Coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection

Abstract

The invention discloses a coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection. The method comprises the following steps: collecting biochemical examination index data of CWP patients and healthy people; an SMOTE algorithm is adopted to carry out oversampling processing; a basic prediction model is selected and optimized; performing feature selection by using an integrated feature selection method, and constructing a new feature subset; and inputting the new feature subset into the prediction model to carry out CWP staging prediction. According to the invention, feature selection is carried out by using an integrated feature selection method, so that the classification performance and generalization ability of a prediction model can be significantly improved; based on conventional biochemical inspection data, the detection cost can be remarkably reduced, operation is easy and convenient, stage diagnosis can be conducted on CWP patients more accurately and rapidly, doctors are helped to formulate more effective treatment and prevention strategies, and the method is particularly suitable for CWP early screening and diagnosis of mining area workers in the medical environment with limited resources.

Description

Technical Field

[0001] The present invention relates to a coal worker's pneumoconiosis staging prediction method, in particular to a coal worker's pneumoconiosis staging prediction method based on small sample clinical data and integrated feature selection, belonging to the technical field of machine learning. Background Art

[0002] Coal worker's pneumoconiosis (CWP) is a general term for lung lesions caused by long-term inhalation of dust in the production environment by coal miners. It is an occupational disease characterized by pulmonary tissue fibrosis caused by long-term inhalation of dust particles in the production environment, which leads to dust particles being retained in the lungs. The disease has a certain degree of lateness, a high mortality rate, and can cause irreversible and huge damage to the lungs. According to the severity of the disease, CWP is usually divided into stages I, II, and III. So far, there is no effective treatment for CWP, so its early screening and diagnosis play a vital role in preventing the patient's condition from further deteriorating. The existing diagnosis and staging of CWP mainly rely on high-kilovoltage chest X-rays, and professional doctors perform manual analysis and staging based on the density and distribution characteristics of the image shadows. However, this method is time-consuming and easily affected by subjective factors, which may lead to misdiagnosis. In addition, the cost of imaging diagnostic methods is high, which also brings certain economic pressure to patients.

[0003] In recent years, with the development of artificial intelligence and machine learning technology, automated technology for clinical data analysis and disease diagnosis has been widely used. Especially in the case of small sample data sets, machine learning can effectively predict the staging and diagnosis of the disease through the construction of models. For example, the Chinese invention patent with publication number CN117954094A discloses a method, device, electronic device and storage medium for predicting the incidence of pneumoconiosis in coal mines. This method inputs dust information such as dust concentration and dust composition in the coal mine into a pre-trained SOM-BP neural network, so that the SOM-BP neural network outputs the incidence of pneumoconiosis within a specified time period based on the dust information. However, the dust information involved in this method is difficult to obtain, the time range involved is wide, and the investigation is relatively difficult. Another example is the Chinese invention patent with the authorization announcement number CN117315378B, which discloses a classification determination method for pneumoconiosis and related equipment. This method trains a stable diffusion model by preprocessing and labeling training samples, then normalizes and models the DR chest radiographs, obtains the segmented images, and adds them to the training set for classification. However, the DR images involved in this invention are expensive, and the performance of the deep learning model depends largely on the quality and quantity of the DR images of the training data. When the training data is insufficient or of low quality, the prediction accuracy of the model will be affected. Therefore, how to achieve more accurate, more convenient and lower-cost CWP staging prediction is an urgent problem to be solved in the industry. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection, which can solve the problems existing in the existing CWP staging prediction, such as uneven data distribution, judgment results affected by subjective factors and high cost, and realize accurate, convenient and low-cost CWP staging prediction.

