Multi-source multi-label medical clinical data feature selection method containing deletion

Abstract

The invention discloses a missing-containing multi-source multi-label medical clinical data feature selection method. The method comprises the following steps: 1, inputting multi-source medical clinical text data; 2, performing missing data completion on the text data; 3, performing oversampling on the complemented text data to obtain class balance data; 4, performing in-source clinical feature selection on each data source by adopting an improved infinite feature selection algorithm; 5, inter-source redundancy elimination is carried out on the clinical candidate feature sets obtained by different data sources, and a clinical feature selection set with redundant features eliminated is obtained. According to the invention, clinical features with high importance can be screened out from a large amount of medical clinical data containing missing, so that the efficiency and accuracy of abnormality discovery can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of feature selection in data mining, and in particular to multi-source multi-label clinical data in the medical field containing missing data, and proposes a method for selecting features of multi-source multi-label medical clinical data containing missing data. Background Art

[0002] With the rapid development of the information age and the continuous advancement of technology, the medical field has gradually entered an era of big data. In this context, multi-source and multi-label data has become more and more common. For example, a patient may have multiple test data from different sources, including physical characteristics such as weight, blood pressure, heart rate, blood sugar level, etc., as well as diagnostic characteristics such as symptoms, test results, etc. These data are usually associated with multiple abnormal labels.

[0003] However, this type of multi-source, multi-label data also brings many challenges, especially when performing feature selection. First, physical and diagnostic features are often high-dimensional and often accompanied by a large amount of missing data. In practical applications, due to various reasons, such as patients failing to undergo certain examinations, incomplete data records, etc., some features may be missing, resulting in the inability to effectively integrate information between different data sources, thereby affecting the overall performance and accuracy of the model. For example, some patients may lack certain diagnostic features, such as imaging test results, or physical feature data may be missing, such as unmeasured weight. This requires us to not only select the most representative features from different data sources when performing feature selection, but also consider how to effectively handle these missing data to avoid bias in model results.

[0004] Secondly, there is often a problem of label imbalance in medical data. For example, some abnormalities may be relatively rare, resulting in extremely uneven label distribution. In the case of multi-source multi-label data, each patient may be associated with multiple labels at the same time, such as a combination of multiple abnormalities, and the correlation between these labels also needs to be considered in the feature selection process. If the imbalance problem is not effectively handled, the model may tend to predict more common abnormalities and ignore rarer abnormalities. Summary of the invention

[0005] The present invention aims to overcome the deficiencies in the prior art and proposes a method for selecting features of multi-source and multi-label medical clinical data containing missing information, in order to screen out clinical features of high importance from a large amount of medical clinical data containing missing information, thereby improving the efficiency and accuracy of abnormality detection.

[0006] In order to achieve the above-mentioned invention object, the present invention adopts the following technical solutions:

[0007] The feature selection method of the multi-source multi-label medical clinical data with missing data of the present invention is characterized in that it is performed according to the following steps:

[0008] Step 1: Obtain medical clinical data ,in, Indicates Clinical data from various sources, Indicates the number of data sources, and , ,in, Indicates the Data sources The clinical characteristics matrix of patients, express The abnormal category label matrix of , represents the field of real numbers, Indicates Data sources The number of patients, Indicates Data sources The number of clinical features, Indicates the Data sources The clinical feature vector of the jth patient, , Indicates Data sources The jth patient The sth clinical characteristic value, when the Data sources The jth patient The sth clinical characteristic value When missing, ; , express The abnormal category label of , T represents transposition;

[0009] Let the Data sources The clinical feature space is denoted as ,in, Indicates the i-th data source No. Clinical characteristic values;

[0010] Let the Data sources The abnormal category label space is recorded as ,in, Indicates the i-th data source No. Anomaly category labels , ;

[0011] Step 2: Complete the missing clinical feature values ​​in each data source to obtain the The completed clinical feature matrix ,in, Indicates the Data sources The non-missing clinical feature vector of the j-th patient;

[0012] Step 3: Right Perform multi-label oversampling to obtain class-balanced Data sources ;

[0013] Step 3.1: Calculate the Data sources The abnormal category label space The average imbalance ratio of all abnormal category labels in , and obtain all abnormal category labels belonging to the minority class;

