Occupational health examination result prediction method based on expert system and machine learning

Abstract

The application discloses a professional health examination result prediction method based on an expert system and machine learning, and the method is: after adding the number of visits information to the physical examination archive information to form an original data set after cleaning and standardizing, important features are extracted through deep learning to obtain a new data set by recombination; all hazard factor types of each physical examination personnel in the new data set are identified; the corresponding prediction label is obtained through logical judgment of each hazard factor according to the occupational health standard to form a decision set; the decision set is marked with occupational contraindications and suspected occupational diseases respectively, and the corresponding feature SOD and feature OC of the marked results are added to the new data set; the new data set after the expert system link is randomly divided into a training set and a test set, and more than two machine learning models are selected to train the training set to obtain respective prediction results; the prediction results of the more than two models are weighted and averaged through a weighted voting method to obtain the final output result. The application can more effectively identify the occupational health risk.

Description

Technical Field

[0001] This invention relates to the field of predictive medicine, and more particularly to a method for predicting occupational health examination results based on expert systems and machine learning. Background Technology

[0002] A fundamental task of predictive medicine is predicting an individual's likelihood of contracting a disease. This is crucial for developing effective preventative measures aimed at completely preventing disease or at least minimizing its impact on patients. In recent years, data mining and machine learning techniques have made significant progress in addressing problems in predictive medicine and bioinformatics. However, due to the diversity of datasets and the specific nature of predictive medicine tasks, research on data analysis, processing, mining, and predictive models for occupational health examination results is relatively limited. While some analysis and mining work has been conducted on other relevant datasets, these studies often lack in-depth exploration of data characteristics, and the models used are not ideal in predictive performance, exhibiting limitations in practical applications.

[0003] The performance of existing k-Nearest Neighbors (KNN) algorithms largely depends on the quality and quantity of training data. If the training data is biased or incomplete, the model's prediction accuracy will be affected. Secondly, when dealing with large datasets, the KNN algorithm suffers from high computational complexity due to the need to calculate the distances between the sample to be predicted and all training samples, resulting in a slow prediction process. Furthermore, the KNN algorithm is highly sensitive to feature scale, therefore feature normalization is typically required before application.

[0004] Support Vector Machine (SVM) algorithms are highly sensitive to parameter selection, especially the regularization parameter C and the kernel function parameters. These parameters typically require optimization through methods such as cross-validation. Secondly, SVMs may encounter computational efficiency issues when dealing with large-scale datasets, particularly with very high feature dimensions. Furthermore, SVM models have relatively poor interpretability, the so-called "black box" problem, which can be a challenge in fields requiring interpretability such as occupational health examinations and diagnoses.

[0005] Furthermore, while the Random Forest (RF) algorithm has demonstrated significant advantages in predicting occupational health examination and occupational disease diagnoses, it also faces several challenges and limitations in practical applications. First, due to the complexity of the Random Forest model, it may contain a large number of decision trees, which not only increases training time but may also lead to high memory consumption. Secondly, the model's interpretability is affected, as integrating the predictions of numerous decision trees makes the decision-making process less transparent and difficult to explain to non-expert users. The performance of Random Forests largely depends on the quality and representativeness of the training data; bias or incompleteness of the data can affect the model's predictive accuracy. Simultaneously, although Random Forests reduce the risk of overfitting through ensemble methods, inappropriate adjustment of model parameters, such as the number of trees and the number of features selected, can still lead to overfitting. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting occupational health examination results based on expert systems and machine learning.

[0007] The technical solution adopted in this invention is:

[0008] A method for predicting occupational health examination results based on expert systems and machine learning includes the following steps:

[0009] Step 1, Data Processing: After cleaning and standardizing the physical examination records, visit frequency information is added to form the original dataset; some feature data of the original dataset are extracted by deep learning through a neural network model to obtain important features; the important features and the remaining feature data of the original dataset together form a new feature dataset;

[0010] Furthermore, step 1 specifically includes the following steps:

[0011] Step 1-1: Traverse the physical examination records and delete redundant feature information that is irrelevant to the physical examination conclusion;

[0012] Furthermore, the redundant feature information is personal information features, which include name, medical examination number, employer name, and employer's region.

[0013] Steps 1-2: Extract hazard factor information from the physical examination records; convert each individual hazard factor in the hazard factor information into a corresponding dummy variable to form a dummy variable set, and map the individual hazard factor name of each dummy variable to the hazard factor name classified in the specified occupational health standard, namely the "Technical Specification for Occupational Health Surveillance"; wherein, the hazard factor information includes physical examination hazard factors and exposure hazard factors;

[0014] Specifically, we obtain two features: physical examination hazard factors and exposure hazard factors. We use commas as separators to identify single hazard factors and convert them into dummy variables F. We then create a dictionary to map the single hazard factor names of F to the hazard factor names classified in the occupational health standard, namely the "Technical Specification for Occupational Health Surveillance". We delete the two features: physical examination hazard factors and exposure hazard factors, and add the dummy variable F.

[0015] Steps 1-3: Identify and obtain time feature information from the physical examination records, convert it into months, and save it as an integer.

[0016] Steps 1-4 involve extracting information from the physical examination records and adding information on the number of medical visits to form the original dataset.

[0017] Steps 1-5: Construct a neural network model. Input some feature data from the original dataset into the neural network model to perform deep learning on the original dataset to extract important features. The important features and the remaining feature data from the original dataset together form a new feature dataset.

[0018] Furthermore, the time characteristic information in steps 1-3 includes total length of service, length of service during the disaster, and start and end dates.

[0019] Furthermore, the neural network model in steps 1-5 includes an input layer, hidden layers, and an output layer. The input layer receives data with a dimension equal to the number of features in the training set. The hidden layer consists of two linear layers. The first linear layer maps the input features to a 256-dimensional feature vector. Dropout regularization is added after the first linear layer, with a dropout rate of 0.2. The second linear layer compresses the 256-dimensional feature vector output from the first linear layer into a 128-dimensional feature vector. Each linear layer is followed by a ReLU activation function to introduce non-linearity and enhance the model's expressive power. The linear layer of the output layer maps the 128-dimensional feature vector to the number of target categories, i.e., 5 dimensions.

