A multi-classification method for mood disorders based on dual feature selection

Abstract

The present invention belongs to the fields of machine learning, data representation, and classification technology, and particularly relates to a multi-classification method for mood disorders based on dual feature selection. The method extracts features from objective data, obtaining a feature data for each piece of objective data; performs dual feature selection on the obtained feature data to obtain data to be classified; the first feature selection of the dual feature selection uses the KS test and the maximum mutual information coefficient (MIC) to calculate a comprehensive score of the features, selecting effective features between each two categories; the comprehensive score is a weighted sum of the difference score and the correlation score; the second feature selection of the effective features between each two categories serves as the data to be classified after the dual feature selection; the data to be classified is trained to obtain a trained model; the test set is fed into the trained model for testing to obtain a classification result; the method relies on feature selection to improve classification accuracy and achieves high-accuracy identification of mood disorders.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, data representation and classification, and in particular relates to a multi-classification method for mood disorders based on dual feature selection. Background Art

[0002] With the development of society and the accelerating pace of people's work and life, the psychological pressure caused by various social factors has caused the prevalence of mental illness among the public to continue to surge. Among the many mental illnesses, depression, bipolar disorder and anxiety disorders are typical types of mood disorders, and have become one of the main risk factors affecting human physical and mental health in this century. According to statistics from the World Health Organization, in 2019, 280 million people and 45 million people worldwide were affected by depression and bipolar disorder respectively. However, incorrect diagnoses are sometimes obtained in the initial consultation because there are currently no clear biomarkers (such as genetic testing) that can accurately diagnose mood disorders.

[0003] There are two common diagnostic tools for mood disorders: diagnostic and treatment manuals and symptom rating scales. These methods are highly subjective, leading to non-objective diagnostic and treatment standards for mood disorders and serious problems of missed and misdiagnoses. Currently, research using objective signals combined with machine learning to identify mood disorders is attracting significant attention.

[0004] As we all know, feature selection is crucial to machine learning. We believe that the features used in most machine learning tasks determine the upper limit of the effect, and the selection and combination of models are only infinitely close to this upper limit. Therefore, effective feature selection and statistical methods are very important and necessary; among them, the KS test (i.e. Kolmogorov-Smirnov test) is a non-parametric test; the maximum mutual information coefficient, i.e., the Maximum Information Coefficient, is referred to as MIC. Using MIC to measure the degree of association between two genes, linear or nonlinear relationship, has higher accuracy than mutual information (MI). MIC is an excellent way to calculate data correlation. This application is committed to exploring feature extraction and feature selection, and then used for classification of mood disorders, and is committed to improving the effect of feature selection, thereby improving the accuracy of classification, and achieving the distinction between various mood disorders and normal people through objective data, assisting clinical diagnosis and treatment and reducing the workload of doctors. Summary of the Invention

[0005] The purpose of the present invention is to address the defect that the feature selection method is poor in effect and leads to low classification accuracy in studies using objective signals combined with machine learning to identify multiple categories of patients with mood disorders and healthy people. A multi-classification method for mood disorders based on dual feature selection is proposed, wherein the number of categories of the multi-classification is not less than 3 and one of the categories is healthy people; the method first extracts feature data from objective data, and the feature data is subjected to dual feature selection to obtain data to be classified, and the data to be classified is classified to obtain a classification result; the objective data includes but is not limited to speech data, magnetic resonance imaging data, functional magnetic resonance imaging data, electroencephalogram data and muscle electricity data; the dual feature selection includes two feature selection processes, wherein the first feature selection uses the KS test and the maximum mutual information coefficient to calculate the comprehensive score of the feature, and selects the effective features between each two categories; the second feature selection is based on the effective features, and the effective features for multi-classification are selected according to the number of times the feature is selected and the ranking; and the ranking is from high to low according to the comprehensive score.

[0006] In order to achieve the above objectives, the following technical solutions are adopted:

[0007] The multi-classification method for mood disorders based on dual feature selection includes:

[0008] S1. Extract features from objective data, and obtain a feature data for each objective data;

[0009] S2, perform double feature selection on the feature data obtained in S1 to obtain data to be classified;

[0010] The dual feature selection includes a first feature selection and a second feature selection;

[0011] The first feature selection uses the KS test and the maximum mutual information coefficient MIC to calculate the comprehensive score of the features and select the effective features between each two categories;

[0012] The comprehensive score is the weighted sum of the difference score and the relevance score;

[0013] The second feature selection screens the effective features between all categories as the data to be classified after the double feature selection;

[0014] S3, divide the data to be classified obtained in S2 into a training set and a test set according to the proportion;

[0015] S4, send the training set obtained in S3 into the classifier for training to obtain a trained model;

[0016] S5: Send the test set obtained in S3 to the trained model in S4 for testing to obtain the classification result.