[0014] Step 3.2: Generate synthetic clinical feature vectors of abnormal class labels belonging to the minority class and add them to the clinical feature matrix In the above example, a synthetic anomaly class label vector belonging to the minority class is generated and added to the anomaly class label matrix and constitute the Data sources Class-balanced data source ;

[0015] Let the Class-balanced data sources The clinical feature space is denoted as ,in, Represents the i-th class balanced data source No. Clinical characteristic values, ;

[0016] Let the Class-balanced data sources The abnormal category label space is recorded as ,in, Represents the i-th class balanced data source No. Anomaly category labels, ;

[0017] Step 4: Use the improved infinite feature selection algorithm to select Class-balanced data sources Clinical feature selection was performed to obtain the Candidate feature sets on a class-balanced data source ;

[0018] Step 4.1: Calculate clinical feature values ​​based on mutual information and Fisher score and abnormal category labels The correlation between The combined weight of all clinical characteristic values ​​in ;

[0019] Step 4.2: Calculate simultaneous selection and The amount of information that can be generated , thereby constructing the Class-balanced data sources Special Information Matrix ,in, Indicates the Class-balanced data sources Clinical characteristic value The comprehensive weight of

[0020] Step 4.3: Use formula (10) to calculate the Class-balanced data sources The feature importance matrix on :

[0021] (10)

[0022] In formula (10), r is a hyperparameter term, is the identity matrix;

[0023] Step 4.4: Calculate the feature importance matrix The sum of each column in the first k columns is selected as the clinical characteristic value with the largest sum of each column. Class-balanced data sources The candidate feature set , thus obtaining balanced data sources for all classes The candidate feature set ,in, express The candidate feature values, ;

[0024] Step 5: Traverse the candidate feature set All candidate features of Perform redundancy judgment between sources to obtain a candidate feature set that eliminates redundancy , thus obtaining a clinical feature subset without redundant features , used to judge abnormalities in the patient's clinical characteristic data;

[0025] (11)

[0026] In formula (11), Indicates Class-balanced data sources The abnormal category label space, Indicates the Class-balanced data sources The candidate feature set The candidate feature values, , Indicates that the candidate feature set is removed and The collection composed of Indicates that the candidate feature set is removed Candidate feature value The collection composed later; Represents candidate feature values With the abnormal category label space The mutual information of .

[0027] The feature of the method for selecting missing multi-source multi-label medical clinical data features described in the present invention is that step 2 is performed according to the following steps:

[0028] Step 2.1: Calculate the i-th data source by cosine similarity The correlation between the clinical feature values ​​of the Laplacian matrix The Laplacian matrix of the graph of the association between and anomaly labels ;

[0029] Step 2.2: Use equation (1) to solve the i-th data source Clinical feature correlation matrix and the i-th data source Upper clinical feature space To the abnormal label space The mapping matrix ,and ,in, Represents the i-th data source The sth clinical characteristic value With Clinical characteristic values The correlation between , , let the i-th data source The sth clinical characteristic value Its own relevance :

[0030] (1)

[0031] In formula (1), , , , There are 4 hyper parameters respectively;

[0032] Step 2.3: Data sources Patient clinical characteristics matrix Right multiplication Get the first The completed clinical feature matrix ,in, Indicates the Data sources The non-missing clinical feature vector of the j-th patient.

[0033] Furthermore, the step 3.1 is performed according to the following steps:

[0034] Step 3.1.1: Use formula (2) to calculate the Data sources Abnormal label space Middle Class imbalance ratio of abnormal class labels :

[0035] (2)

[0036] In formula (2), Indicates the i-th data source No. Anomaly category labels, , It is judgment Is it present in In the indicator function, max() represents the maximum value function; It is judgment Is it present in The indicator function in , and has:

[0037] (3)

[0038] Step 3.1.2: Use formula (4) to calculate the Data sources Abnormal label space The average imbalance ratio :

[0039] (4)

[0040] In formula (4), when > , indicating that Data sources Middle Anomaly category labels is a minority class; otherwise, continue to judge the next abnormal class label, so as to obtain all the abnormal class labels belonging to the minority class.

