Risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering

Abstract

The present invention discloses a risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering; the patient's symptoms and medical test results are input through a data input module to form symptom and medical test result features; the features suitable for binary classification machine learning problems are generated through an automatic feature engineering module, and the class conditional probability density of continuous features is evaluated through the submodules of the module, features that do not meet the requirements are discarded, new features are automatically combined and operated to generate new features, new features are screened and retained according to thresholds, and the new feature generation process is recorded; the automatically generated features are trained and the model is saved based on a tree model through a model training module; the risk of interstitial lung disease in patients with rheumatoid arthritis is predicted through a disease prediction module according to the risk prediction model for interstitial lung disease in patients with rheumatoid arthritis and the patient's input data. The present invention performs automatic feature engineering on symptoms and medical test results, which is less time-consuming and more efficient.

Description

Technical Field

[0001] The present invention belongs to the field of medical information technology, and in particular relates to a risk prediction system for interstitial lung lesions in patients with rheumatoid arthritis based on automatic feature engineering. Background Art

[0002] Rheumatoid arthritis is a common chronic autoimmune disease in clinical practice. In severe cases, extra-articular manifestations may occur. Interstitial lung disease is a common manifestation of lung involvement in rheumatoid arthritis, which can increase the risk of death in patients. Therefore, early screening of patients with rheumatoid arthritis and interstitial lung disease has important clinical significance.

[0003] Interstitial lung disease is associated with multiple factors, and traditional single factors or a few factors cannot fully describe its risk. With the help of automatic feature engineering, features can be automatically extracted and combined from a large amount of data to more comprehensively describe the patient's condition. This comprehensiveness is particularly critical for the early screening of patients with rheumatoid arthritis and interstitial lung disease, and the quality of automatic feature engineering plays a decisive role in the final prediction results. Existing automatic feature engineering methods include: automatic feature engineering methods based on optimization model results, automatic feature engineering methods based on meta-learning, etc. The automatic feature engineering method based on the optimization model results automatically combines features to generate new features, and then trains the model with the new features. This method evaluates the quality of new features based on the model prediction results (AUC value or information gain index). The automatic feature engineering method based on meta-learning automatically performs feature engineering on variable features in new machine learning problems by learning the feature engineering methods used in previous machine learning cases.

[0004] However, these methods have some limitations in predicting rheumatoid arthritis with interstitial lung disease. First, the automatic feature engineering method based on meta-learning requires a large number of previous machine learning cases as training sets, and the degree of correlation between these previous cases and the new machine learning problem will affect the quality of the feature engineering results. Secondly, when evaluating the newly generated features, the traditional method is to use the new features to train the model and use the model results as a criterion. When the amount of data is large, some complex models require a long time to train the model, which will reduce the efficiency of automatic feature engineering. In addition, in the process of combining the original features to generate new features, the existing technology will try all possible calculation methods, and will not analyze the original features before the calculation. A large number of useless features will be generated, which will take up a lot of memory space, and the subsequent evaluation of these features will consume more time. There are many clinical features that need to be paid attention to in patients with rheumatoid arthritis, and their medication history is often very complicated. In addition, multiple laboratory tests are required, so a more efficient method is needed to generate features that are highly correlated with labels. Finally, traditional automatic feature engineering methods do not consider subsequent machine learning models during operation, resulting in the generated features not being sufficiently in line with the actual situation, reducing the accuracy and generalization ability of the prediction model. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the prior art and to provide a risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering, comprising a data input module, an automatic feature engineering module, a model training module and a disease prediction module;

[0008] The data input module is used to input the patient's symptoms and medical test result data to form symptom and medical test result characteristics;

[0009] The automatic feature engineering module is used to automatically generate a feature suitable for a binary classification machine learning problem; the automatic feature engineering module includes a continuous feature evaluation submodule for evaluating the class conditional probability density of continuous features, a feature screening submodule for discarding features that do not meet the requirements, a new feature automatic generation submodule for automatically combining operations to generate new features, a candidate feature screening and updating submodule for screening and retaining new features according to a threshold, and an automatic feature engineering reproduction submodule for recording the new feature generation process;

[0010] The model training module is used to train the automatically generated features based on the tree model and save the model;

[0011] The disease prediction module is used to predict the risk of interstitial lung lesions in patients with rheumatoid arthritis based on the risk prediction model of interstitial lung lesions in patients with rheumatoid arthritis and the patient's input data.

