A causal feature extraction method for dynamic medical data

Abstract

The present invention discloses a causal feature extraction method for dynamic medical data, including: 1 extracting features from a medical data set and constructing / updating a feature set of medical data; 2 summarizing the medical data information in the form of statistical information by constructing / updating a full-dimensional tree structure based on the extracted features and current medical data; 3 assigning an existing key feature (factor) set to an initial feature subset for this causal feature extraction, i.e., initializing the optimal subset; 4 using the causal relationship inference theory and algorithm, selecting the optimal feature subset through iteration of adding and deleting operations, i.e., obtaining the key feature information of the current medical data; 5 when new medical data arrives, repeating steps 1-4, and selecting real-time key features in the dynamic medical data. The present invention can realize the discovery of key features from dynamic medical data, and make the extraction of key features more interpretable and robust.

Description

Technical Field

[0001] The invention belongs to the field of data mining, and involves technologies such as causal discovery and artificial intelligence, and is specifically a causal feature extraction method for dynamic medical data. Background Art

[0002] With the rapid development of big data technology, using information technology to process health and medical big data to protect the health of residents has become one of the hottest topics. Health and medical big data refers to the data related to health and medical care generated in the process of disease prevention and health management. However, most medical data is generated dynamically, which is huge and messy. Medical staff cannot get key information from numerous entries in a timely and accurate manner. Therefore, while capturing and summarizing dynamic medical data, it is of great significance to extract key feature information from current medical data in real time in order to better assist doctors in their work.

[0003] Feature selection is required to select key features. As a data processing technology, feature selection can reduce the number of features, remove redundant features and noise data, and is widely used in various fields of machine learning and data mining. Feature selection is to select a necessary and as small a set of feature subsets as possible from a high-dimensional data set for various machine learning and data mining tasks, and has similar or better results than using complete data features.

[0004] However, the traditional feature selection method, which mainly selects feature subsets by calculating the correlation between features and categorical variables, has some disadvantages:

[0005] (1) Lack of interpretability. Correlation aims to capture the co-occurrence between features and categorical variables, so features selected by correlation cannot provide convincing explanations for the prediction model.

[0006] (2) Lack of robustness. The performance of the prediction model constructed by selecting a feature subset based on the correlation between features and categorical variables may be significantly reduced when applied to other datasets with the same distribution, and the effect will be much worse than that during verification, that is, the model is unreliable. Summary of the invention

[0007] In order to overcome the shortcomings of the prior art and cope with dynamic medical data, the present invention proposes a causal feature extraction method for dynamic medical data, in order to more accurately select key features to assist doctors in their work, while continuously following up on the patient's condition, selecting key factors in real time, and improving the interpretability and robustness of the selection.

[0008] The present invention adopts the following technical solutions to solve the technical problems:

[0009] The causal feature extraction method for dynamic medical data of the present invention is characterized by being performed according to the following steps:

[0010] Step 1: Get the medical data record of the ith time and record it as Among them, record j,i represents the jth medical data record of the i-th time, 1≤j≤R i ; R i represents the number of medical data records at the i-th time, and the j-th medical data record at the i-th time j,i Includes M i Features, denoted as Among them, f m,j,i Represents the jth medical data record of the i-th time j,i The mth feature in ; and the mth feature f m,j,i YesN m Different values; M j,i Represents the jth medical data record of the i-th time j,i The number of features in , let the jth record record j,i The category label is L j,i ;

[0011] Step 2: Build and update the structure of the full-dimensional tree, and summarize all data information in the form of statistical information in the full-dimensional tree:

[0012] Step 2.0, initialize i=1; k=1;

[0013] Step 2.1, create a new root node root and use it as the k-th layer statistical node, and record Data i The number of records;

[0014] Step 2.2, press Data i Starting from the mth feature, create a new child node of the statistical node under the kth statistical node and use it as the kth feature node;

[0015] Step 2.3: According to the number of values ​​of each feature, under each feature node of the kth layer, a new child node of the feature node is created downward, and used as the k+1th layer statistical node. The k+1th layer statistical node records the number of entries in the data set that have the same value as the kth layer feature node, and serves as statistical information;

[0016] Step 2.4: After assigning m+1 to m and k+1 to k, return to step 2.2 and execute sequentially until the full-dimensional tree is constructed, thus obtaining the i-th full-dimensional tree, recorded as ADT. i , used to store the i-th medical data record;