[0041] Further, step 3.2 is performed as follows:

[0042] Step 3.2.1: Traverse the Data sources Abnormal label space For each abnormal category label in For the minority class, Remove all the The non-missing clinical feature vector is denoted as ,in, for Middle Contains The non-missing clinical feature vector, contain The total number of non-missing clinical feature vectors, ;

[0043] Step 3.2.2: Calculate based on Euclidean distance Zhongyu The closest clinical feature vectors and used as The nearest clinical feature vector of ;

[0044] Step 3.2.3: Traversal Each clinical characteristic value in Is it a numeric type? If so, calculate the corresponding clinical characteristic value and The difference between the neighbor clinical feature values ​​at the corresponding positions in the neighbor clinical feature vectors, and The sum of the differences and the product of the random number is used as a synthetic clinical feature value; otherwise, the nearest clinical feature value with the highest frequency in the nearest clinical feature vector is selected and used as the synthetic clinical feature value, so that Composite clinical features Synthetic clinical feature vector , thus obtaining the synthetic clinical feature vector set And add the Clinical feature matrix after data source completion middle;

[0045] choose The number of times the abnormal category vector corresponding to the nearest clinical feature vector appears exceeds The abnormal category label is used as The synthetic anomaly label vector , thus obtaining the synthetic anomaly category label set And add the The abnormal category label matrix of the data sources In the Data sources Class-balanced data source .

[0046] Further, step 4.1 is performed as follows:

[0047] Step 4.1.1: Calculate the Class-balanced data sources Middle Clinical characteristic values Variance , clinical characteristic values exist All labels with abnormal categories The mean value of patients , all labels that do not contain abnormal categories The mean value of patients , and clinical characteristic values exist The mean of all patients ;

[0048] Step 4.1.2: Use formula (5) to calculate the Class-balanced data sources Middle Clinical characteristic values With Anomaly category labels Fisher score , thus obtaining the Class-balanced data sources Clinical feature space With abnormal label space The correlation matrix :

[0049] (5)

[0050] Step 4.1.3: Keep the abnormal category label space Middle Anomaly category labels The first m largest Fisher scores are taken, and the other Fisher scores are set to 0, so as to obtain the processed correlation matrix and accumulate them in the column direction, and the accumulated value obtained is used as the first Clinical characteristic values The weight of the Fisher score ;

[0051] Step 4.1.4: Calculate the first Class-balanced data sources Middle Clinical characteristic values and abnormal category labels Mutual information between :

[0052] (6)

[0053] In formula (6), Clinical characteristic value When the abnormal category label The probability of Clinical characteristic value The probability of Indicates abnormal category label The probability of

[0054] Step 4.1.5: Calculate the Clinical characteristic values and the abnormal category label pair ( ) , No. Clinical characteristic values and abnormal category labels and The ternary mutual information between And known abnormal labels In the case and The conditional mutual information ,in, Indicates Class-balanced data sources Middle Anomaly category labels ;

[0055] Step 4.1.6: Use formula (7) to calculate the Class-balanced data sources Middle Clinical characteristic values The mutual information weight :

[0056] (7)

[0057] In formula (7), Indicates Clinical characteristic values The weight of is obtained by formula (8):

[0058] (8)

[0059] Step 4.1.7: Calculate the Class-balanced data sources No. Clinical characteristic values The standard deviation of ;

[0060] Step 4.1.8: Calculate using formula (9) The comprehensive weight of , thus obtaining The combined weight of all clinical characteristic values ​​in ;

[0061] (9)

[0062] In formula (9), are 3 hyperparameters, and .

[0063] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the method, and the processor is configured to execute the program stored in the memory.

[0064] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the method when executed by a processor.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] 1. In medicine, the patient's physical characteristics and diagnostic characteristics complement each other and provide different information about the patient's health status. When selecting clinical features, if only a single data source is considered, some key information may be missed, thus affecting the feature screening effect. The present invention can effectively solve the problem of clinical feature selection under multi-source and multi-label data by comprehensively considering the information of multiple data sources. It can also process multiple related abnormal category labels at the same time, further improving the flexibility and applicability of feature selection.

[0067] 2. In medical data, patients' physical characteristics and diagnostic characteristics often have missing data, such as missing test results or some physiological indicators not collected. In addition, some abnormal labels are unevenly distributed, resulting in label imbalance. The present invention fully considers the impact of feature missing and label imbalance on the feature selection process, and solves the problem of feature missing and label imbalance by constructing a feature completion matrix and using multi-label oversampling technology, thereby improving the accuracy of clinical feature selection.