[0012] Furthermore, the symptoms and medical test result characteristics constituted by the data input module include continuous characteristics and discrete characteristics; among them, the continuous characteristics include "age", "C-reactive protein test value", and "erythrocyte sedimentation rate test value", and the discrete characteristics include "whether there is a history of smoking" and "whether the rheumatoid factor is positive"; the input label data is whether the patient has concurrent interstitial lung disease.

[0013] Furthermore, in the features generated by the automatic feature engineering module and applicable to the binary classification machine learning problem, the probability density function of the data corresponding to the label value of 0 is separated from the probability density function of the data corresponding to the label value of 1.

[0014] Furthermore, the execution operation of the continuous feature evaluation submodule is as follows: in the training set of the binary classification machine learning problem, the label corresponding to each feature is known, and the class conditional probability density of the feature corresponding to each label is calculated by the Bayesian conditional probability formula and ; By separating the two posterior class conditional probability densities and , thereby optimizing the prior conditional probability between features and labels.

[0015] Furthermore, the execution operation of the feature screening submodule is:

[0016] For discrete features, if the difference between the conditional probabilities of the two labels of the feature is greater than the first threshold ,Right now , then the feature is taken as the original feature; otherwise, the feature is discarded;

[0017] For continuous features, when the quantile difference of the class conditional probability density function is greater than the second threshold , then the feature is used as both the original feature and the candidate feature.

[0018] Furthermore, the new feature automatic generation submodule includes a feature self-operation unit and a feature combination operation unit; the feature self-operation unit performs feature self-operation to generate new features, or the feature combination operation unit performs feature combination operation to generate new features.

[0019] Furthermore, the feature operations themselves include exponential operations of natural bases, logarithmic operations of natural bases and absolute value operations; the feature combinations include combinations of continuous features with continuous features, and combinations of continuous features with discrete features.

[0020] Furthermore, the execution operation of the feature combination operation unit is:

[0021] When a candidate feature is combined with other candidate features or continuous features in the original features, the operations between the features are performed through four arithmetic operations and norm operations;

[0022] When the candidate features are combined with the discrete features in the original features, the operation between the features is performed by combining the continuous features with the discrete features. For addition and subtraction operations, all the data of the discrete features are subtracted by 0.5, that is, [0, 1] is mapped to [-0.5, 0.5], and then the operation is performed with the normalized continuous features. For multiplication, [0, 1] is mapped to [-1, 1] or [1,-1], and then the operation is performed with the normalized continuous features.

[0023] Furthermore, the candidate feature screening and updating submodule screens candidate features based on the quantile difference of the class conditional probability density function, sorts the screened candidate features according to the quantile difference of the class conditional probability density function, and retains the candidate features with the largest quantile difference of the class conditional probability density function. candidate features.

[0024] Furthermore, the automatic feature engineering reproduction submodule is used to record the generation process and method of new features; when reproducing the automatic feature engineering, the automatic feature engineering reproduction submodule performs corresponding operations on the features in the test set to generate the same candidate features.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The present invention optimizes continuous features in binary classification machine learning problems (especially using tree-based prediction models), and proposes an automatic feature engineering method that can automatically extract and combine continuous variables to generate new features.

[0027] The object of optimization of the present invention is the continuous features in the training set; the objective function of optimization is the quantile difference of the class conditional probability density function. When automatically generating new features, the present invention will first determine the "sameness / oppositeness" of the two features, and then based on this, specifically select the combined operation to control the number of features generated in each round, saving the consumption of memory space. For the newly generated features, it is only necessary to evaluate based on the quantile difference, without the need to evaluate based on the model and the prediction results each time after the training model, saving time. The combination of these two methods makes the automatic feature engineering less time-consuming and more efficient.

