Method and system for predicting individual exposure level of air pollutants

Abstract

The invention provides a method and system for predicting the individual exposure level of air pollutants, and the method comprises the steps: S1, collecting the original sample data information of a subject population to form an original data set, and dividing the original data set into a training subset and a test subset; s2, constructing an air pollutant crowd exposure level initial model; s3, training the initial model by using the training subset, and optimizing the initial model based on the combination of a regularization method of Lasso regression, an SHAP value weighting method and a recursive feature elimination algorithm; step S4-S5, obtaining model prediction data and evaluating prediction accuracy; and step S6, collecting input characteristic variables of a target population different from the subject population, and substituting the input characteristic variables of the target population into the prediction model subjected to accuracy evaluation to realize prediction of the air pollutant individual exposure level of the target population. The prediction model is better in fitting performance and high in accuracy.

Description

Technical Field

[0001] This invention belongs to the field of environmental exposure and human health technology, and in particular relates to a method and system for predicting individual exposure levels to air pollutants. Background Technology

[0002] Environmental exposure refers to the effects of environmental toxins on organisms through respiratory, digestive, or skin contact, or their influence on abiotic factors through migration and transformation. Health exposure routes refer to the ways in which chemical substances enter the human body after contact (such as ingestion, inhalation, or skin contact). In reality, environmental exposure is a common phenomenon, and it receives widespread attention due to the health risks it poses to the population.

[0003] Air pollutant exposure, as a crucial component of environmental exposure, has long been considered a limiting factor in the progress of air pollution epidemiology research due to measurement errors. Previous air pollution epidemiology studies have generally tended to use ambient air pollutant concentration data measured at the nearest fixed monitoring station to represent the actual air pollutant exposure levels of the subjects. However, the ambient air pollutant concentration data measured at these fixed monitoring stations differs from the individuals' actual air pollutant exposure levels. Using this simple substitution method may lead to significant errors in air pollution epidemiological predictions or even render predictions invalid.

[0004] Currently, researchers have proposed predictive models for individual exposure levels to air pollutants. These models include input characteristic variables such as meteorological factors and individual measured data. However, these models still have shortcomings in selecting the input characteristic variables that contribute most to the prediction results, leaving room for optimization.

[0005] Based on the above, this application provides a technical solution to solve the above technical problems. Summary of the Invention

[0006] In a first aspect, the present invention provides a method for predicting an individual's exposure level to air pollutants, comprising: Step S1: Collect raw sample data of the subject population and organize the raw sample data into a raw dataset. Divide the raw dataset into a training subset and a test subset according to the random principle. Both the training subset and the test subset include input feature variables and individual air pollutant exposure level monitoring data. Step S2: Based on the training subset, construct an initial model of the population exposure level to air pollutants using a nonlinear fitting algorithm; Step S3: Train the initial model using the training subset to obtain the air pollutant exposure level prediction model, and optimize the air pollutant exposure level prediction model based on the combination of Lasso regression regularization method, SHAP value weighting method and recursive feature elimination algorithm. Step S4: Input the input feature variables from the test subset into the optimized air pollutant exposure level prediction model, predict the individual air pollutant exposure level of the subject population, obtain the model prediction data, compare the individual air pollutant exposure level monitoring data from the test subset with the model prediction data, and evaluate the accuracy of the prediction model. Step S5: If the accuracy evaluation of the prediction model passes, proceed to step S6; otherwise, repeat steps S2 to S4 until the accuracy evaluation of the prediction model passes. Step S6: Collect input characteristic variables from a target population that is different from the subject population, and substitute the input characteristic variables of the target population into the prediction model that has been evaluated for accuracy, so as to predict the individual exposure level of air pollutants for the target population.

[0007] Furthermore, in step S1, the original sample data information includes: basic information of the subject, housing information, activity pattern data, environmental parameter data, and individual air pollutant exposure level monitoring data.

[0008] Furthermore, in step S1, basic information, housing conditions, and activity patterns of the subject population are collected through community surveys and questionnaires, and environmental parameter data and individual air pollutant exposure level monitoring data are monitored through monitoring stations and active sampling equipment.