[0005] To achieve the above purpose, the coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection specifically includes the following steps:

[0006] Step 1, data collection: collect clinical biochemical examination data of CWP patients and control biochemical examination data of healthy people;

[0007] Step 2, data preprocessing: use the SMOTE algorithm to oversample the data obtained in Step 1;

[0008] Step 3, build a prediction model: select a basic prediction model and optimize it;

[0009] Step 4, feature selection: Use the integrated feature selection method to select features from the clinical biochemical examination data in the oversampled data obtained in Step 2, and construct a new feature subset, as follows:

[0010] a. Perform multiple feature importance evaluations on the clinical biochemical examination data in the oversampled data obtained in Step 2 using the random forest algorithm, average the values ​​after multiple random forest importance evaluations, and obtain the importance value of each feature;

[0011] b. Perform correlation analysis on the clinical biochemical test data in the oversampled data obtained in Step 2 by using a correlation analysis algorithm to obtain correlation coefficients between different features, add the correlation coefficients of each feature with other features and calculate the average value to obtain the independence value of each feature;

[0012] c. Construct the following linear relationship model that comprehensively considers feature importance and relevance:

[0013] y=ω1x1+ω2(-x2)

[0014] Where: y is the comprehensive evaluation value of each feature; ω1 and ω2 are the feature importance weight and feature independence weight respectively, and ω1+ω2=1; x1 is the importance value; x2 is the independence value;

[0015] d. Assign values ​​to ω1 and ω2 respectively, obtain weight combinations, and substitute the importance value obtained in step a and the independence value obtained in step b into the linear relationship model to obtain the comprehensive evaluation value y of each feature variable, and then perform dimensionality reduction sorting on the comprehensive evaluation value y under the weight combination, and select features within the set percentage range according to the sorting results to construct a new feature subset;

[0016] Step 5, CWP stage prediction: Input the new feature subset obtained in Step 4 into the prediction model constructed in Step 3 to perform CWP stage prediction.

[0017] Furthermore, in step d of Step 4, when assigning values ​​to ω1 and ω2, ω1:ω2=4:6.

[0018] Furthermore, in step b of Step 4, the correlation analysis algorithm is the Spearman correlation analysis algorithm.

[0019] Furthermore, in Step 3, the basic prediction model selected is the SVM model, and the SVM model is optimized by the PSO algorithm to generate the SVM-PSO prediction model.

[0020] Furthermore, in Step 1, the clinical biochemical examination data are derived from routine biochemical examinations of blood routine, liver function, kidney function, and blood lipids.

[0021] Compared with the prior art, the coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection uses an integrated feature selection method for feature selection, which can significantly improve the classification performance and generalization ability of the prediction model, and is particularly suitable for scenarios with small sample sizes and unbalanced categories of medical data; compared with the traditional CWP staging prediction method based on imaging diagnosis, the present invention is based on conventional biochemical examination data, which can significantly reduce the detection cost, and is easy to operate. It can provide a theoretical basis and data support for the early screening and diagnosis of CWP, and can more accurately and quickly diagnose CWP patients by stage, helping doctors to develop more effective treatment and prevention strategies, and is particularly suitable for the early screening and diagnosis of CWP among mining workers in a medical environment with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of the present invention;

[0023] Figure 2 It is a schematic diagram of data distribution before and after oversampling of the present invention;

[0024] Figure 3 It is the flow chart of Adaboost algorithm;

[0025] Figure 4 It is the PSO algorithm flow chart;

[0026] Figure 5 It is the importance assessment diagram of 22 biochemical indicators;

[0027] Figure 6 This is a diagram of the prediction results based on the selection of important features;

[0028] Figure 7 It is a heat map of correlation coefficients among 22 biochemical indices;

[0029] Figure 8 This is a diagram showing the prediction results based on correlation feature selection. DETAILED DESCRIPTION

[0030] This coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection collects routine biochemical examination data of CWP patients and uses SMOTE oversampling technology to solve the imbalance problem of CWP staging data; combines the machine learning algorithm of support vector machine (SVM) and uses particle swarm optimization (PSO) algorithm to optimize the model to improve the accuracy of CWP staging prediction; comprehensively considers the importance and relevance of features, and adopts an integrated feature selection method to select features to improve the classification performance and generalization ability of the prediction model.