[0020] Furthermore, in steps 1-5, the neural network model is trained using the cross-entropy loss function and the Adam optimizer.

[0021] Specifically, the neural network model performs the following actions: The input layer receives data of the same dimensionality as the training set features; then, the data undergoes feature mapping through two linear layers. The first linear layer maps the data to 256 dimensions, and the second linear layer further compresses it to 128 dimensions. After each linear layer, a ReLU activation function is applied to increase the model's non-linearity. Dropout regularization is added after the first linear layer, with a dropout rate of 0.2. Finally, a linear layer maps the 128-dimensional feature vector to 5 dimensions, completing the forward propagation process of the entire network. During training, the network is trained using the cross-entropy loss function and the Adam optimizer. After training, the trained network extracts features from new data. Finally, a new dataset is obtained by combining the features extracted from the second linear layer by the neural network with other features from the original dataset.

[0022] Specifically, the data features extracted by the second linear layer of the neural network represent new features that affect the prediction results after the neural network has been trained. These can be understood as new latent features discovered through the correlation between certain key features in the dataset.

[0023] Step 2: Identify all types of hazard factors for each examinee in the new dataset; based on the provisions of the specified occupational health standard (specifically, the National Occupational Health Standard of the People's Republic of China GBZ188-2014 "Technical Specifications for Occupational Health Surveillance"), perform logical judgment on each hazard factor to obtain the corresponding prediction label, and summarize them to form a decision set; label the decision set with occupational contraindications and suspected occupational diseases respectively, and add the corresponding features SOD and OC to the new dataset;

[0024] Furthermore, the predictive labels include "No abnormalities found so far", "Re-examination", "Suspected occupational disease", "Occupational contraindication" and other diseases or abnormalities.

[0025] Furthermore, the specific steps of step 2 are as follows:

[0026] Step 2-1: Identify all risk factor types for individuals undergoing physical examinations in the dataset: F = {f1, f2, ..., f n}, where n represents the number of single hazard factor features; for each physical examination record, iterate through the single hazard factor feature set F to obtain the single hazard factor feature set G corresponding to each physical examination record. The feature set G of the i-th physical examination record is G_i. i G = {g1, g2, ..., g} k}, where k is the number of single hazard features with a value of 1;

[0027] Specifically, as a feasible implementation method, the number of single hazard factor features n = 81; obtain the single hazard factor features with a value of 1 in the single hazard factor feature set F to form the single hazard factor feature set G for each physical examination record.

[0028] Step 2-2: Identify and obtain the physical examination feature t corresponding to the hazard factor f for each physical examination record, where T is the set of physical examination features t; the set of physical examination features T for the i-th physical examination record is Ti. i : T = {t1, t3, ..., t d ,d is the number of physical examination features};

[0029] Steps 2-3 involve constructing a logical function H for each hazard factor f, with the specific expression as follows:

[0030]

[0031] Steps 2-4: Establish a corresponding decision set D, D = {d1, d2, ..., dn} for the single hazard factor characteristic set G. k ,d i =H(f) i ,T i )},d i This represents the single hazard factor g corresponding to the single hazard factor characteristic set G. i The value determined by the logical function H; the set of physical examination characteristics D of the j-th physical examination record is D. j :

[0032] Steps 2-5: Label the decision set with occupational contraindications and suspected occupational diseases respectively to obtain new features SOD and OC, and add the new features SOD and OC to the new dataset;

[0033] Among them, the feature SOD represents the expert system's simulation of expert decision-making in the judgment of suspected occupational diseases. 1 indicates the presence of a suspected occupational disease, and 0 indicates its absence; the feature OC represents the expert system's simulation of expert decision-making regarding the judgment of occupational contraindications. 1 indicates the presence of occupational contraindications, and 0 indicates the absence of contraindications.

[0034] Step 3, Hybrid Model Training: The new dataset, after passing through the expert system, is randomly divided into a training set and a test set. Two or more machine learning models are selected and trained on the training set to obtain their respective prediction results. The prediction results of the two or more models are weighted and averaged to obtain the final output result through a weighted voting method.

[0035] Furthermore, four models were selected: Random Forest, AdaBoost, XGBoost, and LightGBM.

[0036] Furthermore, using the accuracy of the test set on each learning model as an evaluation metric, a grid search was employed to find the hyperparameters of each learning model.

[0037] Specifically, using accuracy as the evaluation metric, we employed grid search to find the hyperparameters of Random Forest, AdaBoost, XGBoost, and LightGBM models, respectively. Then, we used a weighted voting method to average the prediction results of the four models to obtain the final output.

[0038] Furthermore, the specific steps of step 3 are as follows:

[0039] Step 3-1: The new dataset that has passed through the expert system is randomly divided into a training set and a test set in a 7:3 ratio. Two or more machine learning models are trained on the training set. The hyperparameter range of each learning model is set using grid search. The accuracy of the predictions made by the model on the test set is used as the evaluation metric to find the hyperparameters of the model.

[0040] Specifically, the selected machine learning models were Random Forest, AdaBoost, XGBoost, and LightGBM.

[0041] Step 3-2 involves using a weighted voting method to average the predictions of two or more models to obtain the final output. The accuracy of each model on the test set is used as the evaluation metric to assign weights to each model. The formula for weighted voting is as follows:

[0042]

[0043] in, It is the prediction result of the i-th model, ω i These are the weights of the i-th model, which are allocated based on the performance of each model on the test set.

[0044] This invention, employing the above technical solutions, proposes an occupational health examination result diagnostic conclusion analysis and prediction model that integrates machine learning models and expert systems. This improves the intelligence level of occupational health management and provides strong support for the prevention and control of occupational diseases. Through this innovative model, occupational health risks can be identified more effectively, providing a scientific basis for formulating preventive measures and health policies. Compared with traditional models based on k-Nearest Neighbors, Support Vector Machines, and Random Forests, this invention improves the model's ability to identify small samples and significantly enhances the negative predictive value (NPV), ensuring the accuracy of predictions for key categories. It also enhances the model's adaptability and generalization ability to unknown situations, ensuring the model's effectiveness and reliability in practical applications. Attached Figure Description

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0046] Figure 1 This is a flowchart illustrating the occupational health examination result prediction method based on expert systems and machine learning according to the present invention.