[0068] 3. The present invention improves the feature selection algorithm and can effectively fuse label-specific features and other label information. A new inter-source feature fusion method is used to effectively remove redundancy and improve the purity of the clinical feature set for redundant features between different data sources. This not only improves the accuracy of clinical feature selection, but also reduces the calculation time, ensuring the dual optimization of accuracy and efficiency, and is suitable for large-scale analysis and abnormal judgment in medical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of the framework of the multi-source multi-label feature selection method of the present invention. DETAILED DESCRIPTION

[0070] In this embodiment, a method for selecting features of multi-source multi-label medical clinical data with missing data is specifically used for feature selection of clinical data with missing data. By completing the missing data and performing intra-source feature selection on the completed clinical data, some clinical features with high importance are selected from clinical features of multiple sources, and finally the inter-source clinical feature fusion is realized, and redundant features are removed, thereby improving the efficiency and accuracy of abnormal judgment. Specifically, if Figure 1 As shown, the following steps are included:

[0071] Step 1: Obtain medical clinical data ,in, Indicates Clinical data from various sources, Indicates the number of data sources, and , ,in, Indicates Data sources The clinical characteristics matrix of patients, express The abnormal category label matrix of , represents the field of real numbers, Indicates Data sources The number of patients, Indicates Data sources The number of clinical features, Indicates Data sources The clinical feature vector of the jth patient, , Indicates Data sources The jth patient The sth clinical characteristic value, when the Data sources The jth patient The sth clinical characteristic value When missing, ; , express The abnormal category label of , T represents transposition;

[0072] Order Data sources The clinical feature space is denoted as ,in, Indicates the i-th data source No. Clinical characteristic values;

[0073] Order Data sources The abnormal category label space is recorded as ,in, Indicates the i-th data source No. Anomaly category labels , .

[0074] Step 2: Complete the missing clinical feature values ​​in each data source;

[0075] Step 2.1: Calculate the correlation matrix Si between clinical features by cosine similarity , where Si ,in Indicates Data sources Middle Candidate feature values With Candidate feature values The cosine similarity of , and then calculate Si The degree matrix ,in Represents candidate feature values The degree, when 'season Finally, we get the graph Laplacian matrix ,and Similarly, the Laplacian matrix of the graph of the correlation between abnormal labels is calculated .

[0076] Use formula (1) to solve the i-th data source Clinical feature correlation matrix and the i-th data source Upper clinical feature space To the abnormal label space The mapping matrix ,and ,in, Indicates the i-th data source The sth clinical characteristic value With Clinical characteristic values The correlation between , , let the i-th data source The sth clinical characteristic value Its own relevance :

[0077] (1)

[0078] In formula (1), , , , are four hyperparameters respectively. Formula (1) can be solved by the existing gradient descent algorithm to obtain the clinical feature correlation matrix ;

[0079] Step 2.3: Data sources Patient clinical characteristics matrix Right multiplication Get the first The completed clinical feature matrix ,in, Indicates Data sources The non-missing clinical feature vector of the j-th patient.

[0080] Step 3: Right Perform multi-label oversampling to obtain class-balanced Data sources ; By sampling the anomaly category labels belonging to the minority class, the method can pay more attention to these rare anomalies, thereby increasing the probability of discovering these rare anomalies.

[0081] Step 3.1: Calculate the Data sources The abnormal category label space The average imbalance ratio of all abnormal category labels in , and obtain all abnormal category labels belonging to the minority class;

[0082] Step 3.1.1: Use formula (2) to calculate the Data sources Abnormal label space Middle Class imbalance ratio of abnormal class labels :

[0083] (2)

[0084] In formula (2), Indicates the i-th data source No. Anomaly category labels, , It is judgment Is it present in In the indicator function, max() represents the maximum value function; It is judgment Is it present in The indicator function in , and has:

[0085] (3)

[0086] Step 3.1.2: Use equation (4) to calculate Data sources Abnormal label space The average imbalance ratio :

[0087] (4)

[0088] In formula (4), when > When Data sources Middle Anomaly category labels is a minority class; otherwise, continue to judge the next abnormal class label, so as to obtain all the abnormal class labels belonging to the minority class.