[0017] Step 2.5, assign i+1 to i;

[0018] Step 2.6: In the i-1th full-dimensional tree ADT i-1 The number of records in the updated root node root is

[0019] Step 2.7, initialize k=1;

[0020] Step 2.8, Update the k-th layer feature node under the k-th layer statistical node: If Data i If a new feature appears in the , a corresponding feature node is created under the k-th layer statistical node;

[0021] Step 2.9: Statistical Data i The number of entries with the same value as the feature node of the kth layer is accumulated into the statistical information of each statistical node of the kth layer under each feature node of the kth layer, thereby completing the update of the statistical node. At the same time, if Data i If a new value appears in the feature, a corresponding statistical node is created under the k-th layer feature node to record Data i The number of entries in the same as the new value;

[0022] Step 2.10: After assigning k+1 to k, loop through step 2.8 in sequence until the update of the i-th full-dimensional tree is completed and the medical data records of the previous i times are stored;

[0023] Step 2.11, return to step 2.5 and execute until the storage of all medical data records is completed;

[0024] Step 3: Construct the optimal feature subset, that is, the key feature set F′:

[0025] Step 3.1, initialize i=1;

[0026] Step 3.2: Define the medical data record Data for the ith time i The corresponding key feature set selected for the i-th time is F′ i ={f′ 1,i ,f′ 2,i ,…,f′ m′,i ,…,f′ M′,i}, where f′ m′,i Represents the key feature set F′ selected for the i-th time i The m′th key feature in the ith time, M′ represents the key feature set F′ selected for the ith time i The number of key features in ;

[0027] Step 3.3, Initialization

[0028] Step 3.4: Use the causal relationship inference method to obtain the medical data record Data i M i Select possible features from the features as key features and add them to the optimal feature subset F′ i middle;

[0029] Step 3.4.1. Define the candidate feature set C i ={c i,k |c i,k ∈F i \F′ i , 1≤k≤M i}, where F i Represents the medical data record Data for the i-th time i The M included in i A collection of features, and f m,i Represents the medical data record Data for the i-th time i The mth feature, c i,k Represents the medical data record Data for the i-th time i Includes M i A set of features F i Except for the key feature set F′ selected for the i-th time i The kth feature other than ;

[0030] Step 3.4.2, record the medical data of the ith time i The category label L i ={L j,i |j=1,2,…,R i} and candidate feature set C i Each feature in is used as a feature node, and starting from the root node of the full-dimensional tree, all feature nodes and their values ​​are traversed to obtain the first contingency table between all feature nodes, which is used to characterize the number of occurrences of different combinations of values ​​between each feature node;

[0031] Step 3.4.3: Take the key feature set F′ selected for the i-th time i is the condition set, and the candidate feature set C is calculated using formula (1) i Each feature in and the class label L i Conditional correlation between:

[0032]

[0033] In formula (1), G 2 Represents statistical values, Represents the medical data record Data at the i-th timei In the equation, the m′th key feature f′ is satisfied. m′,i =a and label L j,i = the number of records of b, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i =a, label L j,i = b and condition set F′ i The number of records whose value is set β; Represents the medical data record Data at the i-th time i In the condition set F′, i The value is the number of records in set β, Represents the medical data record Data at the i-th time i , satisfying label L j,i = b and condition set F′ i The value is the number of records in set β, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i = a and condition set F′ i The number of records whose value is set β;

[0034] Step 3.4.4: Use formula (2) to calculate the key feature set F′ selected for the i-th time i The m′th key feature f′ m′,i and the class label L i The degrees of freedom between m′,i :

[0035]

[0036] In formula (2), Represents the medical data record Data for the i-th time i The m′th key feature f′ m′,i The number of different values ​​of Represents the medical data record Data for the i-th time i The number of different values ​​of the category label in ;

[0037] Step 3.4.5: Take the m′th key feature f′ m′,i and the class label L i The statistical value G 2 In degrees of freedom df m′,i The next asymptotically becomes a chi-square distribution, and a chi-square test is performed on the chi-square distribution to obtain the probability value P m′,i , the probability value P m′,irepresents the probability of falsely rejecting the null hypothesis when the null hypothesis is true, where the null hypothesis is the m′th key feature f′ m′,i and the class label L i The conditions between them are independent;