[0047] Figure 2 This is a schematic diagram illustrating the process of simulating expert decision-making in the expert system of this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0049] This invention explores a predictive model for occupational health examination results diagnosis that integrates machine learning models and expert systems. XGBoost, as a highly efficient machine learning algorithm, excels in handling imbalanced data and improving prediction accuracy. By constructing and optimizing multiple decision trees, it effectively handles various complex data patterns. Expert systems, especially those based on knowledge reasoning, are widely used in the healthcare field due to their advantages in simulating expert decision-making processes and providing interpretive reasoning capabilities. The predictive model built by integrating XGBoost and expert system technologies can predict occupational health examination results more accurately and efficiently. Machine learning in the model identifies key patterns in the data through its powerful classification capabilities, while expert systems provide domain-knowledge-based reasoning mechanisms, enabling the model not only to make predictions but also to provide explanations for those predictions. This fusion approach combines the data processing capabilities of machine learning with the decision support capabilities of expert systems, allowing the predictive model to provide more reliable and transparent predictions when faced with complex occupational health data. Furthermore, while providing predictions, expert systems can also offer users possible diagnostic paths and treatment suggestions based on their internal knowledge base, which is invaluable for decision-makers in occupational health management. The introduction of expert systems also helps to improve the interpretability of the model, enabling non-experts to understand the basis of the model's predictions, thereby enhancing their trust in the model's prediction results.

[0050] like Figure 1 As shown in Figure 2, this invention discloses a method for predicting occupational health examination results based on expert systems and machine learning, which includes the following steps:

[0051] Includes the following steps:

[0052] Step 1, Data Processing: After cleaning and standardizing the physical examination records, visit frequency information is added to form the original dataset; some feature data of the original dataset are extracted by deep learning through a neural network model to obtain important features; the important features and the remaining feature data of the original dataset together form a new feature dataset;

[0053] Furthermore, step 1 specifically includes the following steps:

[0054] Step 1-1: Traverse the physical examination records and delete redundant feature information that is irrelevant to the physical examination conclusion;

[0055] Furthermore, the redundant feature information is personal information features, which include name, medical examination number, employer name, and employer's region.

[0056] Steps 1-2: Extract hazard factor information from the physical examination records; convert each individual hazard factor in the hazard factor information into a corresponding dummy variable to form a dummy variable set, and map the individual hazard factor name of each dummy variable to the hazard factor name classified in the specified occupational health standard, namely the "Technical Specification for Occupational Health Surveillance"; wherein, the hazard factor information includes physical examination hazard factors and exposure hazard factors;

[0057] Specifically, the two features of physical examination hazard factors and exposure hazard factors are obtained. Single hazard factors are identified using commas as separators and converted into dummy variables F. A dictionary is built to map the single hazard factor name of F to the hazard factor names classified in the specified occupational health standard, namely the "Technical Specification for Occupational Health Surveillance". The two features of physical examination hazard factors and exposure hazard factors are deleted, and dummy variables F are added.

[0058] Steps 1-3: Identify and obtain time feature information from the physical examination records, convert it into months, and save it as an integer.

[0059] Steps 1-4 involve extracting information from the physical examination records and adding information on the number of medical visits to form the original dataset.

[0060] Steps 1-5: Construct a neural network model. Input some feature data from the original dataset into the neural network model to perform deep learning on the original dataset to extract important features. The important features and the remaining feature data from the original dataset together form a new feature dataset.

[0061] Furthermore, the time characteristic information in steps 1-3 includes total length of service, length of service during the disaster, and start and end dates.

[0062] Furthermore, the neural network model in steps 1-5 includes an input layer, hidden layers, and an output layer. The input layer receives data with a dimension equal to the number of features in the training set. The hidden layer consists of two linear layers. The first linear layer maps the input features to a 256-dimensional feature vector. Dropout regularization is added after the first linear layer, with a dropout rate of 0.2. The second linear layer compresses the 256-dimensional feature vector output from the first linear layer into a 128-dimensional feature vector. Each linear layer is followed by a ReLU activation function to introduce non-linearity and enhance the model's expressive power. The linear layer of the output layer maps the 128-dimensional feature vector to the number of target categories, i.e., 5 dimensions.

[0063] Furthermore, in steps 1-5, the neural network model is trained using the cross-entropy loss function and the Adam optimizer.

[0064] Specifically, the neural network model performs the following actions: The input layer receives data with the same dimensionality as the number of features in the training set; then, the data undergoes feature mapping through two linear layers. The first linear layer maps the data to 256 dimensions, and the second linear layer further compresses it to 128 dimensions. After each linear layer, a ReLU activation function is applied to increase the model's non-linearity. Dropout regularization is added after the first linear layer, with a dropout rate of 0.2. Finally, a linear layer maps the 128-dimensional feature vector to 5 dimensions, completing the forward propagation process of the entire network. During training, the network is trained using the cross-entropy loss function and the Adam optimizer. After model training, the trained network extracts features from new data. Finally, a new dataset is obtained by combining the features extracted from the second linear layer data by the neural network with other features from the original dataset. Specifically, the features extracted from the second linear layer data by the neural network represent new features that influence the prediction results of the trained neural network; these can be understood as new latent features discovered through the correlation between certain key features in the dataset.

[0065] Step 2: Identify all types of hazard factors for each examinee in the new dataset; based on the provisions of the specified occupational health standard (specifically, the National Occupational Health Standard of the People's Republic of China GBZ 188-2014 "Technical Specifications for Occupational Health Surveillance"), perform logical judgment on each hazard factor to obtain the corresponding prediction label, and summarize them to form a decision set; label the decision set with occupational contraindications and suspected occupational diseases respectively, and add the corresponding features SOD and OC to the new dataset;

[0066] Furthermore, the predictive labels include "No abnormalities found so far", "Re-examination", "Suspected occupational disease", "Occupational contraindication" and other diseases or abnormalities.