[0028] By automatically engineering features of symptoms and medical test results and training a tree-based prediction model, a risk prediction system for interstitial lung disease in patients with rheumatoid arthritis was eventually constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0030] Figure 1 A schematic diagram of a risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering provided by an embodiment of the present invention;

[0031] Figure 2 A schematic diagram of the quantile difference of the class conditional probability density function provided by an embodiment of the present invention; wherein, Figure 2 (1) in the figure is a schematic diagram of the probability density function after the original feature is normalized. Figure 2 (2) in is a schematic diagram of the class conditional probability density function of the original feature. Figure 2 (3) in the figure is a schematic diagram of the quantile difference of the class conditional probability density function of the original feature. Figure 2 (4) in the figure is a schematic diagram of the quantile difference of the class conditional probability density function of the new feature;

[0032] Figure 3 A schematic diagram of a screening process for discrete and continuous features provided by an embodiment of the present invention;

[0033] Figure 4 A schematic diagram of a process for automatically combining two features to generate a new feature provided by an embodiment of the present invention;

[0034] Figure 5 A schematic diagram of a candidate feature screening and updating process provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] It should be noted that, in the absence of conflict, the features in the following embodiments and implementations may be combined with each other.

[0037] like Figure 1 As shown, an embodiment of the present invention provides a risk prediction system for interstitial lung lesions in patients with rheumatoid arthritis based on automatic feature engineering, and the system includes: a data input module, an automatic feature engineering module, a model training module and a disease prediction module.

[0038] The data input module is used to input the patient's symptoms and medical test result data to form symptom and medical test result features. These features can be continuous features such as "age", "C-reactive protein test value", "erythrocyte sedimentation rate test value", etc., or discrete features such as "whether there is a history of smoking" and "whether the rheumatoid factor is positive". The input label data is whether the corresponding patient has interstitial lung disease.

[0039] The automatic feature engineering module aims to automatically generate a feature suitable for a binary classification machine learning problem. The probability density function of the data corresponding to the label value of 0 and the probability density function of the data corresponding to the label value of 1 in the feature should be separated from each other as much as possible. The specific schematic diagram is as follows Figure 2 as shown in .

[0040] The process of automatic feature engineering includes: a continuous feature evaluation submodule that evaluates the conditional probability density of continuous features; a feature screening submodule that discards some features that do not meet the requirements; a new feature automatic generation submodule that automatically combines operations to generate new features; a candidate feature screening and updating submodule that selects and retains new features based on thresholds; and an automatic feature engineering reproduction submodule that records the new feature generation process. The specific introduction is as follows:

[0041] 1. Continuous feature evaluation submodule:

[0042] For continuous features, the present invention uses the evaluation criterion of "quantile difference of class conditional probability density function". This evaluation criterion is one of the key parts of the present invention, which will be explained below.

[0043] In the training set of a binary classification machine learning problem, the label corresponding to each feature is known, and the class conditional probability density of the feature corresponding to each label can be calculated and . According to the Bayesian conditional probability formula, we can get:

[0044]

[0045]

[0046] The maximization expression for the difference of prior conditional probabilities can be expressed as:

[0047]

[0048] can be transformed into:

[0049]

[0050] in, Indicates that given a feature Under the condition of , the posterior probability of the sample label category is 0, Indicates that given a feature Under the condition of , the posterior probability that the label category of the sample is 1, Indicates that when the sample label is 0, the feature The class conditional probability density of Indicates that when the sample label is 1, the feature The class conditional probability density of It indicates the prior probability that the sample label is 0 in the entire data set. It represents the prior probability that the sample label is 1 in the entire data set. Representation characteristics The marginal probability density of .

[0051] because and After the training set is determined, the value is fixed. The present invention proposes an optimization method: to "separate" the two posterior class conditional probability densities as much as possible. and , thereby optimizing the prior conditional probability between features and labels.