[0038] Step 3.4.6: Calculate M′ key features and class labels L i ={L j,i |j=1,2,…,R i}, and sort the probability values ​​in the probability value set in descending order;

[0039] Step 3.4.7. Select the first probability value P after descending sorting max,i The corresponding key feature f′ max,i , if the first probability value P max,i If it is less than the significance level α, then the null hypothesis is rejected, that is, f′ max,i and label L i The conditional dependence between them, and f′ max,i Add to the key pathogenic factor set F′ selected for the i-th time i middle;

[0040] Step 3.5: Use causal discovery theory to remove the key pathogenic factor set F′ selected for the i-th time i Key features of errors in

[0041] Step 3.5.1: For the category label L i ={L j,i |j=1,2,…,R i} and the key pathogenic factor set F′ i Each feature in is taken as a feature node, and starting from the root node root of the full-dimensional tree, all feature nodes and their values ​​are traversed to obtain the second contingency table between all feature nodes;

[0042] Step 3.5.2: In the key pathogenic factor set F′ i Select one feature as the current feature, and take the key pathogenic factor set F′ i All features except the current feature are the condition set; the key pathogenic factor set F′ is calculated using formula (1) i The current features are respectively associated with the category label L i The conditional correlation between

[0043] Step 3.5.3: Use formula (2) to calculate the current feature and category label L i The degrees of freedom between the current feature and the category label L i The statistical value G 2The value is asymptotically distributed in the corresponding degrees of freedom, and a chi-square test is performed on the chi-square distribution to obtain the corresponding probability value; if the corresponding probability value is greater than the significance level α, the null hypothesis is accepted, that is, the current feature and category label L i The conditional independence between them indicates that the current feature is not a real key feature, and the current feature is removed from the key feature set F′ i Removed;

[0044] Step 3.6: Repeat steps 3.4 and 3.5 until no features are added to the key pathogenic factor set F′ i In order to obtain the key features of the previous i medical data records;

[0045] Step 4: When new medical data arrives, assign i+1 to i and return to step 2 to execute sequentially.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. The causal feature selection method of key features proposed in the present invention is used to reduce the dimension of the feature space, making the feature space more conducive to further feature learning; the present invention uses a feature selection method based on causal learning to make the key features selected from medical data more interpretable and robust.

[0048] 2. The full-dimensional tree structure used in the present invention can save complete medical data in the form of statistical information in a relatively small space, and connect the tree structure with the contingency table, so that the required contingency table can be obtained from the tree structure, so that G is more accurate when selecting key features. 2 The calculation is more convenient and efficient. At the same time, the medical data continuously generated by patients is aggregated by updating the existing full-dimensional tree structure, and each piece of data is scanned only once when it arrives, which also avoids repeated and redundant data reading caused by calculation.

[0049] 3. The present invention adopts a strategy of initializing the current selection through existing key features, so as to achieve continuous follow-up of residents' health and continuity; and in the selection of a new round of key features, by adding new key features and removing invalid features, the real-time and practicality of key feature extraction is achieved. In addition, by initializing the key feature set, it avoids starting from scratch, speeds up the addition and removal operations, and makes the selection faster.

[0050] 4. The method proposed in the present invention can be used in various data analysis tasks, such as sales data analysis, decision information processing, user social media interest mining, etc., which helps merchants and researchers to mine feature information closely related to the target task, so as to better perform data analysis and information mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic diagram of the method of the present invention;

[0052] Figure 2 A schematic diagram of a data set Data1 for recording the years of the first medical data record;

[0053] Figure 3 A schematic diagram of a full-dimensional tree structure ADT1 constructed from the data set Data1;

[0054] Figure 4 A schematic diagram of a data set Data2 for recording the years of medical data for the second time;

[0055] Figure 5 A partial schematic diagram of ADT2 obtained by updating ADT1 according to the new data set Data2. DETAILED DESCRIPTION

[0056] In this embodiment, a causal feature extraction method for dynamic medical data is to first extract multiple features from the medical data set and construct or update the feature set of the medical data; then, based on the extracted features and the current medical data, the medical data information is summarized in the form of statistical information by constructing or updating a full-dimensional tree structure; then, the existing key feature set is used to assign the initial feature subset of this causal feature extraction, that is, to initialize; using causal inference theory and algorithms, the causal relationship between features is calculated, and through iterative addition and deletion operations, the optimal feature subset is selected to obtain the current optimal key feature; and when new medical data arrives, the above operations are repeated to select real-time key features in the dynamic medical data. Specifically, if Figure 1 As shown, the following steps are included:

