c i ) indicates label l j With center label c i The mutual information of p(a,b) represents the label l j The value is a, the center label c i The joint probability of taking value b; NMI(l j ,c i ) indicates label l j With center label c i The label mutual information of
[0075] This embodiment also proposes a process for label granulation, such as Figure 2 , specifically including the following steps:
[0076] The label space L of the dataset is L={l1,l2,...,l m} Granulation is performed to obtain k information granules;
[0077] Randomly select k labels as cluster centers for clustering operation;
[0078] Use the center label of each cluster as the new cluster center until the cluster center no longer changes;
[0079] Output the label corresponding to the current cluster center as the representative label of the current cluster.
[0080] Algorithm 2. Multi-label feature selection combined with mutual information
[0081] (1) The relationship between features—synergy
[0082] Synergy means that the information in a single feature is irrelevant to the label information, but when used in combination with other features, some information irrelevant to the label will be transformed into information useful for classification. The synergy between features is calculated by using the mutual information. The synergy of features MI(x i ; S; C) is defined as:
[0083]
[0084]
[0085] Among them, MI(x i ;S;c k ) The synergy of the features is based on the feature synergy used for single-label feature selection in DFS-SGII by Wu et al.; MI(x i ;x j ;c k ) indicates that in the candidate feature x i and selected features x j In label c k When the interaction between a feature and the selected feature can provide more information for classification, the feature should be selected. MI(x i ;x j ;c k ) is expressed as:
[0086] MI(x i ;x j ;c k )=MI(x i ;x j )-MI(x i ;x j |c k );
[0087]
[0088]
[0089] Among them, p(a,b,c) represents the candidate feature x i The value of is a, the selected feature x j The value is b, the center label is c k is the joint probability of c; p(a,b|c) represents the probability of the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b, p(a|c) represents the probability of taking the value b at the center label c k The candidate feature x is the premise of c i The probability that the value of is a.
[0090] (2) The relationship between features and labels - correlation
[0091] Feature correlation is the degree of association between the feature and the label. Mutual information MI(x i ; C|S) is the condition of the selected feature set S, the feature to be selected x i Correlation with the representative label set C. MI(x i ; C|S) is defined as:
[0092]
[0093]
[0094] Among them, x j is the selected feature, c k is the representative label, p(a,b|c) represents the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b; p(a|b) is the selected feature x j The candidate feature x is taken as b i The probability of taking the value a; p(c|b) is the selected feature x j The central label c is assumed to be b. k is the probability of c; p(a,c|b) represents the selected feature x j The candidate feature x is taken as b i The value of is a, the selected feature x j The probability of taking the value b.
[0095] (3) The relationship between features - redundancy
[0096] The redundancy of features means that the information they contain can be inferred from other features. i ; S|C) represents the condition that the selected feature x represents the label set C. i The correlation with the selected label set S. Since the selected labels already contain this part of information, it is redundant information for the feature selection system. i ; S|C) is defined as:
[0097]
[0098]
[0099] Among them, p(a|c) represents the center label c k The candidate feature x is the premise of ci The probability of taking the value of a; p(b|c) represents the probability of taking the value of a at the center label c k The selected feature x is c j The probability of taking the value b; p(a,b|c) represents the probability of taking the value b at the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b.
[0100] (4) The relationship between features—dynamic changes
[0101] Dynamic feature relevance means that the amount of information between the selected feature and the label changes as candidate features are considered. While the mutual information between the selected feature and the class remains constant, this information varies with each candidate. Newly selected features should minimize the amount of information provided by existing features for predicting labels. This ensures that the selected features provide more information for predicting labels.
[0102] In this embodiment, the dynamic correlation of the features MI(S; C|x i ) calculation includes:
[0103]
[0104]
[0105] Among them, p(b|a) represents the candidate feature x i The value of a is the selected feature x j The probability of taking the value b; p(c|a) represents the probability of taking the value b in the candidate feature x i The value of is a, the central label c k is the probability of taking value c; p(b,c|a) represents the probability of taking value c in candidate feature x i The feature x is selected under the premise that the value of is a j The value is b, the center label is c k is the probability of taking the value c.