[0089] Step 3.2: Generate synthetic clinical feature vectors of abnormal class labels belonging to the minority class and add them to the clinical feature matrix In the above example, a synthetic anomaly class label vector belonging to the minority class is generated and added to the anomaly class label matrix and constitute the Data sources ;

[0090] Step 3.2.1: Traverse Data sources Abnormal label space For each abnormal category label in For the minority class, Remove all the The non-missing clinical feature vector is denoted as ,in, for Middle Contains The non-missing clinical feature vector, contain The total number of non-missing clinical feature vectors, .

[0091] Step 3.2.2: Calculate based on Euclidean distance Zhongyu The closest clinical feature vectors and used as The nearest clinical feature vector of ;

[0092] Step 3.2.3: Traversal Each clinical characteristic value in Is it a numeric type? If so, calculate the corresponding clinical characteristic value and The difference between the neighbor clinical feature values ​​at the corresponding positions in the neighbor clinical feature vectors, and The sum of the differences and the product of the random number is used as a synthetic clinical feature value; otherwise, the nearest clinical feature value with the highest frequency in the nearest clinical feature vector is selected and used as the synthetic clinical feature value, so that Composite clinical features Synthetic clinical feature vector , thus obtaining the synthetic clinical feature vector set And join the Clinical feature matrix after data source completion middle;

[0093] choose The number of times the abnormal category vector corresponding to the nearest clinical feature vector appears exceeds The abnormal category label is used as The synthetic anomaly label vector , thus obtaining the synthetic anomaly category label set And join the The abnormal category label matrix of the data sources In the Data sources Class-balanced data source ;

[0094] Order Class-balanced data sources The clinical feature space is denoted as ,in, Represents the i-th class balanced data source No. Clinical characteristic values, ;

[0095] Order Class-balanced data sources The abnormal category label space is recorded as ,in, Represents the i-th class balanced data source No. Anomaly category labels, .

[0096] Step 4: Use the improved infinite feature selection algorithm to select Class-balanced data sources Clinical feature selection was performed to obtain Candidate feature sets on a class-balanced data source ;

[0097] Step 4.1: Calculate clinical feature values ​​based on mutual information and Fisher score and abnormal category labels The correlation between The combined weight of all clinical characteristic values ​​in ;

[0098] Step 4.1.1: Calculate the Class-balanced data sources Middle Clinical characteristic values Variance , clinical characteristic values exist All labels with abnormal categories The mean value of patients , all labels that do not contain abnormal categories The mean value of patients , and clinical characteristic values exist The mean of all patients ;

[0099] Step 4.1.2: Use equation (5) to calculate the Class-balanced data sources Middle Clinical characteristic values With Anomaly category labels Fisher score , thus obtaining Class-balanced data sources Clinical feature space With abnormal label space The correlation matrix :

[0100] (5)

[0101] Step 4.1.3: Preserve the abnormal category label space Middle Anomaly category labels The first m largest Fisher scores are taken, and the other Fisher scores are set to 0, so as to obtain the processed correlation matrix and accumulate them in the column direction, and the accumulated value is taken as the first Clinical characteristic values The weight of the Fisher score Different anomaly class labels will rely on different feature combinations. By calculating the Fisher score for each anomaly class separately and retaining the top m maximum values, it can be ensured that the selected features have a high ability to distinguish each anomaly class, thereby enhancing the class specificity of the features.

[0102] Step 4.1.4: Calculate the first Class-balanced data sources Middle Clinical characteristic values and abnormal category labels Mutual information between :

[0103] (6)

[0104] In formula (6), Clinical characteristic value When the abnormal category label The probability of Clinical characteristic value The probability of Indicates abnormal category label The probability of

[0105] Step 4.1.5: Calculate the Clinical characteristic values and the abnormal category label pair ( ) , No. Clinical characteristic values and abnormal category labels and The ternary mutual information between And known abnormal labels In the case and The conditional mutual information ,in, Indicates Class-balanced data sources Middle Anomaly category labels .