[0057] Step 1: Get the medical data record of the ith time and record it as Data i ={record 1,i ,record 2,i ,…,record j,i ,…,record Ri,i}, where record j,i represents the jth medical data record of the i-th time, 1≤j≤R i ; R i represents the number of medical data records at the i-th time, and the j-th medical data record at the i-th time j,i Includes M i Features, denoted as Among them, f m,j,i Represents the jth medical data record of the i-th time j,iThe mth feature in ; and the mth feature f m,j,i YesN m Different values; M j,i Represents the jth medical data record of the i-th time j,i The number of features in , let the jth record record j,i The category label is L j,i ;

[0058] like Figure 2 As shown, it indicates the first medical data record, Data1 = {record 1,1 ,record 2,1 ,record 3,1 ,record 4,1}, where R1=4, feature F={f1}, feature f1 has 2 different values, namely N1=2, and the category label is L.

[0059] Step 2: Build and update the structure of the full-dimensional tree, and summarize all data information in the form of statistical information in the full-dimensional tree:

[0060] like Figure 3 As shown, it represents the full-dimensional tree structure ADT1 built on Data1.

[0061] Step 2.0, initialize i=1; k=1;

[0062] Step 2.1, create a new root node root and use it as the k-th layer statistical node, and record Data i At this time, Data1 has 4 records, that is, the statistical node at the first layer saves the statistical value 4;

[0063] Step 2.2, press Data i Starting from the mth feature, under the kth layer statistical node, a new sub-node of the statistical node is created downward and used as the kth layer feature node, that is, a new first-layer feature node f is created under the first-layer statistical node. 1,1 and L1;

[0064] Step 2.3: According to the number of values ​​of each feature, create a sub-node of each feature node under the k-th layer and use it as the k+1-th layer statistical node. The k+1-th layer statistical node records the number of items in the data set that have the same value as the k-th layer feature node and serves as statistical information. 1,1 There are two value states 0 and 1, and the number of records with f1=0 in Data1 is 1, and the number of records with f1=1 is 3. Then, two new statistical nodes f1=0 and f1=1 are created under the first-level feature nodes, and the statistical values ​​are 1 and 3 respectively.

[0065] Step 2.4: After assigning m+1 to m and k+1 to k, return to step 2.2 and execute sequentially until the full-dimensional tree is constructed, thus obtaining the i-th full-dimensional tree, recorded as ADT. i , used to store the medical data record of the ith time. For example, under the statistical nodes f1=0 and f1=1, a new feature node of the next layer is created starting from feature L;

[0066] Step 2.5, assign i+1 to i;

[0067] like Figure 4 As shown, a new data set Data2 arrives, and ADT2 is obtained by updating ADT1 according to the data records in Data2, as shown in Figure 5 shown.

[0068] Step 2.6: In the i-1th full-dimensional tree ADT i-1 The number of records in the updated root node root is At this time, there are 4 records in Data2, so the statistical value of the root statistical node is updated to 4+4=8;

[0069] Step 2.7, initialize k=1;

[0070] Step 2.8, Update the k-th layer feature node under the k-th layer statistical node: If Data i If a new feature appears in Data2, a corresponding feature node is created under the k-th statistical node. At this time, a new feature f2 appears in Data2, so a corresponding statistical node f2 is created under the first-level statistical node.

[0071] Step 2.9: Statistical Data i The number of entries with the same value as the feature node of the kth layer is accumulated into the statistical information of each statistical node of the kth layer under each feature node of the kth layer, thereby completing the update of the statistical node. At the same time, if Data i If a new value appears in the feature, a corresponding statistical node is created under the k-th layer feature node to record Data i The number of entries in Data2 that have the same value as the new value, for example, if the statistical values ​​of features f1=0 and f1=1 in Data2 are 1 and 2 respectively, the statistical values ​​of the corresponding statistical nodes are updated to 1+1=2 and 3+2=5. At the same time, if a new value state f1=2 appears in f1, a corresponding statistical node f1=2 is created;

[0072] Step 2.10: After assigning k+1 to k, loop through step 2.8 in sequence until the update of the i-th full-dimensional tree is completed and the medical data records of the previous i times are stored;

[0073] Step 2.11, return to step 2.5 and execute until the storage of all medical data records is completed;