[0052] like Figure 2 As shown in (1), is the probability density function of a feature after normalization. According to the label, the feature data corresponding to the label 0 and the feature data corresponding to the label 1 are taken out respectively, and their class conditional probability density functions are plotted respectively. The results are as follows Figure 2 The three probability density function curves represent , , (The probability density function does not necessarily have to be a normal distribution, the figure is just a schematic diagram).

[0053] Assume that the feature data corresponding to the label 1 has a larger mean. Figure 2 As shown in (3) in the figure, we can find its first quartile point, which is denoted as For the probability density function with a smaller mean, find its third quartile, denoted as The difference between these two quantiles is This is the evaluation standard used in the present invention.

[0054]

[0055]

[0056]

[0057] After a feature combination is performed to obtain a new feature, the class conditional probability density function will change. Figure 2 (4) in the figure represents the class conditional probability density function of the new feature under ideal conditions, which is represented by a solid line in the figure and is denoted as and The previous class conditional probability density function is represented by a dotted line. From an intuitive point of view, the two distributions are more separated and the overlap is reduced. Reflected in the numerical value, the first quartile of the group with the larger mean is recorded as , the third quartile of the group with the smaller mean is recorded as At this time, the quantile difference of the class conditional probability density function is .

[0058] 2. Feature screening submodule:

[0059] The processing of discrete features is relatively simple. Take binary features as an example: Since the variable The value of is 0 or 1, and we only need to calculate the conditional probability of the variable and the label, that is and . The conditional probability is calculated by the Bayesian formula:

[0060]

[0061] When the difference between the conditional probabilities of the two labels is greater than the first threshold ,Right now , this feature will be selected as the original feature and used to combine with other features to generate new features in the next module. The value of is usually between 0.1 and 0.5, and the specific value depends on the data characteristics and model requirements.

[0062] For continuous features, the evaluation parameters used are (The quantile difference of the class conditional probability density function). Greater than the second threshold When , the feature will be selected as the original feature and also as a candidate feature to be combined with other features in the next module to generate a new feature. The value of is generally between 0.1 and 1.0, and the specific value depends on the data characteristics and model requirements.

[0063] The process diagram of feature screening is as follows: Figure 3 As shown in the figure, the discrete features that meet the conditions will be selected as the original features, and the continuous features that meet the conditions will be selected as both the original features and the candidate features; the discrete features or continuous features that do not meet the conditions will be discarded.

[0064] 3. New feature automatic generation submodule:

[0065] The method of generating new features from candidate features can be divided into two cases: feature operation itself and feature combination operation.

[0066] (1) Feature operation unit:

[0067] The feature operations in the training set can be: exponential operation (exp) of natural base, logarithmic operation (log) of natural base, absolute value operation, etc. Some data have large values. After exponential operation, they may exceed the upper limit of the computer floating point range, causing overflow. Therefore, the data will be normalized before performing exponential operation. Logarithmic operation requires that the data is greater than 0, and if the data is too close to zero, the result may also exceed the lower limit of the computer floating point range. Therefore, before performing logarithmic operation, it is necessary to make a judgment on the minimum value of the feature data: if the minimum value is less than 0, all data are added to the minimum value as a whole. Then add a small number (such as 0.01) to prevent the result close to negative infinity. In addition, reciprocal operation, square root operation, trigonometric function operation, linear rectifier function operation (ReLU), S-type function (sigmoid) operation, normalized exponential function (softmax), etc. can also be used to generate new features (after the feature data has been processed in a certain way) .

[0068]

[0069] in, Represents the original features The new eigenvalues ​​obtained after preprocessing are intended to ensure that the features are suitable for subsequent operations; Represents the set of all computational operations.