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] like Figure 1 As shown, the coal worker pneumoconiosis staging prediction method based on clinical data and integrated feature selection specifically includes the following steps:

[0033] Step 1: Data Collection

[0034] The purpose of data collection is to establish a broad and rich database for deep learning and training and optimization of CWP staging prediction. By collecting clinical biochemical examination data of CWP patients and control biochemical examination data of healthy people, the model's ability to identify and analyze the degree of CWP lesions can be improved.

[0035] Embodiment The clinical biochemical examination data of 81 CWP patients and the control biochemical examination data of 202 healthy people were obtained from Xuzhou Mining Group General Hospital. Common demographic and common biochemical examinations (such as routine blood tests, liver function, kidney function, blood lipids and other routine biochemical tests) were selected. The clinical biochemical examination data obtained included globulin, albumin / globulin, creatinine, alanine aminotransferase, very low density lipoprotein, hemoglobin, low density lipoprotein, eosinophil absolute value, triglyceride, glucose, lymphocyte absolute value, cholesterol, urea, uric acid, aspartate aminotransferase, neutrophil absolute value, red blood cell, high density lipoprotein, white blood cell, monocyte absolute value, platelet, albumin and other 22 biochemical examination data.

[0036] Step 2: Data preprocessing

[0037] Since the patient data of each stage of CWP is usually unevenly distributed, especially the sample size of stage II and stage III is small, it is easy to cause poor classification performance of the machine learning model. Among the data collected in the embodiment, stage II and stage III only account for 7.4% and 2.1%, respectively, and there is a data imbalance problem. Therefore, the SMOTE algorithm is used to oversample the data to increase the number of minority class samples and ensure the balance of data distribution. For minority class samples, the SMOTE algorithm can generate new sample points by interpolating between the nearest neighbor samples in the feature space to make the number of each type of sample tend to be balanced, solve the model bias problem caused by data imbalance, and ensure that the model has higher recognition ability when processing minority class data.

[0038] After oversampling by SMOTE algorithm, the Phase I data of CWP increased by 148 cases, the Phase II data increased by 181 cases, and the Phase III data increased by 196 cases. The total clinical data was expanded from 283 cases to 808 cases, achieving data balance. The data distribution diagram before and after oversampling is shown in the figure below. Figure 2 shown.

[0039] Step 3: Selection and optimization of prediction model

[0040] The embodiment uses random forest (RF), support vector machine (SVM), K nearest neighbor (KNN), and decision tree (DT) as basic models, and performs stage prediction of CWP based on RF, SVM, KNN, DT, and optimized RF-Adaboost and SVM-PSO machine learning models.

[0041] The RF model consists of multiple decision trees. Although it has good classification performance, it is less sensitive to minority samples when processing unbalanced data. The Adaboost algorithm can overcome the shortcomings of RF by iteratively training multiple weak classifiers to increase the focus on misclassified samples to build a strong classifier. Therefore, the Adaboost algorithm is introduced to optimize the RF model. The optimization mechanism of the Adaboost algorithm is as follows: Figure 3 shown.

[0042] The specific steps of the AdaBoost algorithm are as follows:

[0043] Given a data set D = (x1, y1), (x2, y2), ..., (x N ,y N ), where x i D =(x i1 ,x i2 ,...,x id) is the i-th input sample, d represents the number of features, y i ∈{0,1,2,3}, N is the number of samples.

[0044] ① Initialize sample weights: that is, the initial weight of each sample is set to the same value. The initial weight distribution is expressed as D1 = (ω 11 ,ω 12 ,...,ω 1N ) D , Among them, ω 1i Represents the weight at the beginning of the i-th training sample.