[0074] Step 3: Construct the optimal feature subset, that is, the key feature set F′:

[0075] Step 3.1, initialize i=1;

[0076] Step 3.2: Define the medical data record Data for the ith time i The corresponding key feature set selected for the i-th time is F′ i ={f′ 1,i ,f′ 2,i ,…,f′ m′,i ,…,f′ M′,i}, where f′ m′,i Represents the key feature set F′ selected for the i-th time i The m′th key feature in the ith time, M′ represents the key feature set F′ selected for the ith time i The number of key features in ;

[0077] Step 3.3, Initialization That is, the key feature set is initialized to an empty set;

[0078] Step 3.4: Use the causal relationship inference method to obtain the medical data record Data i M i Select possible features from the features as key features and add them to the optimal feature subset F′ i middle;

[0079] Step 3.4.1. Define the candidate feature set C i ={c i,k |c i,k ∈F i \F′ i , 1≤k≤M i}, where F i Represents the medical data record Data for the i-th time i The M included in i A collection of features, and f m,i Represents the medical data record Data for the i-th time i The mth feature, c i,k Represents the medical data record Data for the i-th time i Includes M i A set of features F i Except for the key feature set F′ selected for the i-th time i The kth feature other than ;

[0080] Step 3.4.2, record the medical data of the ith time i The category label L i ={L j,i |j=1,2,…,R i} and candidate feature set C i Each feature in is used as a feature node, and starting from the root node of the full-dimensional tree, all feature nodes and their values ​​are traversed to obtain the first contingency table between all feature nodes, which is used to characterize the number of occurrences of different combinations of values ​​between each feature node;

[0081] Step 3.4.3: Take the key feature set F′ selected for the i-th time i is the condition set, and the candidate feature set C is calculated using formula (1) i Each feature in and the class label L i Conditional correlation between:

[0082]

[0083] In formula (1), G 2 Represents statistical values, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i =a and label L j,i = the number of records of b, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i =a, label L j,i = b and condition set F′ i The number of records whose value is set β; Represents the medical data record Data at the i-th time i In the condition set F′, i The value is the number of records in set β, Represents the medical data record Data at the i-th time i , satisfying label L j,i = b and condition set F′ i The value is the number of records in set β, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i = a and condition set F′ i The number of records whose value is set β;

[0084] At this time, the contingency table required for calculation can be directly obtained from the constructed full-dimensional tree, so that G is2 Calculation is more convenient and efficient.

[0085] Step 3.4.4: Use formula (2) to calculate the key feature set F′ selected for the i-th time i The m′th key feature f′ m′,i and the class label L i The degrees of freedom between m′,i :

[0086]

[0087] In formula (2), Represents the medical data record Data for the i-th time i The m′th key feature f′ m′,i The number of different values ​​of Represents the medical data record Data for the i-th time i The number of different values ​​of the category label in ;

[0088] Step 3.4.5: Take the m′th key feature f′ m′,i and the class label L i The statistical value G 2 In degrees of freedom df m′,i The next asymptotically becomes a chi-square distribution, and a chi-square test is performed on the chi-square distribution to obtain the probability value P m′,i , probability value P m′,i It represents the probability of falsely rejecting the null hypothesis when the null hypothesis is true, where the null hypothesis is the m′th key feature f′ m′,i and the class label L i The conditions between them are independent;

[0089] Step 3.4.6: Calculate M′ key features and class labels L i ={L j,i |j=1,2,…,R i}, and sort the probability values ​​in the probability value set in descending order;

[0090] Step 3.4.7. Select the first probability value P after descending sorting max,i The corresponding key feature f′ max,i , that is, select the feature with the maximum probability value. If the first probability value P max,i If it is less than the significance level α, then the null hypothesis is rejected, that is, f′ max,i and label L i The conditional dependence between them, and f′ max,i Add to the key pathogenic factor set F′ selected for the i-th time i middle;

[0091] In this embodiment, the significance level α is set to 0.05, and the significance level can be adjusted according to the specific task;

[0092] Step 3.5: Use causal discovery theory to remove the key pathogenic factor set F′ selected for the i-th time i Key features of errors in

[0093] Step 3.5.1: For the category label L i ={L j,i |j=1,2,…,R i} and the key pathogenic factor set F′ i Each feature in is taken as a feature node, and starting from the root node root of the full-dimensional tree, all feature nodes and their values ​​are traversed to obtain the second contingency table between all feature nodes;