[0017] In S1, the number of feature data is x * y, and each piece of feature data includes M features; where x is the number of samples, the number of feature data of one person is 1 sample, and y objective data are collected for each person.

[0018] After the first feature selection described in S2, [a(a + 1) / 2] * N effective features are obtained; a is the category of mood disorder diseases and a is greater than or equal to 2; N is the number of features of each piece of data to be classified.

[0019] The second feature selection filters the effective features between all categories. Specifically, based on the [a(a + 1) / 2] * N effective features, N features are further selected as the final effective features for multi-classification, including the following steps:

[0020] S210: Initialize the feature selection count and the number of selected features to 0 respectively;

[0021] S211: Select the features that are selected [a(a + 1) / 2] - p times in the first feature selection in descending order of comprehensive score;

[0022] The feature selection count is denoted as p; the number of features that are selected [a(a + 1) / 2] - p times is q;

[0023] S212: Add q to the number of selected features and determine whether the sum after adding q is equal to N. If so, jump to S3, then the selection of N features in the data to be classified is completed, that is, the data to be classified with N features is selected from [a(a + 1) / 2] * N effective features; otherwise, add 1 to the feature selection count and jump to S211.

[0024] Among the x people described in S1, there are normal people and patients with a type of mood disorder diseases and a is greater than or equal to 2; the i-th piece of feature data is denoted as L i ; where the value range of i is from 1 to x * y.

[0025] The i-th piece of data to be classified corresponding to the i-th piece of feature data is denoted as, and L i1 contains N features; N << M; the N features are obtained by using the method of double feature selection.

[0026] The first feature selection described in S2 includes the following steps:

[0027] S21: Select any two categories from a + 1 categories of data and initialize the feature selection count to 0;

[0028] a is the category of mood disorder diseases and a is greater than or equal to 2;

[0029] S22: Determine the number of effective features between the two selected categories;

[0030] S23, select the feature with the largest inter-class difference as the first valid feature to be selected, and initialize the feature sequence number to 1;

[0031] S24. Calculate the difference scores of the unselected features;

[0032] S25. Calculate the relevance scores of the unselected features;

[0033] S26. Calculate the comprehensive score of the unselected features;

[0034] S27, select the feature with the highest comprehensive score as the next valid feature to be selected;

[0035] S28, add 1 to the feature number, and determine whether the feature number is equal to the valid feature number. If so, jump to S29; otherwise, jump to S24;

[0036] S29, add 1 to the number of feature selections and determine whether the number of feature selections is equal to a(a+1) / 2. If so, jump to S210; otherwise, select the other two categories in the a+1 category data and jump to S22, and the other two categories are not exactly the same as the two categories selected before.

[0037] The difference score in S24 is the logarithm of the difference of the feature.

[0038] The correlation score in S25 is the logarithm of the difference between 1 and the maximum correlation.

[0039] The maximum correlation is the maximum value of the correlation between the feature and the selected features.

[0040] Beneficial effects

[0041] The multi-classification method for mood disorders based on dual feature selection described in the present invention has the following beneficial effects compared with existing methods:

[0042] 1. The method adopts dual feature selection, which can select the most effective features for multi-classification;

[0043] 2. The method can perform feature selection for classification tasks with more than 2 categories;

[0044] 3. The first feature selection uses KS test and MIC to comprehensively consider the feature difference and correlation, and can select the most effective feature between the two categories;

[0045] 4. Compared with the T test and F test, the KS test does not require the data to be normally distributed and has a wider range of applications;

[0046] 5. Compared with the Pearson correlation coefficient, the MIC can extract nonlinear correlation information between data and has a wider range of applications;

[0047] 6. The second feature selection is based on the effective features obtained in the first feature selection. The effective features finally used for multi-classification are selected according to the comprehensive score of the features and the number of times the features are selected in the first feature selection. The best features are selected from the best, and the classification effect is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is an overall flow chart of a multi-classification method for mood disorders based on dual feature selection in the present invention (taking speech data as an example);