[0045] ② Iterative training and evaluation of weak classifiers: Set the number of iterations to T, and T weak learners will be trained. Specifically, each iteration will update the weight distribution D t , and then use the training data with updated weight distribution to learn to obtain the weak learner in each round.

[0046] Assume that the weak learner obtained from the tth training is represented by h t (x), the performance of the weak learner will be measured by its classification error rate e on the training data set t It is expressed as follows:

[0047]

[0048] Where: I(·) is the indicator function. If the classifier h t (x) is the wrong prediction for the sample, then I(h t (x)≠y i )=1 otherwise it is 0.

[0049] From the above formula, we can know that h t (x) The error rate on the training data set is h t (x) The sum of the weights of the misclassifications.

[0050] ③ Calculate the weight of the weak classifier: use the weak classification weight α t Indicates h t (x) The importance of the final integrated classifier is calculated as follows:

[0051]

[0052] From the above formula, we can see that the smaller the error rate of the classifier, the greater the weight α of the classifier. t The larger it is, the more important the classifier is in the final classifier.

[0053] ④ Update sample weights: adjust sample distribution according to the classification results, so that the training of the next weak learner will focus more on the samples that were misclassified this time, that is, give higher weights to the samples that were misclassified. The adjustment formula for weight distribution is as follows:

[0054] D t+1 =(ω t+1,1 ,ω t+1 ,2,...,ω t+1,N ) D

[0055] ω t+1,i =ω ti exp(α t ·I(h t (x)≠y i ))

[0056] In order to facilitate the later processing, t+1,i The normalization process is as follows:

[0057]

[0058] ⑤Final classifier model: After the iteration is completed, all weak classifiers are combined to obtain the final classifier H(x). The final classification result is obtained by weighted voting of all weak classifiers, as shown below:

[0059]

[0060] The SVM model has good generalization ability in the case of small sample data, but its performance is highly dependent on the selection of penalty parameter c and kernel function parameter g. Traditional parameter tuning methods such as grid search need to be done manually and are time-consuming. The PSO algorithm can use group collaboration and information sharing mechanisms to find the optimal solution by moving and updating particles in multidimensional space. It has the advantages of fast convergence speed and no dependence on gradient information, thus overcoming the shortcomings of SVM. Therefore, the PSO algorithm is introduced to optimize the SVM model. The PSO optimization mechanism is as follows: Figure 4 shown.

[0061] The specific steps of the PSO algorithm are as follows:

[0062] ① Initialize the particle swarm parameters: set the particle swarm size to N, particle dimension to D, number of iterations to T, inertia weight ω, and learning factors c1 and c2. Each particle represents a solution. The learning factor c1 indicates the weight of the next action of the example derived from its own experience, and the acceleration weight that pushes the particle to the optimal position of the individual. c2 indicates the weight of the next action of the particle derived from the experience of other particles, and the acceleration weight that pushes the particle to the optimal position of the group.

[0063] ② Randomly initialize the speed and position of each particle in the search space: For N particles, randomly initialize their position vector X in the search space i =(x i1 ,x i2 ,...,x in ) and the velocity vector V i =(v i1 ,v i2 ,...,v in ), where i = 1, 2, ..., N.

[0064] Initialize the individual best historical position as the current position of each particle, and record the individual best historical position of the particle as Pbest i =(p i1 ,p i2 ,...,p in ), initially Pbest i =X i The individual's historical optimal fitness value is recorded as PbestFitness i , initially it is the classification accuracy corresponding to the current position, that is, the accuracy calculated on the training set using the classifier parameters represented by the current particle position.

[0065] Initialize the best historical position and fitness value of the group. The best historical position of the group is recorded as Gbest = (g1, g2, ..., g n ), the position of the first particle can be used as the initial group optimum. The historical optimal fitness value of the group is recorded as GbestFitness, which is the individual historical optimal fitness value of the first particle at the beginning.