[0067] Furthermore, the specific steps of step 2 are as follows:

[0068] Step 2-1: Identify all risk factor types for individuals undergoing physical examinations in the dataset: F = {f1, f2, ..., f n}, where n represents the number of single hazard factor features; for each physical examination record, iterate through the single hazard factor feature set F to obtain the single hazard factor feature set G corresponding to each physical examination record. The feature set G of the i-th physical examination record is G_i. i G = {g1, g2, ..., g} k}, where k is the number of single hazard features with a value of 1;

[0069] Specifically, as a feasible implementation method, the number of single hazard factor features n = 81; obtain the single hazard factor features with a value of 1 in the single hazard factor feature set F to form the single hazard factor feature set G for each physical examination record.

[0070] Step 2-2: Identify and obtain the physical examination feature t corresponding to the hazard factor f for each physical examination record, where T is the set of physical examination features t; the set of physical examination features T for the i-th physical examination record is Ti. i : T = {t1, t3, ..., t d ,d is the number of physical examination features};

[0071] Steps 2-3 involve constructing a logical function H for each hazard factor f, with the specific expression as follows:

[0072]

[0073] Steps 2-4: Establish a corresponding decision set D, D = {d1, d2, ..., dn} for the single hazard factor characteristic set G. k ,d i =H(f) i ,T i )},d i This represents the single hazard factor g corresponding to the single hazard factor characteristic set G. i The value determined by the logical function H; the set of physical examination characteristics D of the j-th physical examination record is D. j :

[0074] Steps 2-5: Label the decision set with occupational contraindications and suspected occupational diseases respectively to obtain new features SOD and OC, and add the new features SOD and OC to the new dataset;

[0075] Among them, the feature SOD represents the expert system's simulation of expert decision-making in the judgment of suspected occupational diseases. 1 indicates the presence of a suspected occupational disease, and 0 indicates its absence; the feature OC represents the expert system's simulation of expert decision-making regarding the judgment of occupational contraindications. 1 indicates the presence of occupational contraindications, and 0 indicates the absence of contraindications.

[0076] Step 3, Hybrid Model Training: The new dataset, after passing through the expert system, is randomly divided into a training set and a test set. Two or more machine learning models are selected and trained on the training set to obtain their respective prediction results. The prediction results of the two or more models are weighted and averaged to obtain the final output result through a weighted voting method.

[0077] Furthermore, four models were selected: Random Forest, AdaBoost, XGBoost, and LightGBM.

[0078] Furthermore, using the accuracy of the test set on each learning model as an evaluation metric, a grid search was employed to find the hyperparameters of each learning model.

[0079] Specifically, using accuracy as the evaluation metric, we employed grid search to find the hyperparameters of Random Forest, AdaBoost, XGBoost, and LightGBM models, respectively. Then, we used a weighted voting method to average the prediction results of the four models to obtain the final output.

[0080] Furthermore, the specific steps of step 3 are as follows:

[0081] Step 3-1: The new dataset that has passed through the expert system is randomly divided into a training set and a test set in a 7:3 ratio. Two or more machine learning models are trained on the training set. The hyperparameter range of each learning model is set using grid search. The accuracy of the predictions made by the model on the test set is used as the evaluation metric to find the hyperparameters of the model.

[0082] Specifically, the selected machine learning models were Random Forest, AdaBoost, XGBoost, and LightGBM.

[0083] Step 3-2: The prediction results of two or more models are weighted and averaged to obtain the final output result by weighted voting. The accuracy of each model on the test set is used as the evaluation metric to assign weights to each model.

[0084] The principles of this invention will be explained in detail below:

[0085] The dataset used in this invention was obtained from a coastal provincial capital city in China. This dataset contains 71,084 records and 1,535 features. It represents the occupational health examination records of employees in Fuzhou City during their on-the-job work in 2021-2022. These employees worked under different occupational hazard conditions. Occupational hazard factors refer to factors or conditions that arise and / or exist in occupational activities and may adversely affect the health, safety, and work capacity of the working population, including chemical, physical, and biological factors. Therefore, different examination items (represented as features in the dataset) are required based on different occupational hazard factors, and corresponding examination conclusions are drawn according to the "Technical Specifications for Occupational Health Surveillance" (GBZ188).

[0086] Based on the results of occupational health examinations, the conclusions of physical examinations for examinees can be divided into the following five categories: (1) No abnormalities found: All examination indicators in this occupational health examination are within the normal range; (2) Re-examination: If one or more abnormalities related to the target disease are found during the examination, and a re-examination is required for confirmation, the content and time of the re-examination should be specified; (3) Suspected occupational disease: If a suspected occupational disease or possible occupational disease is found during the examination, and further diagnosis by an occupational disease diagnosis institution is required; (4) Occupational contraindications: If a patient with occupational contraindications is found during the examination, the specific disease name should be specified; (5) Other diseases or abnormalities: Other diseases or abnormalities in certain examination indicators besides the target disease. The dataset contains records of five conclusions and their respective percentages: other diseases or abnormalities (34,300 records, 48.25%), no abnormalities found so far (26,950 records, 37.91%), re-examination (9,273 records, 13.0%), occupational contraindications (536 records, 0.75%), and suspected occupational diseases (25 records, 0.03%).

[0087] Data Processing: First, a comprehensive review and evaluation of the collected raw dataset was conducted. This dataset covers diverse physical examination results and related individual characteristic information, aiming to analyze the relationship between examination results and potential health hazard factors through machine learning models. However, due to the complexity of the data sources, the raw dataset contains various forms of anomalies and redundant information. If this information is not properly processed, it will interfere with subsequent analysis and affect the accuracy and reliability of the model.

[0088] To construct a high-quality dataset, this invention employs a series of data cleaning and standardization measures. First, I also removed personal information features not directly related to the medical examination results, such as name and examination number, to protect personal privacy and reduce data redundancy.

[0089] In processing time-based and character-based data, this invention employs standardized procedures. For time-based data, dates and years are converted into a unified format, and useful information (such as months and years) is extracted and represented as numerical data for easier subsequent calculations and analysis. For character-based data, this invention categorizes and encodes the data according to its actual meaning. For example, descriptive text of medical examination results is converted into concise classification labels such as "no abnormalities" and "abnormalities present" to facilitate model recognition and judgment.

