Autism early screening system and device based on MRMR feature selection

Abstract

The invention relates to an autism early screening system and device based on MRMR feature selection, and the system integrates a data preprocessing module, a data coding module, a data imbalance processing module, a feature filtering module, a feature selection module and a model construction module. Text is converted into numerical values through coding, and then a data type proportion is balanced by applying a strategy. Subsequently, the feature filtering and selection process ensures that only highly correlated and low-redundancy features are retained, forming an optimized feature set. And finally, based on the feature set, systematically constructing an autism prediction model integrating a plurality of prediction classifiers, and realizing accurate autism prediction by training each classifier and selecting the optimal classifier. The invention aims to improve the accuracy and efficiency of early screening of autism and provide powerful support for early intervention treatment of child patients.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to an early screening system and device for autism based on MRMR feature selection. Background Art

[0002] CHAT-23 (Checklist for Autism in Toddlers, 23 Months) is a widely used early screening scale for autism, and its effectiveness has been verified in clinical practice. The scale helps medical staff to preliminarily assess whether a child is at risk of autism through a series of questions targeting the behavioral characteristics of infants and young children. However, the traditional CHAT-23 screening process relies on manual judgment, which is not only time-consuming, but also may be affected by subjective factors, resulting in limited accuracy and efficiency of the screening results.

[0003] Many scholars have made useful attempts in the study of machine learning for early diagnosis of autism. For example, Thabtah et al. used linear regression to analyze the significant differences in different characteristics between adults and adolescents with autism; Duda et al. recorded multiple behavioral characteristics and built multiple machine learning models to distinguish between autism and attention deficit hyperactivity disorder; Abbas et al. combined parent-report questionnaires and home video behavioral characteristics to create a low-cost and accurate autism screening tool based on a random forest model. These studies provide valuable experience and reference for the application of machine learning in early screening of autism.

[0004] However, it is worth noting that most of the current machine learning research on early screening for autism is based on foreign data, and there is still relatively little early screening work for domestic children with autism. Due to differences in geography, culture, and social environment, the behavioral manifestations of children with autism may vary. Therefore, it is particularly important to build scale tools and machine learning models suitable for local children. The problems with existing early screening devices for autism are as follows:

[0005] First, there are few early screening tools for autism in China;

[0006] Second, the CHAT-23 scale tool not only requires the respondents to complete all 23 questions in the questionnaire, but also requires medical staff to manually make a positive judgment on autism based on the standards provided by the CHAT-23 standard, which has low diagnostic efficiency;

[0007] Third, in actual diagnosis, the prediction model built based on the CHAT-23 scale will have a very obvious data imbalance phenomenon. Specifically, the number of negative confirmed objects far exceeds that of positive confirmed objects. The model trained by this sample data will be more inclined to learn the characteristics of negative samples, resulting in a higher prediction success rate for negative samples, but the prediction effect on positive samples is greatly reduced, thereby reducing the overall prediction accuracy. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides an early screening system for autism based on MRMR feature selection, comprising the following modules:

[0009] A data preprocessing module, wherein the data preprocessing module obtains sample data that meets a specified age condition;

[0010] A data encoding module, wherein the data encoding module encodes the text-type sample data into numerical data based on the sample data;

[0011] A data imbalance processing module, which adjusts the quantity ratio of different types of data based on the numerical data to obtain balanced sample data;

[0012] A feature filtering module, wherein the feature filtering module removes sample data whose characterization capability for the final diagnosis result is lower than a threshold value based on the sample data after the balancing process, and obtains highly correlated feature data;

[0013] A feature selection module, wherein the feature selection module selects feature data with maximum correlation and minimum redundancy based on the strongly correlated feature data, and constructs a feature sequence;

[0014] A model building module, wherein the model building module builds an autism prediction model, wherein the autism prediction model integrates multiple prediction classifiers, trains the multiple prediction classifiers separately based on the feature sequence, selects the prediction classifier with the best performance according to the evaluation index, and uses its output result as the disease prediction result of the autism prediction model.

[0015] In one embodiment of the present invention, the model building module is also used to determine the optimal number of features, including: based on the feature sequence, adding features to the best performing prediction classifier in order according to their importance to the prediction results, so as to determine the number of features of the prediction classifier when it performs best.

[0016] In one embodiment of the present invention, encoding the text-type sample data into numerical data specifically includes: encoding the frequency words "never, occasionally, sometimes, often" in the sample data as 0, 1, 2, and 3 respectively, and encoding the result of diagnosing whether the subject has autism symptoms as 0 or 1, "0" being a negative result, indicating that no autism symptoms exist, and "1" being a positive result, indicating that autism symptoms exist.