[0066] ③Calculate the fitness value of each particle: For each particle i, its current position x i Substitute the objective function f(x) and calculate the fitness value Fitness i =f(x i ). The fitness value indicates the quality of the particle's current position.

[0067] ④ Update the historical optimal fitness value and position of individual particles and the optimal fitness value and position of the group:

[0068] Update the best individual history of particles: For each particle i, compare its current fitness value with the best individual history fitness value. If the current fitness value is better, that is, Fitness i >PbestFitness i , then update the individual's historical best position to the current position, that is, Pbest i =X i, update the individual's historical best fitness value to the current fitness value, that is, PbestFitness i =Fitness i .

[0069] Update the historical optimum of the group: find the optimal fitness value among the individual historical optimal fitness values ​​of all particles. The particle position corresponding to this fitness value is the optimal position in the history of the group, that is,

[0070]

[0071] ⑤Update the speed and position of each particle:

[0072] The speed update formula is expressed as: i (t+1)=ωv i (t)+c1r1(p i -x i (t))+c2r2(gx i (t))

[0073] Where: v i (t+1) is the next iteration speed of particle i, v i (t) is the current speed, x i (t) is the current position, p i is the best historical position of the individual particle, g is the best historical position of the group, r1 and r2 are random numbers uniformly distributed in the interval [0,1].

[0074] The position update formula is expressed as: i (t+1)=x i (t)+v i (t+1)

[0075] After updating the position, check whether the particle exceeds the reasonable value range of the parameter. If so, perform boundary processing.

[0076] ⑥Update other parameters:

[0077] According to the design of the algorithm, parameters such as inertia weight and learning factor can be updated to dynamically adjust the search behavior of the algorithm. For example, the inertia weight can be linearly decreased, that is, where ω max and ω min are the maximum and minimum values ​​of the inertia weight respectively, and t is the current iteration number.

[0078] ⑦ Check whether the end condition is met: Determine whether the maximum number of iterations has been reached, or whether the optimal fitness of the group has not been significantly improved after several consecutive iterations. If the end condition is met, output the global optimal result and end the program, otherwise go to step ③ to continue execution.

[0079] Step 4, feature selection:

[0080] The embodiment uses a variety of feature selection methods to reduce the dimension of the oversampled data. After comparing the accuracy of the results, the optimal feature selection method is selected to reduce the dimension, so as to achieve the prediction of CWP stages with the least feature variables. Specifically:

[0081] Feature selection method 1: feature selection based on random forest importance evaluation method:

[0082] Random forest is an ensemble learning algorithm based on decision trees. It selects training sets from the sample set by sampling with replacement, then randomly selects features at each node for partitioning, generates multiple decision trees, and finally obtains the final prediction result by voting or taking the average. Random forest performs well in classification and regression tasks, and has high accuracy and stability.

[0083] The random forest importance assessment method is to score each feature based on its contribution to the decision tree classification and derive its importance to the target variable. The higher the feature importance, the greater the contribution of the feature to the CWP stage prediction. The results of the random forest feature importance assessment of 22 biochemical indicator features are shown in Figure 2. Figure 5 The results show that albumin, platelets, absolute monocyte count, white blood cells, and high-density lipoprotein are the five indicators that contribute most to the prediction of CWP staging. Based on this, the key features that have a significant impact on the prediction of CWP staging were screened out. The features ranked in the top 70% of the importance data were selected as input and combined with the machine learning model for prediction. The prediction results are shown in the figure below. Figure 6 shown.

[0084] Feature selection method 2: feature selection based on correlation analysis:

[0085] Since the input variables are all clinical data, there are often highly correlated multicollinearity. The 22 biochemical index characteristics were analyzed by Spearman correlation analysis algorithm, and the correlation coefficient heat map is shown in the figure below. Figure 7 In order to eliminate the influence of multicollinearity on the prediction ability of the model, the correlation coefficient threshold was set to 0.7. After screening, 6 feature pairs that met the conditions were obtained: glycerol and very low-density lipoprotein, cholesterol and low-density lipoprotein, red blood cells and hemoglobin, globulin and globulin / albumin, white blood cells and absolute value of neutrophils, alanine aminotransferase and aspartate aminotransferase.