[0009] Furthermore, the step of organizing the original sample data information into an original dataset includes: Step S1.1: Clean the raw sample data of the collected subject population and remove missing values ​​and outliers; Step S1.2: Transform the unstructured data in the original sample data into structured data to form an original dataset that can be directly used for modeling.

[0010] Furthermore, step S1 also includes: performing multiple rounds of cross-validation on the data in the training subset using a grid search method. Each round of cross-validation includes the number of variables and trees that attempt to split the data, thereby selecting the initial basic parameters of the final air pollutant exposure level model.

[0011] Furthermore, in step S2, the nonlinear fitting algorithm is an ensemble learning algorithm based on decision trees, which includes any one of random forest, gradient boosting decision tree, lightweight gradient boosting machine, extreme gradient boosting tree, and categorical feature boosting.

[0012] Furthermore, in step S3, the air pollutant population exposure level prediction model is optimized by combining the Lasso regression regularization method, the SHAP value weighting method, and the recursive feature elimination algorithm, including: Step S3.1: Using the regularization method of Lasso regression, an objective function with an L1 norm penalty term is introduced to compress the regression coefficients of the input feature variables in the initial model. By adjusting the penalty coefficient, the automatic dimensionality reduction of the input feature variable space is achieved, resulting in input feature variables that can explain the prediction of the population exposure level of air pollutants. Step S3.2: Use the SHAP value weighting method to quantify the marginal contribution of the input feature variables that have explanatory power for predicting the population exposure level of air pollutants to the initial model, and establish an importance ranking matrix of the input feature variables; Step S3.3: Use the recursive feature elimination algorithm to carry out iterative optimization. Use SHAP weights as the basis for feature removal. Evaluate the predictive performance of each input feature variable in the importance ranking matrix through cross-validation. Eliminate the input feature variables with the lowest contribution in each round until the optimal subset of input feature variables for prediction is determined. Step S3.4: Reconstruct the air pollutant exposure level prediction model based on the optimal subset of predictor variables, and output the ranking results of the importance of the input feature variables.

[0013] Furthermore, step S4 also includes: Step S4.1: Obtain individual air pollutant exposure level monitoring data and outdoor air pollutant concentration at the same time to obtain the first ratio of individual air pollutant exposure level monitoring data to outdoor air pollutant concentration; Step S4.2: Obtain the individual exposure level data output by the air pollutant population exposure level prediction model after accuracy assessment and the outdoor air pollutant concentration at the same time to obtain the second ratio of individual exposure level data to outdoor air pollutant concentration; Step S4.3: Verify the accuracy of the air pollutant population exposure level prediction model in predicting individual exposure levels based on the first ratio and the second ratio.

[0014] In a second aspect, the present invention provides a system for predicting an individual's exposure level to air pollutants, comprising: The data acquisition and preprocessing module collects raw sample data from the subject population and organizes the raw sample data into a raw dataset. Based on a random principle, the raw dataset is divided into a training subset and a test subset. Both the training subset and the test subset include input feature variables and individual air pollutant exposure level monitoring data. An initial model building module, which constructs an initial model of the population exposure level to air pollutants based on the training subset using a nonlinear fitting algorithm; An initial model training module trains the initial model using the training subset to obtain an air pollutant exposure level prediction model, and optimizes the air pollutant exposure level prediction model based on a combination of Lasso regression regularization, SHAP value weighting, and recursive feature elimination algorithm. The testing module inputs the input feature variables from the test subset into the optimized air pollutant exposure level prediction model, predicts the individual air pollutant exposure level of the subject population, obtains model prediction data, and compares the individual air pollutant exposure level monitoring data from the test subset with the model prediction data to evaluate the accuracy of the prediction model. The prediction module collects input characteristic variables from a target population that is different from the subject population, and substitutes the input characteristic variables of the target population into a prediction model that has been evaluated for accuracy, so as to predict the individual exposure level of air pollutants for the target population.

[0015] This invention can bring at least one of the following beneficial effects: This invention provides a method for predicting individual air pollutant exposure levels. The process of predicting air pollutant exposure levels in the population involves model training, optimization, and accuracy verification. At the model optimization level, this invention uses a combination of Lasso regression regularization, SHAP value weighting, and recursive feature elimination algorithm to optimize the air pollutant exposure level prediction model. This model demonstrates superior optimization performance compared to existing models that use the Boruta algorithm alone.