[0017] In one embodiment of the present invention, the balanced sample data specifically includes: counting the number of negative sample data and positive sample data, randomly selecting a larger number of sample data using an adaptive sampling synthesis downsampling method, and for a smaller number of sample data, copying or synthesizing new sample data using a boundary synthesis oversampling method, and finally obtaining negative sample data and positive sample data after the number ratio is balanced; wherein the negative sample data is the sample data of the negative result, and the positive sample data is the sample data of the positive result.

[0018] In one embodiment of the present invention, obtaining the strongly correlated feature data specifically includes: based on the sample data after the balancing process, first calculating the variance of each feature data, removing the feature data with a variance lower than a threshold, and then calculating the P value of each feature data after the removal operation, removing all feature data with a P value greater than 0.05, to obtain the strongly correlated feature data.

[0019] In one embodiment of the present invention, constructing a characteristic sequence specifically includes:

[0020] Define c as the output category, and select m data sets S with the greatest correlation from the strongly correlated feature data according to formula (1):

[0021]

[0022] Where D is the feature f in the dataset S i and the average mutual information between the output category c, I(f i ,c) represents feature f i and the mutual information between the output category c; |S| represents the number of selected features in the dataset S;

[0023] According to formula (2), calculate any feature f in the data set S m and another feature f n The redundancy R between

[0024]

[0025] Among them, I(f m ,f n ) represents the feature fm and feature f n The mutual information of

[0026] Based on the redundancy R, formula (3) is constructed, and according to formula (3), a feature set S' with maximum correlation and minimum redundancy is obtained from the m data sets S with the largest average mutual information:

[0027] MRMR=max(DR) (3);

[0028] Based on the feature set S' with the maximum correlation and the minimum redundancy, a feature sequence is constructed.

[0029] In one embodiment of the present invention, the feature f i and the mutual information between the output category c I(f i ,c) or the feature f m and the feature f n The mutual information I(f m ,f n ) are obtained by the following formula:

[0030]

[0031] Among them, p(x,y) is the joint probability density of variables x and variables y, p(x) is the probability density of variable x, and p(y) is the probability density of variable y.

[0032] In one embodiment of the present invention, the feature sequence sorts the features according to their importance to the prediction result.

[0033] In one embodiment of the present invention, the evaluation index includes specificity and sensitivity. The closer the value of the sensitivity is to 1, the lower the missed diagnosis rate of the model prediction result as positive. The closer the value of the specificity is to 1, the lower the misdiagnosis rate of the model.

[0034] The present invention also provides an early screening device for autism based on MRMR feature selection, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, it can be used to run the early screening system for autism based on MRMR feature selection.

[0035] The above technical solution of the present invention has the following advantages compared with the prior art:

[0036] The present invention constructs an early screening model for autism in domestic children. Under the premise of maintaining the prediction effect equivalent to that of the original expert method, it can complete the prediction work of early screening for autism using only less than half of the question features. While reducing the workload of medical staff, it increases the number of questions required to build the model and improves the efficiency of early screening. In view of data imbalance, adaptive synthetic sampling technology is used to alleviate the impact of negative and positive imbalances in actual data on the prediction of model results. In summary, the present invention improves the accuracy of autism screening and can assist doctors in discovering signs of autism in patients earlier and intervening in treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein

[0038] Figure 1 is a structural schematic diagram of an autism early screening system based on MRMR feature selection provided in an embodiment of the present invention;

[0039] Figure 2 It is a relationship diagram between the number of features, specificity and sensitivity in the LightGBM model in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0041] Embodiment 1

[0042] Reference Figure 1 As shown, the present invention provides an early screening system for autism based on MRMR feature selection, comprising the following modules:

[0043] A data preprocessing module, wherein the data preprocessing module obtains sample data that meets the 3-month age condition of the infant autism screening scale CHAT-2;

[0044] A data encoding module, wherein the data encoding module encodes the text-type sample data into numerical data based on the sample data;

[0045] A data imbalance processing module, which adjusts the quantity ratio of different types of data based on the numerical data to obtain balanced sample data;

[0046] A feature filtering module, wherein the feature filtering module removes sample data whose characterization capability for the final diagnosis result is lower than a threshold value based on the sample data after the balancing process, and obtains highly correlated feature data;

[0047] A feature selection module, wherein the feature selection module selects feature data with maximum correlation and minimum redundancy based on the strongly correlated feature data, and constructs a feature sequence;

[0048] A model building module, wherein the model building module builds an autism prediction model, wherein the autism prediction model integrates multiple prediction classifiers, and the multiple prediction classifiers are trained separately based on the feature sequence, and the best performing prediction classifier is selected according to the evaluation index, and its output result is used as the disease prediction result of the autism prediction model; at the same time, based on the feature sequence, features are added to the best performing prediction classifier in order according to their importance to the prediction result, so as to determine the number of features of the prediction classifier when the performance is the best.