[0086] The six feature pairs are arranged and combined to obtain 64 feature combinations that need to be eliminated. Next, the six features contained in each combination are eliminated from the original data set to obtain 64 new data sets. The 64 new data sets are then combined with the machine learning model for prediction to obtain the data set with the best performance. The prediction results are as follows: Figure 8 shown.

[0087] Feature selection method three: feature selection based on integrated feature selection method:

[0088] The integrated feature selection method is to linearly fit the importance and independence of features, so as to identify the key features that have a significant impact on the model performance and thus improve the predictive ability of the model. It includes the following steps:

[0089] a. Perform multiple feature importance evaluations on 22 biochemical indicator features using the random forest algorithm, average the values ​​after multiple random forest importance evaluations, and obtain the importance value of each biochemical indicator feature;

[0090] b. Perform correlation analysis on the 22 biochemical index characteristics using the Spearman correlation analysis algorithm, add the correlation coefficients of each biochemical index characteristic with other biochemical index characteristics and calculate the average value to obtain the independence value of each biochemical index characteristic;

[0091] c. Construct the following linear relationship model that comprehensively considers feature importance and relevance:

[0092] y=ω1x1+ω2(-x2)

[0093] Where y is the comprehensive evaluation value of each feature, which is used to measure the total score after the importance and independence of the feature are combined; ω1 and ω2 are the feature importance weight and feature independence weight, respectively, representing the relative importance of importance and independence in the comprehensive evaluation, and ω1+ω2=1; x1 is the importance value; x2 is the independence value.

[0094] d. Assign values ​​to ω1 and ω2 respectively, obtain weight combinations, and substitute the importance value obtained in step a and the independence value obtained in step b into the linear relationship model to obtain the comprehensive evaluation value y of each biochemical indicator characteristic variable, and then perform dimensionality reduction sorting on the comprehensive evaluation value y under the weight combination, select the biochemical indicator features within the set percentage range according to the sorting result order, and construct a new feature subset for subsequent model training.

[0095] In the embodiment, the values ​​of ω1 and ω2 are assigned with a gradient of 0.1, 0.1 to 0.9, and a total of 9 weight combinations. The comprehensive evaluation value y is sorted by dimensionality reduction under the 9 weight combinations, and the features in the top 70% are taken according to the sorting results, and finally 5 new feature subsets are constructed.

[0096] Step 5, CWP phase prediction

[0097] The new feature subset was used as input and combined with the machine learning model for CWP stage prediction.

[0098] The embodiment divides the training data set and the test data set into a ratio of 7:3, and uses multi-dimensional evaluation indicators such as accuracy (ACC), precision (P), recall (R), and F1 score to comprehensively evaluate the performance of the model. The evaluation results of the five new feature subsets are shown in Table 1 below.

[0099] Table 1 Evaluation results of each new feature subset

[0100]

[0101]

[0102] It can be seen from Table 1 that the model performance is best when ω1:ω2=4:6. Figure 6 , Figure 8 The prediction results are compared with the optimal prediction results in Table 1 through the F1 score, and the results are shown in Table 2 below.

[0103] Table 2 Prediction results of three feature selection methods

[0104]

[0105] Table 2 shows the performance results of various machine learning models based on single-dimensional feature selection methods and integrated feature selection methods. The higher the F1 score, the more accurate the model is in predicting CWP stages, and the higher the improvement percentage, the better the algorithm improvement effect. It can be seen that the SVM-PSO model shows the best accuracy in predicting CWP stages, with an F1 score of 0.9547.