[0106] Step 4.1.6: Use equation (7) to calculate the Class-balanced data sources Middle Clinical characteristic values The mutual information weight :

[0107] (7)

[0108] In formula (7), Indicates Clinical characteristic values The weight of is obtained by formula (8):

[0109] (8)

[0110] Step 4.1.7: Calculate the Class-balanced data sources No. Clinical characteristic values The standard deviation of ;

[0111] Step 4.1.8: Calculate using formula (9) The comprehensive weight of , thus obtaining The combined weight of all clinical characteristic values ​​in ;

[0112] (9)

[0113] In formula (9), are 3 hyperparameters, and ; is a measure of the correlation between the clinical feature value and the abnormal class label, and It is a measure of the amount of information in the clinical feature value itself. It comprehensively considers the importance of the clinical feature value from multiple angles, thereby improving the accuracy of clinical feature selection.

[0114] Step 4.2: Calculate simultaneous selection and The amount of information that can be generated , thus constructing the Class-balanced data sources Special Information Matrix ,in, Indicates Class-balanced data sources Clinical characteristic value The comprehensive weight of

[0115] Step 4.3: Use equation (10) to calculate the Class-balanced data sources The feature importance matrix on :

[0116] (10)

[0117] In formula (10), r is a hyperparameter term, is the unit matrix, and the above formula can capture the characteristic information matrix The high-order information can be obtained through the computation, and the calculation is convenient and efficient.

[0118] Step 4.4: Calculate the feature importance matrix The sum of each column in the table is calculated, and the clinical characteristic value with the largest sum of each column in the first k columns is selected as the first Class-balanced data sources The candidate feature set , thus obtaining balanced data sources for all classes The candidate feature set ,in, express The candidate feature values, .

[0119] Step 5: Traverse the candidate feature set All candidate features of Perform redundancy judgment between sources to obtain a candidate feature set that eliminates redundancy , thus obtaining a clinical feature subset without redundant features , used to judge abnormalities in the patient's clinical characteristic data;

[0120] (11)

[0121] In formula (11), Indicates Class-balanced data sources The abnormal category label space, Indicates Class-balanced data sources The candidate feature set The candidate feature values, , Indicates to remove the candidate feature set and The collection composed of Indicates to remove the candidate feature set Candidate feature value The collection composed later; Represents candidate feature values and abnormal category label space The mutual information of .

[0122] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0123] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0124] The effectiveness of the method of the present invention is verified using real-world medical datasets.

[0125] 1) Prepare standard dataset

[0126] The Medical dataset used in this invention is used as a standard dataset to verify the effectiveness of the feature selection method for multi-source multi-label medical clinical data with missing data. The Medical dataset is a shared dataset for multi-label classification of clinical free text. In the Medical dataset, there are 978 pieces of data corresponding to 978 collected patient samples, as well as 1449 clinical features corresponding to 45 labels corresponding to 45 collected abnormalities. We artificially divide it into 3 data sources, each of which may contain different clinical features, and randomly set 30% of the clinical feature data to 0 to indicate missing data.

[0127] 2) Evaluation indicators

[0128] The AUC area and missing rate are used as evaluation indicators in this example. The AUC area is a more suitable evaluation indicator for multi-label classification to evaluate the performance of feature selection implemented by the present invention under different algorithms. The larger the AUC area value, the better the feature selection effect.

[0129] 3) Experiment on standard datasets

[0130] In order to verify the effectiveness of the method proposed in the present invention, the classification algorithm of the present invention is modeled and predicted on the Medical data set with the classic MDMR (max-dependency and min-redundancy) algorithm, MRMD (manifold regularization and dependence maximization) algorithm and MCLS (manifold-based constraint laplacian sort) algorithm in this field, and their AUC areas and data missing situations are compared. The experimental results are shown in Table 1.

[0131] Table 1 Comparative experimental results

[0132]

[0133] It can be seen from Table 1 that in the Medical dataset, the feature selection algorithm of the present invention can obtain higher prediction performance than other algorithms even when there is data missing, thereby reflecting the effectiveness of the present invention.