[0070] The maximum and minimum values ​​of the candidate feature training set and the operations performed will be recorded by the "New Feature Generation Parameter Recording Module" and used to perform the same operations on the test set data. Special processing methods are required for specific operations: for example, before the logarithmic operation, if the minimum value of the data in the test set Smaller than the minimum value of the data in the training set and is less than 0, then all the values ​​less than The data is replaced by This ensures the consistency of generated data and prevents calculation errors.

[0071]

[0072] in, represents the maximum value of the feature in the training set, represents the minimum value of the feature in the training set, Represents the original feature.

[0073] (2) Feature combination operation unit:

[0074] The combination of features in the training set can be divided into two situations: the combination of continuous features and continuous features, and the combination of continuous features and discrete features.

[0075] When a candidate feature is combined with other candidate features or continuous features in the original features, the operations that can be performed between the features include four arithmetic operations, norm operations (1-norm, 2-norm, infinity-norm) and other operations.

[0076] When the candidate features are combined with the discrete features in the original features, since the discrete features (here taking the binary features as an example) only have two values ​​0 and 1, the traditional four arithmetic operations are not applicable to this type of situation. In this regard, the present invention proposes a method for combining continuous features with discrete features. For addition and subtraction operations: first subtract 0.5 from all the data of the discrete features, that is, map [0, 1] to [-0.5, 0.5], and then perform operations with the normalized continuous features. For multiplication operations: map [0, 1] to [-1, 1] (it can also be [1,-1], depending on the class conditional probability density function of the continuous feature and the conditional probability of the discrete feature, which will be explained later), and then perform operations with the normalized continuous features.

[0077] In order to reduce the computational complexity and running time. The present invention aims to optimize the quantile difference of the class conditional probability density function, and performs selective combination operations instead of traversing all the operation methods. When continuous features are combined with continuous features to create new features, the "sameness / oppositeness of the class conditional probability density functions" of the two features are first analyzed separately. Assume that the mean of the class conditional probability density of the first feature satisfies (i.e. Figure 2 As in (2), the data corresponding to label 1 has a larger mean and is distributed more in the positive direction), if the mean of the class conditional probability density of the second feature also satisfies , then their class conditional probability densities are "isotropic". For features with isotropic properties, when a new feature is generated, operations such as addition or multiplication (when both values ​​are greater than 0) will be selected based on the value range of the two feature data, so that the two class conditional probability density functions of the new feature are far away from each other, thereby increasing the difference in quantiles. On the contrary, if the mean of the class conditional probability density of the second feature satisfies , then their class conditional probability density is "inverse". For features with inverse properties, subtraction or division (depending on the value range of the divided feature) is generally used to make the distributions move away from each other, thereby increasing the difference in quantiles. According to the same principle, when combining continuous features with discrete features, the "sameness / oppositeness of class conditional probability density function and conditional probability" of the two features must also be analyzed first. Assuming that the continuous feature still satisfies , if the mean of the conditional probability of the discrete feature satisfies , then the class conditional probability density and conditional probability can be considered to be "in the same direction", otherwise they can be considered to be "in the opposite direction". Then, according to their same / inverse direction and value range, the appropriate feature operation method is selected to generate new features.

[0078]

[0079] in,

[0080]

[0081]

[0082]

[0083] in, Respectively represent the values ​​of the two features after preprocessing, Represents the operation performed on two features. Indicates preprocessing of features. Respectively represent the addition, multiplication, and normalization of two features. They respectively represent subtraction, division, and normalization of two features.

[0084] 4. Candidate feature screening and updating submodule:

[0085] The generated new features can be added to the candidate features, and the next round of new feature automatic generation can be carried out. However, since the number of candidate features will grow exponentially, if they are not screened, the amount of calculation will be very large. The candidate feature screening method used in the present invention is still based on the evaluation criterion of the quantile difference of the class conditional probability density function. Assume that a new feature The corresponding quantile difference of the class conditional probability density function is , which is characterized by and Features The quantile differences of their class conditional probability density functions are and When the conditions are met When , the newly generated features are retained in the candidate features. Threshold It is not fixed. The initial value of is set to 0. After each candidate feature screening phase, Will be updated. New The value of is the value of all new features retained in this round The smaller of the median and mean of the set is used as the replacement, that is, . Since the newly generated candidate features are retained each time will be greater than the previous threshold , so the threshold after each update will also be greater than the previous threshold. When a round of features is automatically generated, there is no new feature. When it is greater than the threshold, the loop terminates and the candidate feature is output as the result.