[0049] Figure 2 This is a flowchart of a dual feature selection method for multi-classification of mood disorders based on dual feature selection according to the present invention;

[0050] Figure 3 This is a flowchart of the first feature selection in a dual feature selection method for multi-classification of mood disorders based on dual feature selection of the present invention;

[0051] Figure 4 This is a flow chart of the second feature selection in the dual feature selection of the emotional disorder multi-classification method based on dual feature selection of the present invention;

[0052] Figure 5 It is the classification result (confusion matrix) of a multi-classification method for mood disorders based on dual feature selection of the present invention. DETAILED DESCRIPTION

[0053] The following further illustrates and describes in detail the multi-classification method for mood disorders based on dual feature selection according to the present invention in conjunction with the accompanying drawings and examples.

[0054] Example 1

[0055] This embodiment describes in detail the process of classifying a speech dataset containing two mental disorders and healthy people using the method of the present invention.

[0056] The multi-classification method for mood disorders based on dual feature selection includes the following steps:

[0057] S1. Extract features from objective data, and obtain a feature data L for each objective data. i ;

[0058] The objective data includes but is not limited to speech data, magnetic resonance imaging data, functional magnetic resonance imaging data, electroencephalogram data and muscle electrical data;

[0059] The objective data comes from x individuals, i.e., x samples; each of the x individuals obtains y pieces of objective data; the x individuals include normal people and patients with type A mood disorders; a is not less than 2; the characteristic data L i Contains M features; where the value of i ranges from 1 to x*y;

[0060] S2, the x*y feature data L obtained in S1 i Perform double feature selection to obtain x*y pieces of data to be classified L i1 ;

[0061] The data to be classified L i1 Contains N features, and N< <M;

[0062] The N features are obtained using the dual feature selection method;

[0063] The dual feature selection includes two feature selection processes, namely the first feature selection and the second feature selection;

[0064] The first feature selection uses KS test and maximum mutual information coefficient (MIC) to calculate the comprehensive score of the features and select effective features between the two types of data;

[0065] The first feature selection includes the following steps:

[0066] S2.1. Select any two categories from the a+1 category data and initialize the number of feature selections to 0;

[0067] S2.2, determine the number N of valid features between the two selected types of data;

[0068] S2.3. Select the feature with the largest inter-class difference as the first valid feature to be selected, and initialize the feature number to 1;

[0069] S2.4. Calculate the difference scores of the features that were not selected;

[0070] The difference score is the logarithm of the p-value of the KS test of the feature, expressed as log(d), where d is the p-value of the KS test of the feature;

[0071] S2.5. Calculate the relevance scores of the features that were not selected;

[0072] The correlation score is the logarithm of the difference between 1 and the maximum correlation, expressed as log(1-c), where c is the maximum correlation; the maximum correlation is the maximum value of the MIC between the feature and the selected features;

[0073] S2.6. Calculate the comprehensive score of the features that were not selected;

[0074] The comprehensive score is the weighted sum of the difference score and the relevance score, expressed as log(d)g1+log(1-c)g2, where g1 and g2 are the weights of the difference score and the relevance score, respectively;

[0075] S2.7. Select the feature with the highest comprehensive score as the next valid feature to be selected;

[0076] S2.8. Increase the feature number by 1 and determine whether the feature number is equal to the number of features N. If so, skip to S2.9; otherwise, skip to S2.4.

[0077] S2.9. Add 1 to the number of feature selections and determine whether the number of feature selections is equal to a(a+1) / 2. If so, jump to S2.10. Otherwise, select the other two categories in the a+1 category data and jump to S2.2. The other two categories are not exactly the same as the two categories selected previously.

[0078] After the first feature selection, [a(a+1) / 2]*N valid features are obtained;

[0079] The second feature selection further selects N features as the final effective features for multi-classification based on the [a(a+1) / 2]*N effective features;

[0080] The second feature selection includes the following steps:

[0081] S2.10, initialize the number of feature selections p = 0, the number of selected features n = 0;

[0082] S2.11. Select the features selected [a(a+1) / 2]-p times in the first feature selection according to the comprehensive score from high to low. The number of selected features is q.

[0083] S2.12. Determine whether n = n + q is equal to N. If so, jump to S3. Otherwise, p = p + 1, jump to S2.11.

[0084] So far, the data to be classified L has been completed i1 Selection of N features in ;

[0085] S3, the x*y pieces of data to be classified L obtained in S2 i1 Divide into training set and test set according to the proportion;

[0086] S4, send the training set obtained in S3 into the classifier for training to obtain a trained model;

[0087] S5: Send the test set obtained in S3 to the trained model in S4 for testing to obtain the classification result.