[0090] In the actual process of diagnosing occupational health examination results, due to the subjective differences among diagnosing physicians, different judgments may be reached regarding the same examinee and the same examination data. This difference arises because the boundaries between the predicted labels "currently no abnormalities," "re-examination," "suspected occupational disease," "occupational contraindications," and other diseases or abnormalities are not always clear. The diagnostic biases and experience of the diagnosing physicians also influence their judgments. Typically, when an examination indicator shows abnormalities, the physician needs to comprehensively consider whether the abnormality clearly points to an occupational disease, or is more likely a manifestation of other diseases or abnormalities. In cases of uncertain diagnosis, physicians often choose the "re-examination" diagnosis to obtain more information. Therefore, during the initial examination, physicians may be more inclined to suggest a re-examination. To capture this dynamic change in the diagnostic process, this invention introduces the feature of "number of visits" into the dataset to reflect the re-examination status of examinees and the evolution of the diagnosis during the diagnostic process. The addition of this feature helps the model better understand the complexity of diagnostic decisions and improves the accuracy of diagnostic predictions. In this way, the model can learn the decision-making patterns of different diagnosing physicians when facing similar situations, thereby providing more accurate and consistent prediction results in practical applications.

[0091] Furthermore, to ensure data quality, this invention also meticulously screens and processes outliers and missing values ​​in the dataset. For ambiguous, inconsistent, or obviously erroneous data records, this invention corrects or deletes them; for missing medical examination conclusions or other key information, this invention adopts corresponding imputation or deletion strategies based on the severity and distribution of the missing data.

[0092] After data cleaning and normalization, this invention further performs feature engineering to improve the representational power of the dataset and the predictive performance of the model. Specifically, for complex and comprehensive hazard factor features in the original data, this invention decomposes them into multiple independent single hazard factor features. In the original dataset, hazard factors may be composed of combinations of multiple single hazard factors. To enable the model to clearly distinguish and evaluate the impact of each single hazard factor, this invention implements dummy variable transformation. In this process, each single hazard factor is represented as a dummy variable, where 0 indicates that the hazard factor is absent in the individual, and 1 indicates its presence. This step not only helps the model more accurately capture the impact of different hazard factors on physical examination results but also improves the interpretability of the data and the generalization ability of the model.

[0093] To better extract important features from the dataset, this invention employs deep learning techniques, particularly neural networks, to automatically extract features from the dataset. This invention designs a feedforward neural network whose structure includes multiple linear layers and activation layers to achieve efficient mapping from the original data to a high-dimensional feature space.

[0094] A neural network model consists of the following main parts, such as Figure 1 As shown:

[0095] Input layer: The network's input layer receives data with a dimension equal to the number of features in the training set, which ensures that the network can process all features.

[0096] Hidden layers: The network contains two linear layers. The first linear layer maps the input features to a 256-dimensional vector, and the second linear layer further compresses these features to 128 dimensions. Each linear layer is followed by a ReLU activation function to introduce non-linearity and enhance the model's expressive power.

[0097] Regularization: To prevent overfitting, this invention introduces Dropout regularization after the first hidden layer, with a dropout rate of 0.2.

[0098] Output layer: Finally, a linear layer maps the 128-dimensional feature vector to the number of target categories, i.e., 5 dimensions.

[0099] This invention uses the cross-entropy loss function and the Adam optimizer to train the network. After the model training is complete, this invention uses the trained network to extract features from new data. Ultimately, this invention obtains a dataset containing new features, which combines the features extracted by the neural network with other features from the original dataset.

[0100] An expert system is a computer program that simulates the decision-making abilities of human experts. It can handle complex decision-making problems and provide expert-level advice or solutions. The core of an expert system lies in its ability to store vast amounts of professional knowledge and experience, and to use this knowledge to reason and solve problems. In the field of occupational health, expert systems can simulate a doctor's judgment on suspected occupational diseases and occupational contraindications recorded in medical examinations in real-world situations.

[0101] This invention adopts the hazard factor regulations in the "Technical Specifications for Occupational Health Surveillance" to simulate a doctor's judgment on suspected occupational diseases and occupational contraindications in a physical examination record under actual circumstances. This expert system analyzes individual health data, such as lung function test results, biomarker levels in blood samples, and long-term occupational exposure history, to construct a feature space containing these health indicators. Using this information, combined with professional knowledge and experience, the system identifies individuals with suspected occupational diseases and occupational contraindications through logical reasoning by constructing logical functions, thereby achieving the prediction and diagnosis of occupational diseases.

[0102] (1) Identify all risk factor types for individuals undergoing physical examinations in the dataset:

[0103] F = {f1, f2, ..., f n n = 81 represents the number of characteristics of a single hazard factor.

[0104] For each physical examination record, iterate through the single hazard factor feature set F to obtain the single hazard factor feature set G with a value of 1. The feature set G of the i-th physical examination record is Gi. i .

[0105] G = {g1, g2, ..., g} k k is the number of single hazard factor characteristics with a value of 1.

[0106] Then, expert decision-making is simulated for each feature of set G.

[0107] (2) Make a judgment based on the provisions on hazard factors in the National Occupational Health Standard of the People's Republic of China GBZ 188-2014 "Technical Specification for Occupational Health Surveillance" in the knowledge base.