[0016] Specifically, the regularization method of Lasso regression is adopted, and an objective function with an L1 norm penalty term is introduced to compress the regression coefficients of the input feature variables in the initial model. By adjusting the penalty coefficient, the dimensionality of the input feature variable space is automatically reduced, resulting in input feature variables that can explain the prediction of air pollutant exposure levels. The SHAP value weighting method is used to quantify the marginal contribution of the input feature variables that can explain the prediction of air pollutant exposure levels to the initial model, and an importance ranking matrix of the input feature variables is established. The recursive feature elimination algorithm is used for iterative optimization, with SHAP weights as the basis for feature removal. The predictive efficacy of each input feature variable in the importance ranking matrix is ​​evaluated through cross-validation. The input feature variables with the lowest contribution are removed round by round until the optimal subset of input feature variables for prediction is determined. Attached Figure Description

[0017] The preferred embodiments will now be described in a clear and easy-to-understand manner, with reference to the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods.

[0018] Figure 1 This is a schematic diagram illustrating the steps of a method for predicting individual air pollutant exposure levels proposed in an embodiment of the present invention; Figure 2 The graph shows the analysis results of the Bland-Altman method using the Boruta algorithm. Figure 3 The graph shows the analysis results of the Bland-Altman method for the air pollutant exposure level prediction model proposed in this embodiment of the invention. Figure 4 The ratio of predicted individual exposure concentrations to observed individual exposure concentrations obtained using the Boruta algorithm is shown. Figure 5 This is the ratio of the predicted individual exposure concentration to the observed individual exposure concentration based on the air pollutant population exposure level prediction model proposed in this embodiment of the invention. Figure 6 This refers to the ratio of the predicted individual exposure concentration data to the actual outdoor exposure concentration of air pollutants output by the population air pollutant individual exposure concentration prediction model based on the optimized variable screening method system proposed in this embodiment of the invention. Detailed Implementation

[0019] The various aspects of the present invention will be further described in detail below.

[0020] Unless otherwise defined or stated, all technical and scientific terms used herein have the same meaning as are familiar to those skilled in the art. Furthermore, any methods and materials similar to or equivalent to those described herein may be used in this invention.

[0021] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] Terminology Explanation SHAP (SHapley Additive exPlanations) is a model interpretability algorithm derived from the "Shapley Value" in game theory. Originally used to quantify "the contribution of each participant to the final payoff in a multiplayer cooperative game", it has been adapted in the field of machine learning to "quantify the marginal contribution of each input feature to the model's prediction results" and to highlight the role of high-contribution features through "weighting".

[0023] The core logic of the Recursive Feature Elimination (RFE) algorithm is to "start with all features, and gradually reduce the feature set through the process of 'training the model, evaluating the importance of features, removing the least important features, and repeating the iteration', and finally find the 'smallest feature subset with the best prediction effect'".

[0024] Example 1

[0025] Firstly, see Appendix Figure 1 This invention provides a method for predicting an individual's air pollutant exposure level, comprising: Step S1: Collect raw sample data of the subject population and organize the raw sample data into a raw dataset. Divide the raw dataset into a training subset and a test subset according to the random principle. Both the training subset and the test subset include input feature variables and individual air pollutant exposure level monitoring data.

[0026] The original sample data information includes: (1) basic information of the subjects, such as age, gender, and whether they smoke; (2) housing information, such as ventilation and whether air conditioning is installed; (3) activity pattern information, such as commuting time, indoor time, and outdoor exercise frequency; (4) environmental parameter information, such as outdoor air pollutant concentration, temperature, humidity, and wind speed; (5) individual air pollutant exposure level monitoring data, where the air pollutants are: particulate matter (PM2.5, PM10), gaseous pollutants (sulfur dioxide, nitrogen dioxide, ozone, carbon monoxide, volatile organic compounds), and special pollutants; the above original sample data information comprehensively covers the dimensions of the subjects themselves, individual behavior, and individual environment, avoiding deviations in the initial model construction of the population exposure level of air pollutants due to the lack of original sample data information.

[0027] The individual air pollutant exposure level monitoring data were obtained by active sampling devices (MicroPEM sampler, DustTrak II sampler), passive samplers (diffusion samplers), or real-time sensor monitoring devices (Plantower PMS5003 portable PM2.5 sensor).