[0086] The features that pass the screening will be retained in the candidate features. After several rounds of feature generation, the number of candidate features will become large. When the number of candidate features exceeds a threshold, the quantile difference of the class conditional probability density function of all candidate features needs to be calculated. Sort and keep The largest candidate features. Among the remaining features, randomly select All other candidate features are discarded.

[0087] 5. Automatic feature engineering reproduction submodule:

[0088] like Figure 1As shown, there is a new feature generation parameter recording module after the new feature automatic generation module. This module records the generation process and method of the new feature (for example, which / which two features are generated by which operation, what operations need to be performed on the features before the operation, etc.). When the automatic feature engineering is reproduced, the module performs corresponding operations on the features in the test set to generate the same candidate features. This module ensures the usability of the present invention in actual machine learning problems.

[0089] One thing to note here is that when performing some scaling and translation operations on the test set feature data, the training set is used as the benchmark, not the test set itself. Operation, where and are the maximum and minimum values ​​of the feature, The test set feature data is based on the upper and lower limits of the training set. So when normalizing, the operation is , ( is the normalized test set feature data) This is consistent with the normalization operation done in the training set.

[0090] The model training module uses a tree-based model to train automatically generated features and saves a risk prediction model for interstitial lung lesions in patients with rheumatoid arthritis.

[0091] The disease prediction module predicts the risk of interstitial lung disease in patients with rheumatoid arthritis based on the trained risk prediction model for interstitial lung disease in patients with rheumatoid arthritis and the patient's input data.

[0092] Example:

[0093] For example, the symptoms and medical test results of rheumatoid arthritis patients in the training set include "age", "body mass index (BMI)", "C-reactive protein", "erythrocyte sedimentation rate", etc. These continuous features can be selected as candidate features and original features after screening. The set of candidate features is recorded as Other discrete features such as “whether there is cough” and “whether there is morning stiffness” can be added to the original features after screening. The set of original features is recorded as .

[0094] In the first round of feature generation, the candidate features will perform self-operation to generate some features such as "normalized age" and "the inverse of C-reactive protein". At the same time, the candidate features will perform inter-feature combination operations with other features to generate some features such as "the result of the 2-norm operation of age and erythrocyte sedimentation rate" and "the result of the multiplication of age and whether the patient is in the morning". After screening, the set of candidate features is updated to Here Compared with the previous Although the notation is the same, they only represent the sequence number of the candidate features and are not necessarily the same feature. In the second round of feature generation, more complex features such as "the result of multiplying age by whether or not the patient has morning stiffness and then multiplying the result by the inverse of the erythrocyte sedimentation rate" will be generated. The set of candidate features will be updated again. Finally, the candidate features and original features are used to train the tree-based model. represents the number of initial candidate features (based on the screened continuous features), represents the number of original features (based on the filtered discrete features), represents the number of candidate features after the first round of feature generation, Indicates the number of candidate features after the second round of feature generation.

[0095] Often, tree-based models provide a property called "feature importance" after training. If the newly generated feature has a high importance and the model obtains good prediction results, the user can find its generation process and its closed expression in the "New Feature Generation Parameter Recording Module". Users with a medical research background can delve into the medical significance and other medical uses behind the new feature.