[0088] The speech dataset contains speech data from 150 subjects, including 50 patients with major depressive disorder (MDD), 50 patients with bipolar disorder (BD), and 50 healthy subjects (HC). Each person read 7 passages according to the paradigm and recorded 7 speech passages accordingly.

[0089] The classification method is verified using the speech data set; the overall flow chart of the classification method is as follows Figure 1 As shown; the flowchart is divided into a training phase and a testing phase; in the training phase, feature data is first extracted from the speech data in the training set, and then the feature data is subjected to dual feature selection to obtain valid features and index numbers of valid features, and then the valid features are used to train the model to obtain a trained model; in the testing phase, feature data is first extracted from the speech data in the test sample, and then the corresponding features are retained according to the index numbers of the valid features to obtain valid features, and then the trained model is used to classify the valid features to obtain classification results.

[0090] The multi-classification method for mood disorders based on dual feature selection comprises the following steps:

[0091] S1. Extract features from the speech data set, and obtain a feature data L for each speech data. i ;

[0092] The speech data set is from 150 people, that is, 150 samples; each of the 150 people has 7 speech data; the 150 people include normal people and patients with 2 types of mood disorders; the feature data L i Contains 6373 features; the value of i ranges from 1 to 1050;

[0093] S2, the 1050 feature data L obtained in S1 i Perform double feature selection to obtain 1050 pieces of data to be classified L i1 ;

[0094] The data to be classified L i1 Contains 120 features;

[0095] The 120 features are obtained using the dual feature selection method;

[0096] The flowchart of the dual feature selection is as follows Figure 2 As shown, it includes the first feature selection and the second feature selection;

[0097] The flowchart of the first feature selection is as follows Figure 3 As shown, the comprehensive score of the features was calculated using KS test and MIC to select the effective features between the two types of data;

[0098] The first feature selection includes the following steps:

[0099] S2.1. Select any two categories from the three categories of data and initialize the number of feature selections to 0;

[0100] S2.2, determining the number of valid features between the two selected data types 120;

[0101] S2.3. Select the feature with the largest inter-class difference as the first valid feature to be selected, and initialize the feature number to 1;

[0102] S2.4. Calculate the difference scores of the features that were not selected;

[0103] The difference score is the logarithm of the p-value of the KS test of the feature, expressed as log(d), where d is the p-value of the KS test of the feature;

[0104] S2.5. Calculate the relevance scores of the features that were not selected;

[0105] The correlation score is the logarithm of the difference between 1 and the maximum correlation, expressed as log(1-c), where c is the maximum correlation; the maximum correlation is the maximum value of the MIC between the feature and the selected features;

[0106] S2.6. Calculate the comprehensive score of the features that were not selected;

[0107] The comprehensive score is the weighted sum of the difference score and the relevance score, expressed as log(d)g1+log(1-c)g2, where g1 and g2 are the weights of the difference score and the relevance score, respectively;

[0108] S2.7. Select the feature with the highest comprehensive score as the next valid feature to be selected;

[0109] S2.8. Increase the feature number by 1 and determine whether the feature number is equal to the feature quantity 120. If so, skip to S2.9; otherwise, skip to S2.4.

[0110] S2.9. Add 1 to the number of feature selections and determine whether the number of feature selections is equal to 3. If so, jump to S2.10. Otherwise, select the other two categories of the three categories of data and jump to S2.2. The other two categories are not exactly the same as the two categories selected previously.

[0111] After the first feature selection, 360 valid features are obtained;

[0112] The flowchart of the second feature selection is as follows Figure 4 It is shown that based on the 360 ​​effective features, 120 features are further selected as the effective features for the final multi-classification;

[0113] The second feature selection includes the following steps:

[0114] S2.10, initialize the number of feature selections p = 0, the number of selected features n = 0;

[0115] S2.11. Select the features that were selected 3-p times in the first feature selection according to the comprehensive score from high to low. The number of selected features is recorded as q.

[0116] S2.12. Determine whether n = n + q is equal to 120. If so, jump to S3; otherwise, p = p + 1, jump to S2.11.