[0094] Step 3.5.2: In the key pathogenic factor set F′ i Select one feature as the current feature, and take the key pathogenic factor set F′ i All features except the current feature are the condition set; the key pathogenic factor set F′ is calculated using formula (1) i The current features are respectively associated with the category label L i The conditional correlation between

[0095] Step 3.5.3: Use formula (2) to calculate the current feature and category label L i The degrees of freedom between the current feature and the category label L i The statistical value G 2 The value is asymptotically distributed in the corresponding degrees of freedom, and a chi-square test is performed on the chi-square distribution to obtain the corresponding probability value; if the corresponding probability value is greater than the significance level α, the null hypothesis is accepted, that is, the current feature and category label L i The conditional independence between them indicates that the current feature is not a real key feature, and the current feature is removed from the key feature set F′ i Removed;

[0096] Step 3.6: Repeat steps 3.4 and 3.5 until no features are added to the key pathogenic factor set F′ i In order to obtain the key features of the previous i medical data records;

[0097] Step 4: When new medical data arrives, assign i+1 to i and return to step 2 to execute sequentially.

Claims

1. A causal feature extraction method for dynamic medical data, characterized by Proceed as follows: Step 1: Get the medical data record of the ith time and record it as Among them, record j,i represents the jth medical data record of the i-th time, 1≤j≤R i ; R i represents the number of medical data records at the i-th time, and the j-th medical data record at the i-th time j,i Includes M i Features, denoted as Among them, f m,j,i Represents the jth medical data record of the i-th time j,i The mth feature in ; and the mth feature f m,j,i YesN m Different values; M j,i Represents the jth medical data record of the i-th time j,i The number of features in , let the jth record record j,i The category label is L j,i ; Step 2: Build and update the structure of the full-dimensional tree, and summarize all data information in the form of statistical information in the full-dimensional tree: Step 2.0, initialize i=1; k=1; Step 2.1, create a new root node root and use it as the k-th layer statistical node, and record Data i The number of records; Step 2.2, press Data i Starting from the mth feature, under the kth level statistical node, create a new child node of the statistical node downward and use it as the kth level feature node; Step 2.3: According to the number of values ​​of each feature, under each feature node of the kth layer, a new child node of the feature node is created downward, and used as the k+1th layer statistical node. The k+1th layer statistical node records the number of entries in the data set that have the same value as the kth layer feature node, and serves as statistical information; Step 2.4: After assigning m+1 to m and k+1 to k, return to step 2.2 and execute sequentially until the full-dimensional tree is constructed, thus obtaining the i-th full-dimensional tree, recorded as ADT. i , used to store the i-th medical data record; Step 2.5, assign i+1 to i; Step 2.6: In the i-1th full-dimensional tree ADT i-1 The number of records in the updated root node root is Step 2.7, initialize k=1; Step 2.8, Update the k-th layer feature node under the k-th layer statistical node: If Data i If a new feature appears in the , a corresponding feature node is created under the k-th layer statistical node; Step 2.9: Statistical Data i The number of entries with the same value as the feature node of the kth layer is accumulated into the statistical information of each statistical node of the kth layer under each feature node of the kth layer, thereby completing the update of the statistical node. At the same time, if Data i If a new value appears in the feature, a corresponding statistical node is created under the k-th layer feature node to record Data i The number of entries in the same as the new value; Step 2.10: After assigning k+1 to k, loop through step 2.8 in sequence until the update of the i-th full-dimensional tree is completed and the medical data records of the previous i times are stored; Step 2.11, return to step 2.5 and execute until the storage of all medical data records is completed; Step 3: Construct the optimal feature subset, that is, the key feature set F′: Step 3.1, initialize i=1; Step 3.2: Define the medical data record Data for the ith time i The corresponding key feature set selected for the i-th time is F i ′={f′ 1,i ,f′ 2,i ,…,f′ m′,i ,…,f′ M′,i }, where f′ m′,i represents the key feature set F selected for the i-th time i ′ represents the m′th key feature in ′, M′ represents the key feature set F selected for the i-th time i The number of key features in ′; Step 3.3, Initialization Step 3.4: Use the causal relationship inference method to obtain the medical data record Data i M i Select possible features from the features as key features and add them to the optimal feature subset F i 'middle; Step 3.4.