[0049] Furthermore, the data preprocessing module determines the respondents who meet the age requirements of the CHAT-23 infant autism screening scale, collects the questionnaire contents filled out by the respondents, filters out data that do not meet the model construction conditions (for example, incomplete answers to questions, age does not meet the CHAT-23 question standards, etc.), and obtains the original sample data for subsequent construction of an autism prediction model.

[0050] In this embodiment, the data encoding module uniformly encodes the original sample data that conforms to the model construction, and encodes the text-type sample data into numerical data, specifically including: according to the different frequency words in the sample data, "never", "occasionally", "sometimes", and "often" are encoded as 0, 1, 2, and 3 respectively, and the result of diagnosing whether the subject has autism symptoms is encoded as 0 or 1, "0" is a negative result, indicating that there are no autism symptoms, and "1" is a positive result, indicating that there are autism symptoms.

[0051] Specifically, since the number of negative confirmed objects in the original sample data far exceeds that of positive confirmed objects, the model trained with this sample data will be more inclined to learn the characteristics of negative samples, resulting in a higher performance of the model in terms of negative prediction success rate, but a poor prediction effect on positive samples. Therefore, the data imbalance processing module adjusts the quantitative proportion relationship of different types of data based on the numerical data to obtain balanced sample data, specifically including: counting the number of negative sample data and positive sample data, using an adaptive sampling synthesis downsampling method to randomly select sample data with a large number, and for sample data with a small number, using a boundary synthesis oversampling method to copy or synthesize new sample data, and finally obtaining negative sample data and positive sample data after quantitative proportion balance processing.

[0052] Furthermore, the feature filtering module removes the sample data whose characterization ability for the final diagnosis result is lower than a threshold value based on the sample data after the balancing process, and obtains the highly correlated feature data, which specifically includes:

[0053] Based on the sample data after the balancing process, the variance of each feature data is first calculated, and the feature data with variance lower than the threshold is removed. Then, the P value of each feature data after the removal operation is calculated, and all feature data with P values ​​greater than 0.05 are removed to obtain the strongly correlated feature data.

[0054] In this embodiment, the feature selection module selects feature data with maximum correlation and minimum redundancy based on the strongly correlated feature data by using a maximum correlation and minimum redundancy (MRMR) algorithm to construct a feature sequence, specifically including:

[0055] Define c as the output category, and select m data sets S with the greatest correlation from the strongly correlated feature data according to formula (1):

[0056]

[0057] Where D is the feature f in the dataset S i and the average mutual information between the output category c, I(f i ,c) represents feature f i and the mutual information between the output category c; |S| represents the number of selected features in the dataset S;

[0058] According to formula (2), calculate any feature f in the data set S m and another feature f n The redundancy R between

[0059]

[0060] Among them, I(f m ,f n ) represents the feature f m and feature f n The mutual information of

[0061] Based on the redundancy R, formula (3) is constructed, and according to formula (3), a feature set S' with maximum correlation and minimum redundancy is obtained from the m data sets S with the largest average mutual information:

[0062] MRMR=max(DR) (3);

[0063] Based on the feature set S' with the maximum correlation and the minimum redundancy, the features are sorted according to their importance to the prediction result to construct a feature sequence.

[0064] In this embodiment, the feature f i and the mutual information between the output category c I(f i ,c) or the feature f m and the feature f n The mutual information I(f m ,f n ) are obtained by the following formula:

[0065]

[0066] Among them, p(x,y) is the joint probability density of variables x and variables y, p(x) is the probability density of variable x, and p(y) is the probability density of variable y.

[0067] In this embodiment, the evaluation index includes specificity and sensitivity. The closer the value of the sensitivity is to 1, the lower the missed diagnosis rate of the model prediction result is positive. The closer the value of the specificity is to 1, the lower the misdiagnosis rate of the model.

[0068] The effectiveness of the system of the present invention is described below using specific examples:

[0069] The data preprocessing module was used to remove missing values ​​and data with an age below 18 or above 24 months from the 657 data obtained, and the data of children aged 18 to 24 months who met the CHAT-23 age conditions were obtained, including 48 positive data that could be used for the experiment, 323 negative data, and a total of 371 questionnaire data.