[0117] So far, the data to be classified L has been completed i1 Selection of 120 features;

[0118] S3, the 1050 pieces of data to be classified obtained in S2 i1 Divide into training set and test set in a ratio of 7:3;

[0119] S4, send the training set obtained in S3 into the classifier for training to obtain a trained model;

[0120] S5: Send the test set obtained in S3 to the model trained in S4 for testing to obtain the classification result;

[0121] The classification result is represented by a confusion matrix, which is as follows: Figure 5 As shown; from the confusion matrix, it can be seen that the classification accuracy is 94.0%.

[0122] The above description is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the accompanying drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A multi-classification method for mood disorders based on dual feature selection, characterized in that: include: S1. Extract features from objective data, and obtain a feature data for each objective data; The objective data includes speech data, magnetic resonance imaging data, functional magnetic resonance imaging data, electroencephalogram data and muscle electrical data; S2, perform double feature selection on the feature data obtained in S1 to obtain data to be classified; The dual feature selection includes a first feature selection and a second feature selection; The first feature selection uses the KS test and the maximum mutual information coefficient MIC to calculate the comprehensive score of the features and select the effective features between each two categories; The comprehensive score is the weighted sum of the difference score and the relevance score; The first feature selection includes: S21, select any two categories from the a+1 category data and initialize the number of feature selections to 0; S22, determining the effective number of features between the two selected categories; S23, select the feature with the largest inter-class difference as the first valid feature to be selected, and initialize the feature sequence number to 1; S24. Calculate the difference scores of the unselected features; S25. Calculate the relevance scores of the unselected features; S26. Calculate the comprehensive score of the unselected features; S27, select the feature with the highest comprehensive score as the next valid feature to be selected; S28, add 1 to the feature number, and determine whether the feature number is equal to the valid feature number. If so, jump to S29; otherwise, jump to S24; S29, add 1 to the number of feature selections and determine whether the number of feature selections is equal to a(a+1) / 2. If so, jump to S210; otherwise, select the other two categories in the a+1 category data and jump to S22, and the other two categories are not exactly the same as the two categories selected previously; After the first feature selection in S2, [a(a+1) / 2]*N valid features are obtained; a is the category of mood disorder; N is the number of features of each data to be classified; The second feature selection screens the effective features among all categories, and further selects N features based on the [a(a+1) / 2]*N effective features as the effective features for the final multi-classification as the data to be classified after the dual feature selection; S3, divide the data to be classified obtained in S2 into a training set and a test set according to the proportion; S4, send the training set obtained in S3 into the classifier for training to obtain a trained model; S5: Send the test set obtained in S3 to the trained model in S4 for testing to obtain the classification result.

2. The multi-classification method for mood disorders according to claim 1, characterized in that: include: In S1, the number of feature data is x*y, and each feature data includes M features; where x is the number of samples, the number of feature data for one person is 1 sample, and each person collects y pieces of objective data.

3. The multi-classification method for mood disorders according to claim 1, characterized in that: In S2, a is greater than or equal to 2.

4. The multi-classification method for mood disorders according to claim 1, characterized in that: The second feature selection screens the effective features among all categories, including the following steps: S210, initializing the number of feature selections and the number of selected features to 0; S211. Select the features that were selected [a(a+1) / 2]-p times in the first feature selection according to the comprehensive scores from high to low; The number of feature selections is denoted as p; the number of features selected [a(a+1) / 2]-p times is q; S212, add q to the number of selected features and determine whether the sum after adding q is equal to N. If so, jump to S3, and the selection of N features in the data to be classified is completed, that is, N features of the data to be classified are selected from [a(a+1) / 2]*N valid features; otherwise, add 1 to the number of feature selections and jump to S211.

5. The multi-classification method for mood disorders according to claim 2, characterized in that: S1 The x people include normal people and patients with type a mood disorder and a is greater than or equal to 2; the i-th feature data is recorded as L i ; The value range of i is 1 to x*y.

6. The multi-classification method for mood disorders according to claim 5, characterized in that: The i-th piece of data to be classified corresponding to the i-th piece of feature data is denoted as, and L i1 includes N features; N << M; the N features are obtained by using the method of dual feature selection.

7. The multi-classification method for mood disorders according to claim 1, characterized in that: The difference score in S24 is the logarithm of the difference of the feature.

8. The multi-classification method for mood disorders according to claim 1, characterized in that: The correlation score in S25 is the logarithm of the difference between 1 and the maximum correlation.

9. The multi-classification method for mood disorders according to claim 8, characterized in that: The maximum correlation is the maximum value of the correlation between the feature and the selected features.

Images