[0096] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering, characterized in that: It includes data input module, automatic feature engineering module, model training module and disease prediction module; The data input module is used to input the patient's symptoms and medical test result data to form symptom and medical test result characteristics; The automatic feature engineering module is used to automatically generate a feature suitable for a binary classification machine learning problem; the automatic feature engineering module includes a continuous feature evaluation submodule for evaluating the class conditional probability density of continuous features, a feature screening submodule for discarding features that do not meet the requirements, a new feature automatic generation submodule for automatically combining operations to generate new features, a candidate feature screening and updating submodule for screening and retaining new features according to a threshold, and an automatic feature engineering reproduction submodule for recording the new feature generation process; The execution operation of the continuous feature evaluation submodule is as follows: in the training set of the binary classification machine learning problem, the label corresponding to each feature is known, and the class conditional probability density of the feature corresponding to each label is calculated by the Bayesian conditional probability formula and ; By separating the two posterior class conditional probability densities and , thereby optimizing the prior conditional probability between features and labels; is the probability density function of a feature after normalization. According to the label, the feature data corresponding to the label 0 and the feature data corresponding to the label 1 are taken out respectively, and their class conditional probability density functions are plotted respectively. Assuming that the feature data corresponding to the label 1 has a larger mean, its first quartile is obtained, which is recorded as ; For the probability density function with a smaller mean, find its third quartile, denoted as ; The difference between these two quantiles That is the quantile difference of the class conditional probability density function: ； ； ； The execution operation of the feature screening submodule is: for a discrete feature, if the difference between the conditional probabilities of the two labels of the feature is greater than the first threshold ,Right now , then the feature is used as the original feature; otherwise, the feature is discarded; for continuous features, when the quantile difference of the class conditional probability density function is greater than the second threshold , then the feature is used as both the original feature and the candidate feature; The model training module is used to train the automatically generated features based on the tree model and save the model; The disease prediction module is used to predict the risk of interstitial lung lesions in patients with rheumatoid arthritis based on the risk prediction model of interstitial lung lesions in patients with rheumatoid arthritis and the patient's input data.

2. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 1, characterized in that: The symptom and medical test result features constituted by the data input module include continuous features and discrete features; among them, the continuous features include "age", "C-reactive protein test value", and "erythrocyte sedimentation rate test value", and the discrete features include "whether there is a history of smoking" and "whether the rheumatoid factor is positive"; the input label data is whether the patient has concurrent interstitial lung disease.

3. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 1, characterized in that: In the features generated by the automatic feature engineering module and applicable to the binary classification machine learning problem, the probability density function of the data corresponding to the label value of 0 is separated from the probability density function of the data corresponding to the label value of 1.

4. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 1, characterized in that: The new feature automatic generation submodule includes a feature self-operation unit and a feature combination operation unit; the feature self-operation unit performs feature self-operation to generate a new feature, or the feature combination operation unit performs feature combination operation to generate a new feature.

5. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 4, characterized in that: The feature operations themselves include exponential operations of natural bases, logarithmic operations of natural bases and absolute value operations; the feature combinations include combinations of continuous features with continuous features, and combinations of continuous features with discrete features.

6. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 5, characterized in that: The execution operation of the feature combination operation unit is: When a candidate feature is combined with other candidate features or continuous features in the original features, the operations between the features are performed through four arithmetic operations and norm operations; When the candidate features are combined with the discrete features in the original features, the operation between the features is performed by combining the continuous features with the discrete features. For addition and subtraction operations, all the data of the discrete features are subtracted by 0.5, that is, [0, 1] is mapped to [-0.5, 0.5], and then the operation is performed with the normalized continuous features. For multiplication, [0, 1] is mapped to [-1, 1] or [1,-1], and then the operation is performed with the normalized continuous features.

7. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 1, characterized in that: The candidate feature screening and updating submodule screens candidate features based on the quantile difference of the class conditional probability density function, sorts the screened candidate features according to the quantile difference of the class conditional probability density function, and retains the candidate features with the largest quantile difference of the class conditional probability density function. candidate features.

8. The risk prediction system for interstitial lung disease in patients with rheumatoid arthritis based on automatic feature engineering according to claim 1, characterized in that: The automatic feature engineering reproduction submodule is used to record the generation process and method of new features; when the automatic feature engineering is reproduced, the automatic feature engineering reproduction submodule performs corresponding operations on the features in the test set to generate the same candidate features.

Images