Claims

1. A method for feature selection of multi-source multi-label medical clinical data with missing data, characterized in that: The steps are as follows: Step 1: Obtain medical clinical data ,in, Indicates Clinical data from various sources, Indicates the number of data sources, and , ,in, Indicates the Data sources The clinical characteristics matrix of patients, express The abnormal category label matrix of , represents the field of real numbers, Indicates Data sources The number of patients, Indicates Data sources The number of clinical features, Indicates the Data sources The clinical feature vector of the jth patient, , Indicates Data sources The jth patient The sth clinical characteristic value, when the Data sources The jth patient The sth clinical characteristic value When missing, ; , express The abnormal category label of , T represents transposition; Let the Data sources The clinical feature space is denoted as ,in, Indicates the i-th data source No. Clinical characteristic values; Let the Data sources The abnormal category label space is recorded as ,in, Indicates the i-th data source No. Anomaly category labels , ; Step 2: Complete the missing clinical feature values ​​in each data source to obtain the The completed clinical feature matrix ,in, Indicates the Data sources The non-missing clinical feature vector of the j-th patient; Step 3: Right Perform multi-label oversampling to obtain class-balanced Data sources ; Step 3.1: Calculate the Data sources The abnormal category label space The average imbalance ratio of all abnormal category labels in , and obtain all abnormal category labels belonging to the minority class; Step 3.2: Generate synthetic clinical feature vectors of abnormal class labels belonging to the minority class and add them to the clinical feature matrix In the above example, a synthetic anomaly class label vector belonging to the minority class is generated and added to the anomaly class label matrix and constitute the Data sources Class-balanced data source ; Let the Class-balanced data sources The clinical feature space is denoted as ,in, Represents the i-th class balanced data source No. Clinical characteristic values, ; Let the Class-balanced data sources The abnormal category label space is recorded as ,in, Represents the i-th class balanced data source No. Anomaly category labels, ; Step 4: Use the improved infinite feature selection algorithm to select Class-balanced data sources Clinical feature selection was performed to obtain the Candidate feature sets on a class-balanced data source ; Step 4.1: Calculate clinical feature values ​​based on mutual information and Fisher score and abnormal category labels The correlation between The combined weight of all clinical characteristic values ​​in ; Step 4.2: Calculate simultaneous selection and The amount of information that can be generated , thereby constructing the Class-balanced data sources Special Information Matrix ,in, Indicates the Class-balanced data sources Clinical characteristic value The comprehensive weight of Step 4.3: Use formula (10) to calculate the Class-balanced data sources The feature importance matrix on : (10) In formula (10), r is a hyperparameter term, is the identity matrix; Step 4.4: Calculate the feature importance matrix The sum of each column in the first k columns is selected as the clinical characteristic value with the largest sum of each column. Class-balanced data sources The candidate feature set , thus obtaining balanced data sources for all classes The candidate feature set ,in, express The candidate feature values, ; Step 5: Traverse the candidate feature set All candidate features of Perform redundancy judgment between sources to obtain a candidate feature set that eliminates redundancy , thus obtaining a clinical feature subset without redundant features , used to judge abnormalities in the patient's clinical characteristic data; (11) In formula (11), Indicates Class-balanced data sources The abnormal category label space, Indicates the Class-balanced data sources The candidate feature set The candidate feature values, , Indicates that the candidate feature set is removed and The collection composed of Indicates that the candidate feature set is removed Candidate feature value The collection composed later; Represents candidate feature values With the abnormal category label space The mutual information of .

2. A method for selecting features of multi-source and multi-label medical clinical data containing missing data according to claim 1, characterized in that: The step 2 is performed as follows: Step 2.1: Calculate the i-th data source by cosine similarity The correlation between the clinical feature values ​​of the Laplacian matrix The Laplacian matrix of the graph of the association between and anomaly labels ; Step 2.2: Use equation (1) to solve the i-th data source Clinical feature correlation matrix and the i-th data source Upper clinical feature space To the abnormal label space The mapping matrix ,and ,in, Represents the i-th data source The sth clinical characteristic value With Clinical characteristic values The correlation between , , let the i-th data source The sth clinical characteristic value Its own relevance : (1) In formula (1), , , , There are 4 hyper parameters respectively; Step 2.3: Data sources Patient clinical characteristics matrix Right multiplication Get the first The completed clinical feature matrix ,in, Indicates the Data sources The non-missing clinical feature vector of the j-th patient.