[0070] For the collected questionnaire information, there are four different frequency words to record the information of the tested subjects, which are never, occasionally, sometimes, and often. The frequency words "never, occasionally, sometimes, and often" in the 371 text-type questionnaire data are encoded as 0, 1, 2, and 3 respectively through the data encoding module. For whether the patient has autism tendency, 0 and 1 are used to replace the original negative and positive results respectively.

[0071] Since the ratio of negative data to positive data is close to 7:1, the data imbalance processing module uses up-sampling and down-sampling techniques to balance the two, setting the parameter of the adaptive sampling synthesis down-sampling strategy to 0.3, retaining about 160 random negative samples. Combined with the boundary synthesis oversampling technology, some positive samples are copied or synthesized, and the parameter of the up-sampling strategy is set to 0.5, so that about 80 positive samples can be obtained.

[0072] The feature filtering module first sets the variance threshold to 0.1 for the initial 23 problem features and filters out two features based on a statistical method. There are two features with variances below 0.1, namely Q12 and Q16. The features with variances below 0.1 are removed, and the remaining 21 features are retained. The chi-square method is used to calculate the P value of each feature. Considering the correlation between the feature and the diagnosis result, the features with a P value less than 0.05 are considered to be important for the diagnosis result. All features with a P value greater than 0.05 (Q1, Q3, Q4, Q10, Q11, Q18) are removed, and finally 15 features are retained.

[0073] The feature selection module uses the MRMR feature selection method for stepwise feature selection. The goal of MRMR is to give the features that are most relevant to the diagnosis results and have the least redundancy among all features. In MRMR, the MIQ parameter is set to calculate the correlation and difference between features. MIQ calculates the correlation between two features through mutual information and measures the degree of confusion between multiple feature groups through information entropy. According to the MRMR feature selection method, a feature sequence as shown in Table 1 can be obtained, and the features at the front of the sequence are considered to be more important.

[0074] Table 1

[0075]

[0076] The model building module determines the optimal prediction model among the 7 models and determines the optimal number of features. Among them, the steps of determining the optimal prediction model among the 7 models are: using 7 supervised machine learning models to model autism prediction, including two non-integrated learning models: support vector machine (SVM) and decision tree (DT), four integrated learning models: gradient boosted decision tree (GBDT), lightweight gradient boosting machine (LightGBM), random forest (RF) and extreme gradient boosting (XGBoost), and an enhancement algorithm: adaptive boosting (AdaBoost). Before training, feature selection was first performed to determine the most important feature combination as the input of the machine learning model. Subsequently, the trained model was used to predict all test data. Specificity and sensitivity were used as metrics for evaluating the prediction results of autism. Among the tested models, the specificity of LightGBM was 0.933, significantly better than SVM, DT, AdaBoost, RF and XGBoost, while the specificity of GBDT was 0.778. In terms of sensitivity, LightGBM scored 0.856, slightly lower than the 0.911 of the GBDT model. Therefore, after considering the two indicators comprehensively, LightGBM was selected as the basic model. Its corresponding specificity and sensitivity performance are as follows Figure 2 shown.

[0077] The steps to determine the optimal number of features are as follows: In order to determine and select the optimal number of important features for training, the number of input features is gradually increased from 1 to 15, adding them in the order of the MRMR feature selection method described in the feature selection section. The LightGBM model that performs best on this task is used as the base classifier, and different numbers of features are input. The specificity and sensitivity results under different numbers of features are shown in Table 2.

[0078] Table 2

[0079]

[0080]

[0081] From the experimental data in Table 2, it can be seen that when the first 9 features are used as input, the sensitivity of the LightGBM model reaches a peak of 0.909, and the specificity is 0.922. As the number of features gradually increases to 9, the specificity shows an overall upward trend, but it does not continue to increase with the increase in the number of features. When the number of features is less than 5, although the specificity is high, the sensitivity is very low and fails to exceed 0.850. When the number of features exceeds 9, the sensitivity cannot reach above 0.900. Therefore, after comprehensive consideration, the first 9 most important features of the MRMR selection sequence are selected to construct the final prediction model. These 9 features are Q21, Q6, Q9, Q13, Q7, Q14, Q5, Q15 and Q17.

[0082] Embodiment 2

[0083] The present invention also provides an early screening device for autism based on MRMR feature selection, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can be used to run the early screening system for autism based on MRMR feature selection described in Example 1.