[0108] The specific reference is as follows: Chapter 5 concerns occupational health monitoring for workers exposed to hazardous chemicals, including lead and its inorganic compounds, tetraethyl lead, mercury and its inorganic compounds, manganese and its inorganic compounds, beryllium and its inorganic compounds, cadmium and its inorganic compounds, chromium and its inorganic compounds, zinc oxide, arsenic, arsine (arsine trihydrogen), phosphorus and its inorganic compounds, phosphine, barium compounds (barium chloride, barium nitrate, barium acetate), vanadium and its inorganic compounds, trialkyltin, thallium and its inorganic compounds, nickel carbonyl, fluorine and its inorganic compounds, benzene (for exposure to industrial toluene and xylene, refer to the guidelines), carbon disulfide, carbon tetrachloride, methanol, gasoline, bromomethane, and 1,2-dioxane. Ethyl chloride, n-hexane, amino and nitro compounds of benzene, trinitrotoluene, benzidine, chlorine, sulfur dioxide, nitrogen oxides, ammonia, phosgene, formaldehyde, monomethylamine, carbon monoxide, hydrogen sulfide, vinyl chloride, trichloroethylene, allyl chloride, chloroprene, organofluorine compounds, toluene diisocyanate, dimethylformamide, cyanide and nitrile compounds, phenols (phenolic compounds such as cresol, catechol, resorcinol, hydroquinone, etc. refer to the same), pentachlorophenol, chloromethyl ether [bis(chloromethyl) ether refer to the same], acrylamide, unsymmetrical dimethylhydrazine, dimethyl sulfate, organophosphorus pesticides, carbamate pesticides, pyrethroids, acid mist or acid anhydride, asthma-inducing substances, coke oven emissions. Chapter Six concerns occupational health monitoring for workers exposed to dust, including free silica dust (crystalline silica dust), coal dust, asbestos dust, other inorganic dusts that cause pneumoconiosis, cotton dust (including flax, hemp, and jute dust), and organic dust. Chapter Seven concerns occupational health monitoring for workers exposed to harmful physical factors, including noise, hand-transmitted vibration, high temperature, high pressure (see GB 20827), ultraviolet radiation (ultraviolet light), and microwaves. Chapter Eight concerns occupational health monitoring for workers exposed to harmful biological factors, including Brucella and Bacillus anthracis. Chapter Nine concerns occupational health monitoring for workers in special operations, including electrical work, work at heights, pressure vessel work, tuberculosis prevention and control, hepatitis prevention and control, professional motor vehicle driving, video surveillance work, high-altitude work, and aviation work.

[0109] Based on the provisions on hazard factors in the "Technical Specifications for Occupational Health Surveillance", this invention constructs a logical function for each hazard factor. The specific process is as follows:

[0110] Identify and obtain the physical examination characteristics t of the hazard factor f, where T is the set of physical examination characteristics t, and the set of physical examination characteristics T of the i-th physical examination record is T. i :

[0111] T = {t1, t2, ..., t} d d represents the number of physical examination features.

[0112] Construct a function H for each hazard factor f.

[0113]

[0114] (3) Simulated expert decision-making: Establish a corresponding decision set D for set G, and perform logical function judgment on each hazard factor f in set G. The set of physical examination characteristics D of the j-th physical examination record is D. j :

[0115] D = {d1, d2, ..., d} k ,d i =H(f) i ,T i )}

[0116] Add two new feature columns to the dataset: SOD (Suspected occupational disease) and OC (Occupational contraindications).

[0117]

[0118] In this context, feature SOD represents the expert system's simulated expert decision-making judgment on suspected occupational diseases, with 1 indicating the existence of a suspected occupational disease and 0 indicating its absence. Feature OC represents the expert system's simulated expert decision-making judgment on occupational contraindications, with 1 indicating the existence of an occupational contraindication and 0 indicating its absence.

[0119] This paper proposes a hybrid model based on multiple machine learning algorithms. By using four classic machine learning models—Random Forest, AdaBoost, XGBoost, and LightGBM—an ensemble model capable of multi-dimensional prediction is constructed, and the final output is obtained by weighted averaging of the prediction results from each model. To ensure the model's accuracy and generalization ability, the parameters of each model are meticulously tuned, and a weighted fusion strategy is employed to improve model performance. The following is a detailed description of the model training process.

[0120] (1) Dataset Overview and Preprocessing: The dataset used contains 71,084 records, each consisting of 428 features. Its complexity and high dimensionality present two problems: the model is prone to overfitting, and the training time is long. Therefore, this invention implemented a series of data cleaning and normalization measures in the data preprocessing stage described above. The dataset was also divided into training and testing sets.

[0121] (2) Main methods for model hyperparameter training: Grid Search is an exhaustive search method for hyperparameter tuning, aiming to find the optimal combination of hyperparameters for the machine learning model. Its basic principle is to predefine the possible value range of each hyperparameter, traverse all possible combinations within these ranges, train the model one by one and evaluate its performance on the validation set, and finally select the best-performing parameter combination.

[0122] Specifically, grid search first defines a parameter space, which is a set of possible values ​​for each hyperparameter. Then, by iterating through all possible combinations of these values, the model is trained and its performance is evaluated using a validation set. Each trained model is scored according to an accuracy metric. Finally, grid search selects the parameter combination that performs best on the validation set as the final hyperparameter settings.

[0123] (3) Model principle and formula: In order to make full use of the characteristics of different algorithms, this invention selects four mainstream machine learning models: Random Forest, AdaBoost, XGBoost and LightGBM.

[0124] Random Forest is an ensemble learning method based on decision trees. It improves the overall accuracy of the model by constructing multiple decision trees and voting on their results. The core idea of ​​Random Forest is to introduce randomness during training: in the construction of each tree, a subset of data and a subset of features are randomly selected, resulting in multiple independent decision trees. Finally, Random Forest makes a final prediction by voting on or averaging the outputs of all the decision trees.

[0125] The basic formula for random forest is:

[0126]

[0127] Where N is the number of decision trees, h i (x) is the prediction result of the i-th tree for the input x, and the final output is the average of the prediction results of all decision trees or the voting result.

[0128] The main parameters of the random forest model were tuned using grid search, and the specific optimization results are as follows:

[0129] 'n_estimators' (number of decision trees): The optimal value is 3000;

[0130] 'max_depth' (maximum depth of the tree): The optimal value is 13;

[0131] 'min_samples_split' (minimum number of samples required for internal node re-split): optimal value is 2;

[0132] 'min_samples_leaf' (minimum number of samples required for a leaf node): The optimal value is 4.

[0133] After optimization, the random forest achieved an accuracy of 0.9341 on the test set, demonstrating strong performance.

[0134] AdaBoost (Adaptive Boosting): AdaBoost is a boosting method designed to combine multiple weak classifiers (such as decision trees) into a strong classifier. It works by iteratively building the model, with each iteration focusing on the misclassified samples from the previous iteration. In each training round, the weights of misclassified samples increase, while the weights of correctly classified samples decrease, thus making the next weak classifier pay more attention to the misclassified samples from the previous round.