3. A method for selecting features of multi-source and multi-label medical clinical data containing missing data according to claim 2, characterized in that: The step 3.1 is performed as follows: Step 3.1.1: Use formula (2) to calculate the Data sources Abnormal label space Middle Class imbalance ratio of abnormal class labels : (2) In formula (2), Indicates the i-th data source No. Anomaly category labels, , It is judgment Is it present in In the indicator function, max() represents the maximum value function; It is judgment Is it present in The indicator function in , and has: (3) Step 3.1.2: Use formula (4) to calculate the Data sources Abnormal label space The average imbalance ratio : (4) In formula (4), when > , indicating that Data sources Middle Anomaly category labels is a minority class; otherwise, continue to judge the next abnormal class label, so as to obtain all the abnormal class labels belonging to the minority class.

4. A method for selecting missing multi-source multi-label medical clinical data features according to claim 3, characterized in that: Step 3.2 is performed as follows: Step 3.2.1: Traverse the Data sources Abnormal label space For each abnormal category label in For the minority class, Remove all the The non-missing clinical feature vector is denoted as ,in, for Middle Contains The non-missing clinical feature vector, contain The total number of non-missing clinical feature vectors, ; Step 3.2.2: Calculate based on Euclidean distance Zhongyu The closest clinical feature vectors and serve as The nearest clinical feature vector of ; Step 3.2.3: Traversal Each clinical characteristic value in Is it a numeric type? If so, calculate the corresponding clinical characteristic value and The difference between the neighbor clinical feature values ​​at the corresponding positions in the neighbor clinical feature vectors, and The sum of the differences and the product of the random number is used as a synthetic clinical feature value; otherwise, the nearest clinical feature value with the highest frequency in the nearest clinical feature vector is selected and used as the synthetic clinical feature value, so that Composite clinical features Synthetic clinical feature vector , thus obtaining the synthetic clinical feature vector set And add the Clinical feature matrix after data source completion middle; choose The number of times the abnormal category vector corresponding to the nearest clinical feature vector appears exceeds The abnormal category label is used as The synthetic anomaly label vector , thus obtaining the synthetic anomaly category label set And add the The abnormal category label matrix of the data sources In the above, we obtain Data sources .

5. A method for selecting features of multi-source and multi-label medical clinical data containing missing data according to claim 4, characterized in that: Step 4.1 is performed as follows: Step 4.1.1: Calculate the Class-balanced data sources Middle Clinical characteristic values Variance , clinical characteristic values exist All labels with abnormal categories The mean value of patients , all labels that do not contain abnormal categories The mean value of patients , and clinical characteristic values exist The mean of all patients ; Step 4.1.2: Use formula (5) to calculate the Class-balanced data sources Middle Clinical characteristic values With Anomaly category labels Fisher score , thus obtaining the Class-balanced data sources Clinical feature space With abnormal label space The correlation matrix : (5) Step 4.1.3: Keep the abnormal category label space Middle Anomaly category labels The first m largest Fisher scores are taken, and the other Fisher scores are set to 0, so as to obtain the processed correlation matrix and accumulate them in the column direction, and the accumulated value obtained is used as the first Clinical characteristic values The weight of the Fisher score ; Step 4.1.4: Calculate the first Class-balanced data sources Middle Clinical characteristic values and abnormal category labels Mutual information between : (6) In formula (6), Clinical characteristic value When the abnormal category label The probability of Clinical characteristic value The probability of Indicates abnormal category label The probability of Step 4.1.5: Calculate the Clinical characteristic values and the abnormal category label pair ( ) , No. Clinical characteristic values and abnormal category labels and The ternary mutual information between And known abnormal labels In the case and The conditional mutual information ,in, Indicates Class-balanced data sources Middle Anomaly category labels ; Step 4.1.6: Use formula (7) to calculate the Class-balanced data sources Middle Clinical characteristic values The mutual information weight : (7) In formula (7), Indicates Clinical characteristic values The weight of is obtained by formula (8): (8) Step 4.1.7: Calculate the Class-balanced data sources No. Clinical characteristic values The standard deviation of ; Step 4.1.8: Calculate using formula (9) The comprehensive weight of , thus obtaining The combined weight of all clinical characteristic values ​​in ; (9) In formula (9), are 3 hyperparameters, and .

6. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports a processor to execute the method according to any one of claims 1 to 5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are performed.

Images