[0106] Compared with the basic models RF and SVM, the performance of the optimized RF-Adaboost and SVM-PSO has been significantly improved after selecting multiple features. The optimization effect is more obvious on SVM. In particular, the performance optimization after selecting important features is the largest, with an increase of 16.59%. According to the above results, it can be seen that the addition of the optimized model can make up for the shortcomings of the original model, proving the effectiveness of optimizing the model.

[0107] Compared with the feature selection method based on importance, the performance of the integrated feature selection method on RF, RF-Adaboost, SVM, SVM-PSO, DT, and KNN was improved by 1.71%, 1.31%, 4.83%, 2.72%, 5.46%, and 6.86%, respectively. Compared with the feature selection method based on correlation, the performance of the integrated feature selection method on the six machine learning models was improved by 1.13%, 1.17%, 1.2%, 1.2%, 4.23%, and 1.43%, respectively. According to the above results, it can be seen that the integrated feature selection method has a good improvement in model prediction effect compared with the traditional single-dimensional feature selection method while ensuring a high prediction accuracy, which proves the effectiveness of the integrated feature selection method.

Claims

1. A method for predicting the stages of coal workers' pneumoconiosis based on clinical data and integrated feature selection, characterized in that: The specific steps include: Step 1, data collection: collect clinical biochemical examination data of CWP patients and control biochemical examination data of healthy people; Step 2, data preprocessing: use the SMOTE algorithm to oversample the data obtained in Step 1; Step 3, build a prediction model: select a basic prediction model and optimize it; Step 4, feature selection: Use the integrated feature selection method to select features from the clinical biochemical examination data in the oversampled data obtained in Step 2, and construct a new feature subset, as follows: a. Perform multiple feature importance evaluations on the clinical biochemical examination data in the oversampled data obtained in Step 2 using the random forest algorithm, average the values ​​after multiple random forest importance evaluations, and obtain the importance value of each feature; b. Perform correlation analysis on the clinical biochemical test data in the oversampled data obtained in Step 2 by using a correlation analysis algorithm to obtain correlation coefficients between different features, add the correlation coefficients of each feature with other features and calculate the average value to obtain the independence value of each feature; c. Construct the following linear relationship model that comprehensively considers feature importance and relevance: y=ω1x1+ω2(-x2) Where: y is the comprehensive evaluation value of each feature; ω1 and ω2 are the feature importance weight and feature independence weight respectively, and ω1+ω2=1; x1 is the importance value; x2 is the independence value; d. Assign values ​​to ω1 and ω2 respectively, obtain weight combinations, and substitute the importance value obtained in step a and the independence value obtained in step b into the linear relationship model to obtain the comprehensive evaluation value y of each feature variable, and then perform dimensionality reduction sorting on the comprehensive evaluation value y under the weight combination, and select features within the set percentage range according to the sorting results to construct a new feature subset; Step 5, CWP stage prediction: Input the new feature subset obtained in Step 4 into the prediction model constructed in Step 3 to perform CWP stage prediction.

2. The method for predicting coal workers' pneumoconiosis stages based on clinical data and integrated feature selection according to claim 1 is characterized in that: In step d of Step 4, when assigning values ​​to ω1 and ω2, ω1:ω2=4:

6.

3. The method for predicting coal workers' pneumoconiosis stages based on clinical data and integrated feature selection according to claim 1 is characterized in that: In step b of Step 4, the correlation analysis algorithm is the Spearman correlation analysis algorithm.

4. The method for predicting coal workers' pneumoconiosis stages based on clinical data and integrated feature selection according to claim 1, characterized in that: In Step 3, the basic prediction model selected is the SVM model, and the SVM model is optimized by the PSO algorithm to generate the SVM-PSO prediction model.

5. The method for predicting coal workers' pneumoconiosis stages based on clinical data and integrated feature selection according to claim 1, characterized in that: In Step 1, clinical biochemical examination data are derived from routine biochemical examinations of blood routine, liver function, kidney function, and blood lipids.

Images