[0106] The goal of feature selection is to select features that maximize the prediction accuracy of the label set within a limited number of feature selections. Therefore, the patent proposes a new multi-label feature evaluation function. This method introduces feature synergy into the existing multi-label evaluation function to retrieve the occurrence of features to be selected, so that the selected features produce synergy. The multi-label feature selection method is defined as:
[0107]
[0108] The multi-label feature selection method selects the feature with the highest score in the candidate feature set F′ each time and adds it to the selected feature set until the number of features in the selected feature set S reaches the specified number of feature selections, and the algorithm terminates.
[0109] The process of feature selection in this embodiment is as follows: Figure 3 , specifically including the following steps:
[0110] Determine the feature set, representative label set, and number of features to be selected;
[0111] Calculate the score of each feature in the feature set and add the highest feature to the selected feature set until the number of features in the selected feature set meets the feature selection requirement;
[0112] Output the selected feature set as the feature set corresponding to the current dataset selection.
[0113] The present invention also proposes a preprocessing device for complex multi-label medical data containing mutual relationships, which is used to implement a preprocessing method for complex multi-label medical data containing mutual relationships, including a data cleaning module, a data discretization module, a representative label generation module, a feature selection module, and a data output module, wherein:
[0114] The data cleaning module is used to delete features with missing values greater than 99% in the data, delete the collection items in the data set that are not related to the prediction, and fill the vacancies in the data set with 0 values;
[0115] The data discretization module is used to discretize the numerical features in the data set using the minimum entropy method;
[0116] The representative label generation module is used to cluster the labels of the data set, take the label of each cluster close to the cluster center as the representative label of the cluster, and update the labels of all data in each cluster to the corresponding representative label;
[0117] Feature selection module, used to select data features based on the synergy, correlation, redundancy and dynamic changes between data features and labels;
[0118] The data output module is used to use the labels of the processed data and the selected features as the preprocessed data to complete the preprocessing.
[0119] After preprocessing, the feature-selected dataset has a feature set that is more relevant to the label set, and these feature sets contain less redundant information. The processed features can be used to learn the information in the dataset using emerging long-short-term memory networks, transformer models, or traditional machine learning SVM and KNN classifiers. Compared with other feature-selected data, higher accuracy and better learning efficiency can be achieved in prediction.
[0120] The medical data in the present invention can be the patient's historical medical information, such as the data in the UKB database. The Chinese data in this database contains the medical information of 500,000 patients in the past 10 years, including the publication of nearly 2,000 different diseases, dozens of biochemical detection indicators, and various other physical measurement results. The classification is trained through the pre-processed data, and the classifier is used to predict the cause, possibility of onset, prevention measures and treatment measures of the disease corresponding to the corresponding medical data. This application pre-processes the user's historical medical information and inputs it into a pre-trained classifier, which can predict the user's possible disease type, which can be used in scenarios such as user self-diagnosis registration or hospital triage.
[0121] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A method for preprocessing complex multi-label medical data containing interrelationships, characterized in that: The specific steps include: Delete features with missing values greater than 99% in the data, delete the collection items in the data set that are not related to the prediction, and fill the vacancies in the data set with 0 values; The minimum entropy method is used to discretize the numerical features in the data set; Cluster the labels of the data set, take the label of each cluster close to the cluster center as the representative label of the cluster, and update the labels of all data in each cluster to the corresponding representative label; The data features are selected based on the synergy, correlation, redundancy, and dynamic changes between data features and labels, including: Among them, J(x i ) is the feature selection function, the candidate feature x i The larger the value of the corresponding feature selection function, the higher the candidate feature x i The higher the possibility of being selected; MI(x i ; C|S) represents feature correlation; MI(S; C|x i ) represents the dynamic feature correlation; MI(x i ; S|C) represents feature redundancy; MI(x i ; S; C) represents the synergy of features; C represents the label set, c k represents the kth label in the label set; S is the selected feature set; Dynamic correlation of features MI(S;C|x i ) calculation includes: Among them, p(a,b,c) represents the candidate feature x i The value of is a, the selected feature x j The value is b, the center label is c k is the joint probability of c; p(b|a) represents the probability of candidate feature x i The value of a is the selected feature x j The probability of taking the value b; p(c|a) represents the probability of taking the value b in the candidate feature x i The value of is a, the central label c k is the probability of taking value c; p(b,c|a) represents the probability of taking value c in candidate feature x i The feature x is selected under the premise that the value of is a j The value is b, the center label is c k is the probability of taking the value c; The labels of the processed data and the selected features are used as the preprocessed data to complete the preprocessing.