[0135] AdaBoost's update rules are as follows:

[0136]

[0137] Where, ω i h(x) is the weight of the i-th sample. i ) is the classifier's response to sample x i The prediction results, y i α is the true label of the sample, α is the weight of the classifier, and Ι represents the relative importance of the classifier.

[0138] The final strong classifier is the weighted result of all the weak classifiers:

[0139]

[0140] For the AdaBoost model, this invention uses the default base decision tree classifier and adjusts the number of iterations and learning rate based on grid search, ultimately achieving a test set accuracy of 0.9114.

[0141] XGBoost: XGBoost is a reinforcement learning algorithm based on gradient boosting. Compared with the traditional GBDT (Gradient Boosting Decision Tree), XGBoost optimizes regularization, missing value handling, and parallel processing, thereby improving the model's performance and efficiency.

[0142] XGBoost aims to minimize an objective function by iteratively constructing multiple decision trees. This objective function includes a loss function and a regularization term.

[0143]

[0144] in, It is a loss function that measures the model's predicted values. Compared with the true value y i The error between; Ω(f) k ) is a regularization term used to control the complexity of the model and prevent overfitting.

[0145] The core idea of ​​XGBoost is to gradually reduce the model's error by fitting the residuals of the current model with a new decision tree each time. Specifically, in each iteration, XGBoost updates the model parameters using gradient descent to minimize the sum of squares of the current residuals.

[0146] The main parameters of XGBoost were also optimized based on grid search, as follows:

[0147] 'n_estimators' (number of decision trees): The optimal value is 2000;

[0148] 'max_depth' (maximum depth of the tree): The optimal value is 3;

[0149] 'learning_rate': The optimal value is 2.

[0150] The optimized XGBoost model achieved an accuracy of 0.9575 on the test set, making it the best-performing model in this experiment.

[0151] LightGBM (Light Gradient Boosting Machine)

[0152] LightGBM (Light Gradient Boosting Machine) is an efficient gradient boosting framework, particularly suitable for large-scale datasets and high-dimensional features. Similar to XGBoost, LightGBM is also a decision tree-based gradient boosting method, but it is accelerated through histogram-based optimization.

[0153] The core idea of ​​LightGBM is to discretize continuous features into several discrete intervals and perform splitting decisions on the histogram, thereby significantly reducing computation. Its objective function is similar to XGBoost, also including a loss function and a regularization term.

[0154]

[0155] Where, f(x) i ) is the model for sample x i The predicted value is given by Ω, where l is the loss function and Ω is the regularization term used to control the complexity of the model.

[0156] The main parameters of LightGBM were optimized using a grid search, as follows:

[0157] 'n_estimators' (number of decision trees): The optimal value is 2000;

[0158] 'max_depth' (maximum depth of the tree): The optimal value is 3;

[0159] 'min_child_samples' (minimum number of samples for leaf nodes): The optimal value is 8;

[0160] 'min_split_gain' (minimum value of split gain): The optimal value is 0.

[0161] After optimization, the LightGBM model achieved an accuracy of 0.9020 on the test set. Although it performed slightly worse than the other four models, its efficiency in large-scale data processing is commendable.

[0162] (4) Model Fusion and Weighted Voting: Since each model performs differently under varying data characteristics, this invention employs a model fusion method to fully utilize the strengths of each model. Through weighted voting, the prediction results of the four models are weighted and averaged to obtain the final output. The formula for weighted voting is as follows:

[0163]

[0164] in, It is the prediction result of the i-th model, ω i These are the weights of the model, which are allocated based on the performance of each model on the test set.

[0165] Ultimately, the prediction results of the fused model outperformed the single model in both accuracy and robustness, with a test accuracy of 0.9667, demonstrating the advantages of the hybrid model.

[0166] By constructing a hybrid model based on Random Forest, AdaBoost, XGBoost, and LightGBM, the prediction problem on large-scale, high-dimensional datasets was solved. Each model underwent meticulous tuning and optimization, and finally, model fusion was performed through weighted voting to improve overall performance. The combined advantages of the various models resulted in high accuracy and stability for the system, enabling it to effectively handle complex data environments.

[0167] Compared with the prior art, the present invention has the following technical features: (1) The data processing of the present invention effectively solves the problems of complexity, validity and availability of data sources. The present invention extracts the most valuable information from massive data by implementing data cleaning, feature selection and dimensionality reduction techniques, while eliminating those data with strong interference and low credibility; it improves the quality of data and the accuracy of the model, and also significantly reduces the computational cost of data processing, making the data mining process more efficient and goal-oriented. (2) The present invention solves the problem of low prediction accuracy for small samples. Since occupational contraindications and suspected occupational diseases account for too small a proportion of the dataset, and the physical examination items to be tested are different due to different hazard factors, the logical judgment of the diagnosis results is complicated. Common mainstream models have poor prediction effects for small sample classification, and even make the prediction accuracy of occupational contraindications and suspected occupational diseases zero in order to blindly improve the overall accuracy. The present invention introduces an expert system and combines the knowledge and experience of domain experts to make effective judgments for the two key categories of occupational contraindications and suspected occupational diseases. This not only improves the model's ability to identify small samples, but also significantly improves the negative prediction value (NPV), ensuring the prediction accuracy of key categories. (3) This invention addresses the practicality of the model in real-world applications. Common mainstream models are highly likely to produce incorrect predictions when judging samples that do not exist in the dataset used for model training. For example, if the dataset does not contain tetraethyl lead as an occupational contraindication, the model trained on that dataset cannot distinguish tetraethyl lead as an occupational contraindication, which may lead to errors in predicting individuals with this occupational contraindication during physical examinations. This invention, by integrating an expert system, simulates the expert's decision-making process, enabling the model to accurately predict samples that have not appeared in the dataset. The introduction of the expert system enhances the model's adaptability and generalization ability to unknown situations, ensuring the model's effectiveness and reliability in practical applications. Furthermore, the expert system provides interpretability for the model's predictions, making the model's decision-making process more transparent and easier for medical personnel and physical examination personnel to understand and trust.