[0084] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0085] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0086] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0087] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0088] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. An early screening system for autism based on MRMR feature selection, characterized in that: include: A data preprocessing module, wherein the data preprocessing module obtains sample data that meets a specified age condition; A data encoding module, wherein the data encoding module encodes the text-type sample data into numerical data based on the sample data; A data imbalance processing module, which adjusts the quantity ratio of different types of data based on the numerical data to obtain balanced sample data; A feature filtering module, wherein the feature filtering module removes sample data whose characterization capability for the final diagnosis result is lower than a threshold value based on the sample data after the balancing process, and obtains highly correlated feature data; A feature selection module, wherein the feature selection module selects feature data with maximum correlation and minimum redundancy based on the strongly correlated feature data, and constructs a feature sequence; A model building module, wherein the model building module builds an autism prediction model, wherein the autism prediction model integrates multiple prediction classifiers, trains the multiple prediction classifiers separately based on the feature sequence, selects the prediction classifier with the best performance according to the evaluation index, and uses its output result as the disease prediction result of the autism prediction model.

2. The autism early screening system based on MRMR feature selection according to claim 1, characterized in that: The model building module is also used to determine the optimal number of features, including: based on the feature sequence, adding features to the best performing prediction classifier in order according to their importance to the prediction results, so as to determine the number of features of the prediction classifier when it performs best.

3. The autism early screening system based on MRMR feature selection according to claim 1, characterized in that: The encoding of the text-type sample data into numerical data specifically includes: encoding the frequency words "never, occasionally, sometimes, often" in the sample data as 0, 1, 2, and 3 respectively, and encoding the result of diagnosing whether the subject has autism symptoms as 0 or 1, "0" is a negative result, indicating that there are no autism symptoms, and "1" is a positive result, indicating that there are autism symptoms.

4. The autism early screening system based on MRMR feature selection according to claim 3, characterized in that: The obtained balanced sample data specifically includes: counting the number of negative sample data and positive sample data, randomly selecting a larger number of sample data using an adaptive sampling synthesis downsampling method, and for a smaller number of sample data, copying or synthesizing new sample data using a boundary synthesis oversampling method, and finally obtaining negative sample data and positive sample data after the number ratio is balanced; wherein the negative sample data is the sample data of the negative result, and the positive sample data is the sample data of the positive result.

5. The autism early screening system based on MRMR feature selection according to claim 1, characterized in that: The obtaining of the strongly correlated characteristic data specifically includes: based on the sample data after the balancing process, first calculating the variance of each characteristic data, removing the characteristic data whose variance is lower than a threshold, and then calculating the P value of each characteristic data after the removal operation, removing all characteristic data with a P value greater than 0.05, to obtain the strongly correlated characteristic data.

6. The autism early screening system based on MRMR feature selection according to claim 1, characterized in that: The constructing of the feature sequence specifically includes: Define c as the output category, and select m data sets S with the greatest correlation from the strongly correlated feature data according to formula (1): Where D is the feature f in the dataset S i and the average mutual information between the output category c, I(f i ,c) represents feature f i and the mutual information between the output category c; |S| represents the number of selected features in the dataset S; According to formula (2), calculate any feature f in the data set S m and another feature f n The redundancy R between Among them, I(f m ,f n ) represents the feature f m and feature f n The mutual information of Based on the redundancy R, formula (3) is constructed, and according to formula (3), a feature set S' with maximum correlation and minimum redundancy is obtained from the m data sets S with the largest average mutual information: MRMR=max(DR) (3); Based on the feature set S' with the maximum correlation and the minimum redundancy, a feature sequence is constructed.

7. The autism early screening system based on MRMR feature selection according to claim 6, characterized in that: The feature f i and the mutual information between the output category c I(f i ,c) or the feature f m and the feature f n The mutual information I(f m ,f n ) are obtained by the following formula: Among them, p(x,y) is the joint probability density of variables x and variables y, p(x) is the probability density of variable x, and p(y) is the probability density of variable y.

8. The autism early screening system based on MRMR feature selection according to claim 6, characterized in that: The feature sequence sorts the features according to their importance to the prediction results.

9. The autism early screening system based on MRMR feature selection according to claim 1, characterized in that: The evaluation index includes specificity and sensitivity. The closer the value of the sensitivity is to 1, the lower the missed diagnosis rate of the model prediction result is positive. The closer the value of the specificity is to 1, the lower the misdiagnosis rate of the model.

10. An early screening device for autism based on MRMR feature selection, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it can be used to run the early screening system for autism based on MRMR feature selection as described in any one of claims 1 to 9.

Images