2. A method for preprocessing complex multi-label medical data containing mutual relationships according to claim 1, characterized in that: The process of obtaining representative labels for data through clustering includes the following steps: The label space L of the dataset is L={l1,l2,...,l m } Granulation is performed to obtain k information granules; Each label l i are assigned to their nearest representative label c in turn i Thus, a new label information particle is formed. For each new label information particle G i Regenerate representative label c i .
3. A method for preprocessing complex multi-label medical data containing mutual relationships according to claim 2, characterized in that: Generate representative label c i The process includes: Among them, S(G i ,l s ) represents an information particle G i Its representative label l s distance.
4. The method for preprocessing complex multi-label medical data containing mutual relationships according to claim 3, characterized in that: An information particle G i Its representative label l s The distance S(G i ,l s ) is expressed as: Among them, H(l j ) indicates label l j The information entropy of p(a) represents the label l j The probability that the value of is a; MI(l j ;c i ) indicates label l j With center label c i The mutual information of p(a,b) represents the label l j The value is a, the center label c i The joint probability of taking value b; NMI(l j ,c i ) indicates label l j With center label c i The label mutual information of 5. The method for preprocessing complex multi-label medical data containing mutual relationships according to claim 1, characterized in that: The synergy of features MI(x i ; S; C) calculation includes: Among them, MI(x i ;S;c k ) is the feature synergy of single-label feature selection; MI(x i ;x j ;c k ) indicates that in the candidate feature x i and selected features x j In label c k The mutual information under i ;x j ;c k )=MI(x i ;x j )-MI(x i ;x j |c k ), p(a) represents the label l j The probability of taking the value a, p(a,b) represents the candidate feature x i The value of is a, the selected feature x j The joint probability of taking the value b; p(a,b,c) represents the candidate feature x i The value of is a, the selected feature x j The value is b, the center label is c k is the joint probability of c; p(a,b|c) represents the probability of the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b, p(a|c) represents the probability of taking the value b at the center label c k The candidate feature x is the premise of c i The probability that the value of is a.
6. The method for preprocessing complex multi-label medical data containing mutual relationships according to claim 1, characterized in that: Feature correlation MI(x i ; C|S) calculation includes: Among them, p(a|b) is the selected feature x j The candidate feature x is taken as b i The probability of taking the value a; p(c|b) is the selected feature x j The central label c is assumed to be b. k is the probability of c; p(a,c|b) represents the selected feature x j The candidate feature x is taken as b i The value of is a, the selected feature x j The probability of taking the value b.
7. The method for preprocessing complex multi-label medical data containing mutual relationships according to claim 1, characterized in that: Feature redundancy MI(x i ; The calculation of S|C) includes: Among them, p(a|c) represents the center label c k The candidate feature x is the premise of c i The probability of taking the value of a; p(b|c) represents the probability of taking the value of a at the center label c k The selected feature x is c j The probability of taking the value b; p(a,b|c) represents the probability of taking the value b at the center label c k The candidate feature x is the premise of c i The value of is a, the selected feature x j The probability of taking the value b.
8. A preprocessing device for complex multi-label medical data containing mutual relationships, characterized in that: The method for preprocessing complex multi-label medical data containing interrelationships according to claim 1 comprises a data cleaning module, a data discretization module, a representative label generation module, a feature selection module, and a data output module, wherein: The data cleaning module is used to delete features with missing values greater than 99% in the data, delete the collection items in the data set that are not related to the prediction, and fill the vacancies in the data set with 0 values; The data discretization module is used to discretize the numerical features in the data set using the minimum entropy method; The representative label generation module is used to cluster the labels of the data set, take the label of each cluster close to the cluster center as the representative label of the cluster, and update the labels of all data in each cluster to the corresponding representative label; Feature selection module, used to select data features based on the synergy, correlation, redundancy and dynamic changes between data features and labels; The data output module is used to use the labels of the processed data and the selected features as the preprocessed data to complete the preprocessing.