1. Define the candidate feature set C i ={c i,k |c i,k ∈F i \F i ′, 1≤k≤M i }, where F i Represents the medical data record Data for the i-th time i The M included in i A collection of features, and f m,i Represents the medical data record Data for the i-th time i The mth feature, c i,k Represents the medical data record Data for the i-th time i Includes M i A set of features F i Except for the key feature set F selected for the i-th time i The kth feature other than ′; Step 3.4.2, record the medical data of the ith time i The category label L i ={L j,i |j=1,2,…,R i } and candidate feature set C i Each feature in is used as a feature node, and starting from the root node of the full-dimensional tree, all feature nodes and their values ​​are traversed to obtain the first contingency table between all feature nodes, which is used to characterize the number of occurrences of different combinations of values ​​between each feature node; Step 3.4.3: Take the key feature set F selected for the i-th time i ′ is the condition set, and the candidate feature set C is calculated using formula (1) i Each feature in is associated with a class label L i Conditional correlation between: In formula (1), G 2 Represents statistical values, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i =a and label L j,i = the number of records of b, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i =a, label L j,i = b and condition set F i ′ is the number of records in set β; Represents the medical data record Data at the i-th time i In the condition set F i ′ is the number of records in set β, Represents the medical data record Data at the i-th time i In the case where label L j,i = b and condition set F i ′ is the number of records in set β, Represents the medical data record Data at the i-th time i In the equation, the m′th key feature f′ is satisfied. m′,i = a and condition set F i ′ is the number of records in set β; Step 3.4.4: Use formula (2) to calculate the key feature set F selected for the i-th time i The m′th key feature f′ in ′ m′,i and the class label L i The degrees of freedom between m′,i : In formula (2), Represents the medical data record Data for the i-th time i The m′th key feature f′ m′,i The number of different values ​​of N Li Represents the medical data record Data for the i-th time i The number of different values ​​of the category label in ; Step 3.4.5: Take the m′th key feature f′ m′,i and the class label L i The statistical value G 2 In degrees of freedom df m′,i The next asymptotically becomes a chi-square distribution, and a chi-square test is performed on the chi-square distribution to obtain the probability value P m′,i , the probability value P m′,i represents the probability of falsely rejecting the null hypothesis when the null hypothesis is true, where the null hypothesis is the m′th key feature f′ m′,i and the class label L i The conditions between them are independent; Step 3.4.6: Calculate M′ key features and class labels L i ={L j,i |j=1,2,…,R i }, and sort the probability values ​​in the probability value set in descending order; Step 3.4.

7. Select the first probability value P after descending sorting max,i The corresponding key feature f′ max,i , if the first probability value P max,i If it is less than the significance level α, then the null hypothesis is rejected, that is, f′ max,i and label L i The conditional dependence between them, and f′ max,i Add to the key pathogenic factor set F selected for the i-th time i 'middle; Step 3.5: Use causal discovery theory to remove the key pathogenic factor set F selected for the i-th time i The key features of the error in ′; Step 3.5.1: For the category label L i ={L j,i |j=1,2,…,R i } and key pathogenic factor set F i Each feature in ′ is taken as a feature node, and starting from the root node root of the full-dimensional tree, all feature nodes and their values ​​are traversed to obtain the second contingency table between all feature nodes; Step 3.5.2: In the key pathogenic factor set F i ′, select a feature as the current feature, and use the key pathogenic factor set F i All features except the current feature in ′ are the condition set; the key pathogenic factor set F is calculated using formula (1) i The current features in ′ are respectively related to the category label L i The conditional correlation between Step 3.5.3: Use formula (2) to calculate the current feature and category label L i The degrees of freedom between the current feature and the category label L i The statistical value G 2 The value is asymptotically distributed in the corresponding degrees of freedom, and a chi-square test is performed on the chi-square distribution to obtain the corresponding probability value; if the corresponding probability value is greater than the significance level α, the null hypothesis is accepted, that is, the current feature and category label L i The conditional independence between them indicates that the current feature is not a real key feature, and the current feature is removed from the key feature set F i 'removed; Step 3.6: Repeat steps 3.4 and 3.5 until no features are added to the key pathogenic factor set F. i ′, thus obtaining the key features of the previous i medical data records; Step 4: When new medical data arrives, assign i+1 to i and return to step 2 to execute sequentially.

Images