[0168] This invention, employing the above technical solutions, proposes an occupational health examination result diagnostic conclusion analysis and prediction model that integrates machine learning models and expert systems. This model is expected to improve the intelligence level of occupational health management and provide strong support for the prevention and control of occupational diseases. Through this innovative model, occupational health risks can be identified more effectively, providing a scientific basis for formulating preventive measures and health policies.

[0169] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. A method for predicting occupational health examination results based on expert systems and machine learning, characterized in that: Includes the following steps: Step 1, Data Processing: After cleaning and standardizing the physical examination records, visit frequency information is added to form the original dataset; some feature data of the original dataset are extracted by deep learning through a neural network model to obtain important features; The important features, together with the remaining feature data of the original dataset, form a new feature dataset; Step 2: Identify all hazard factors for each examinee in the new dataset; based on the hazard factors specified in the occupational health standards, perform logical judgments on each hazard factor to obtain corresponding prediction labels, and summarize them to form a decision set; label the decision set with occupational contraindications and suspected occupational diseases, and add the corresponding features SOD and OC to the new dataset; the specific steps of Step 2 are as follows: Step 2-1: Identify all risk factor types for individuals undergoing physical examinations in the dataset: , Indicates the number of characteristics of a single hazard factor. Indicates the first A single hazard factor characteristic; For each physical examination record, iterate through all hazard factor feature sets F to obtain the corresponding single hazard factor feature set G for each physical examination record. The single hazard characteristic G in the individual examination record is... ; , The number of single hazard factor characteristics with a value equal to 1; Step 2-2: Identify and obtain the risk factors for each physical examination record. The corresponding physical examination feature t, where T is the set of physical examination features t; the set of physical examination features for the i-th physical examination record. That is : , The number of physical examination features; Steps 2-3, for each hazard factor Construct the logical function H, with the following specific expression: ； Steps 2-4: Establish corresponding decision sets for the single hazard factor characteristic set G. , , , This represents the value of a single hazard factor in the characteristic set G of the corresponding single hazard factor under the logic function H; the decision set D of the j-th individual examination record is... ; Steps 2-5: Label the decision set with occupational contraindications and suspected occupational diseases respectively to obtain new features SOD and OC, and add the new features SOD and OC to the new dataset; Among them, the feature SOD represents the expert system's simulation of expert decision-making in the judgment of suspected occupational diseases. 1 indicates the presence of a suspected occupational disease, and 0 indicates its absence; the feature OC represents the expert system's simulation of expert decision-making regarding the judgment of occupational contraindications. 1 indicates the presence of an occupational contraindication certificate, and 0 indicates the absence of one. Step 3, Hybrid Model Training: The new dataset, after passing through the expert system, is randomly divided into a training set and a test set. Two or more machine learning models are selected and trained on the training set to obtain their respective prediction results. The prediction results of the two or more models are weighted and averaged to obtain the final output result through a weighted voting method.

2. The method for predicting occupational health examination results based on expert systems and machine learning according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1-1: Traverse the medical examination records and delete redundant feature information that is irrelevant to the medical examination conclusion. Redundant feature information is personal information feature, which includes name, medical examination number, employer name, and employer's region. Steps 1-2: Extract hazard factor information from the medical examination records; convert each individual hazard factor in the hazard factor information into a corresponding dummy variable to form a dummy variable set, and map the individual hazard factor name of each dummy variable to the hazard factor name classified in the specified occupational health standard; wherein, the hazard factor information includes medical examination hazard factors and exposure hazard factors; Steps 1-3: Identify and obtain time feature information from the physical examination records, convert it into months, and save it as an integer. Steps 1-4 involve extracting information from the physical examination records and adding information on the number of medical visits to form the original dataset. Steps 1-5: Construct a neural network model. Input some feature data from the original dataset into the neural network model to perform deep learning on the original dataset to extract important features. The important features and the remaining feature data from the original dataset together form a new feature dataset.

3. The method for predicting occupational health examination results based on expert systems and machine learning according to claim 2, characterized in that: The time feature information in steps 1-3 includes total length of service, length of service during the disaster, and start and end dates.

4. The method for predicting occupational health examination results based on expert systems and machine learning according to claim 2, characterized in that: The neural network model in steps 1-5 includes an input layer, hidden layers, and an output layer. The input layer receives data with a dimension equal to the number of features in the training set. The hidden layer consists of two linear layers. The first linear layer maps the input features to a 256-dimensional feature vector. Dropout regularization is applied after the first linear layer with a dropout rate of 0.

2. The second linear layer compresses the 256-dimensional feature vector output from the first linear layer into a 128-dimensional feature vector. The data features output from the second linear layer are used as important features. Each linear layer is followed by a ReLU activation function to introduce non-linearity and enhance the model's expressive power; the linear layers of the output layer map the 128-dimensional feature vector to the number of target categories.

5. The method for predicting occupational health examination results based on expert systems and machine learning according to claim 1, characterized in that: In steps 1-5, the neural network model is trained using the cross-entropy loss function and the Adam optimizer.

6. The method for predicting occupational health examination results based on expert systems and machine learning according to claim 1, characterized in that: In step 2, the predicted labels include "re-examination", "suspected occupational disease", "occupational contraindication" and other diseases or abnormalities.

7. The occupational health examination result prediction method based on expert system and machine learning according to claim 1, characterized in that: In step 3, the accuracy of the test set on each learning model is used as the evaluation metric, and grid search is used to find the hyperparameters of each learning model.

8. The method for predicting occupational health examination results based on expert systems and machine learning according to claim 1, characterized in that: The specific steps for step 3 are as follows: Step 3-1: The new dataset that has passed through the expert system is randomly divided into a training set and a test set in a 7:3 ratio; two or more machine learning models are trained on the training set respectively. The hyperparameter range of each learning model is set by grid search and the accuracy of the model's prediction on the test set is used as the evaluation metric to find the hyperparameters of the model. Step 3-2: The prediction results of two or more models are weighted and averaged to obtain the final output result by weighted voting. The accuracy of each model on the test set is used as the evaluation metric to assign weights to each model.

9. The occupational health examination result prediction method based on expert system and machine learning according to any one of claims 1, 7 or 8, characterized in that: The formula for weighted voting is as follows: ； in, It is the first The prediction results of each model It is the first The weights of the models are assigned based on each model's performance on the test set.

Images