[0028] The process of organizing the original sample data information into an original dataset includes: Step S1.1: Clean the raw sample data of the collected subject population and remove missing values ​​and outliers; Step S1.2: Transform the unstructured data in the original sample data into structured data to form a unified original dataset that can be directly used for modeling.

[0029] It should be noted that the missing values ​​are activity pattern data that was omitted from the subject population, and the outliers are extreme values ​​caused by equipment malfunction. Examples of converting unstructured data into structured data include converting the textual data information of "yes" or "no" into "1" or "0", where "1" represents "smoking" or "air conditioning installed", and "0" represents "non-smoking" or "air conditioning not installed"; or converting the textual data information of gender, such as "male" or "female", into "1" or "0", where "1" represents "male" and "0" represents "female".

[0030] The original dataset is divided into a training subset and a test subset according to a random principle, wherein the training subset accounts for 90% of the original dataset and the test subset accounts for 10% of the original dataset, that is, the ratio of the number of sample data in the training subset to the number of sample data in the test subset is 9:1. It should be understood that the division of the training subset and the test subset includes, but is not limited to, the above ratio. Any addition or deletion of any form should be included in the spirit of the present invention as long as it does not limit the purpose of the invention.

[0031] Step S2: Based on the training subset, construct an initial model of the population exposure level to air pollutants using a nonlinear fitting algorithm; The nonlinear fitting algorithm is a decision tree-based ensemble learning algorithm. Preferred ensemble learning algorithms are selected from any one of Random Forest, Gradient Boosted Decision Trees (GBDT), Lightweight Gradient Boosting Machine (LGB), eXtreme Gradient Boosting (XGB), and Categorical Boosting (CATB). It should be understood that the aforementioned decision tree-based ensemble learning algorithm can fit the complex relationships that superimpose the interaction between the individual, their behavior, and their environment within the subject population.

[0032] Step S3: Train the initial model using the training subset to obtain the air pollutant exposure level prediction model, and optimize the air pollutant exposure level prediction model based on the combination of Lasso regression regularization method, SHAP value weighting method and recursive feature elimination algorithm. In step S3, the air pollutant population exposure level prediction model is optimized by combining the Lasso regression regularization method, the SHAP value weighting method, and the recursive feature elimination algorithm, including: Step S3.1: Using the regularization method of Lasso regression, an objective function with an L1 norm penalty term is introduced to compress the regression coefficients of the input feature variables in the initial model. By adjusting the penalty coefficient, the automatic dimensionality reduction of the input feature variable space is achieved, resulting in input feature variables that can explain the prediction of the population exposure level of air pollutants. Specifically, before step S3.1, a preprocessing step for the input feature variables is included. For example, the input feature variables with specific values ​​in each category, such as age, commuting time, outdoor PM2.5 concentration, and outdoor temperature, in the training subset are processed using the Z-score standardization formula to obtain Z-score values.

[0033] in, Input the true values ​​of the feature variables for the corresponding categories. Input the mean of the feature variables for the corresponding category. Input the standard deviation of the feature variables for the corresponding category; For example: the actual outdoor PM2.5 concentration is 35 μg / m³, if its corresponding... , The Z-score obtained after standardization It should be understood that applying the Z-score standardization formula to input feature variables with specific numerical values ​​for each category can eliminate the magnitude difference between input feature variables with specific numerical values ​​in different categories, ensuring the fairness of the Lasso regression regularization method in penalizing all input feature variables. Secondly, after completing the preprocessing step of the input feature variables, a Lasso regression model can be constructed, using the individual air pollutant exposure level monitoring data in the training subset as the target variable; and the input feature variables after Z-score standardization and structured processing as independent variables; thus obtaining the objective function with an L1 norm penalty term:

[0034] in, For the intercept term, For all A vector consisting of the regression coefficients of the input feature variables. For the first vector The regression coefficients of the input feature variables, The number of sample data items in the training subset. For the first The true value of the target variable in each sample data set. For the first The sample in the first The values ​​of each input feature variable The penalty coefficient is... The value is not less than 0; Secondly, considering the size of the input feature variables (see step S1), the size of the input feature variables is between 20 and 30, and the penalty coefficient... The candidate values ​​are generated by logarithmic uniform sampling, resulting in 50-100 values, with a range of [value range missing]. This ensures coverage of the entire range from "weak penalty (retaining more input feature variables)" to "strong penalty (retaining fewer input feature variables)"; for each penalty coefficient The mean absolute error is calculated, and the optimal penalty coefficient is determined based on the principle that the regression coefficient with the smallest and non-zero mean absolute error corresponds to a moderate number of input feature variables. For example: optimal penalty coefficient At that time, the number of input feature variables retained was 18, resulting in the final Lasso regression model; Finally, with the optimal penalty coefficient Train the final Lasso regression model and output the regression coefficients of all input feature variables. ,reserve Input feature variables, remove The input feature variables are those that can be used to predict the level of public exposure to air pollutants.

[0035] Step S3.2: Use the SHAP value weighting method to quantify the marginal contribution of the input feature variables that have explanatory power for predicting the population exposure level of air pollutants to the initial model, and establish an importance ranking matrix of the input feature variables; Specifically, based on the regression coefficients selected in step S3.1 The input feature variables are used as input to calculate the global SHAP value of all sample data in the training subset. The larger the global SHAP value, the greater the marginal contribution of the corresponding input feature variable that has explanatory power for predicting the level of air pollutant exposure to the population to the initial model, and vice versa. The input feature variables are sorted in descending order using the global SHAP value as the weight, resulting in an importance ranking matrix for the input feature variables.

[0036] Step S3.3: Use the recursive feature elimination algorithm to carry out iterative optimization. Use SHAP weights as the basis for feature removal. Evaluate the predictive performance of each input feature variable in the importance ranking matrix through cross-validation. Eliminate the input feature variables with the lowest contribution in each round until the optimal subset of input feature variables for prediction is determined. Specifically, based on the global SHAP value in step S3.2, a recursive feature elimination algorithm is used to eliminate input feature variables with low contribution through multiple rounds of iteration. The predictive efficacy of each input feature variable in the importance ranking matrix is ​​evaluated through cross-validation. The input feature variables with the lowest contribution are eliminated round by round until the optimal subset of predictive variables with the largest coefficient of determination and the smallest mean absolute error is determined. Step S3.4: Reconstruct the air pollutant exposure level prediction model based on the optimal subset of predictor variables, and output the ranking results of the importance of the input feature variables.

[0037] Step S4: Input the input feature variables from the test subset into the optimized air pollutant exposure level prediction model, predict the individual air pollutant exposure level of the subject population, obtain the model prediction data, compare the individual air pollutant exposure level monitoring data from the test subset with the model prediction data, and evaluate the accuracy of the prediction model. Step S5: If the accuracy evaluation of the prediction model passes, proceed to step S6; otherwise, repeat steps S2 to S4 until the accuracy evaluation of the prediction model passes. Step S6: Collect input characteristic variables from a target population that is different from the subject population, and substitute the input characteristic variables of the target population into the prediction model that has been evaluated for accuracy, so as to predict the individual exposure level of air pollutants for the target population.

[0038] As a preferred implementation, step S1 further includes: performing multiple rounds of cross-validation on the data in the training subset using a grid search method, with each round of cross-validation including the number of variables and trees for validation split attempts, thereby selecting the initial basic parameters of the final air pollutant exposure level model.

[0039] It should be understood that cross-validation based on a training subset, by testing different combinations of the number of variables and the number of trees in the splitting attempts, can evaluate and select the initial basic parameters of the final air pollutant exposure level model, and thus adapt to the initial model construction in the subsequent step S2, avoiding overfitting or underfitting caused by blindly setting the basic parameters of the initial model. Specifically, the grid search method is used to perform multiple rounds of cross-validation on the data in the training subset, including: If the training subset in step S1 contains N (N is a positive integer and N is not a prime number) sample data, the training subset is split into K mutually exclusive subsets (K is a positive integer), and each subset contains N / K sample data. The distribution of input feature variables and the distribution of individual air pollutant exposure level monitoring data in each subset are consistent with the overall training subset. For example, the proportion of smokers in each subset differs from the proportion of smokers in the training subset by ≤5%, and the mean outdoor PM2.5 concentration in each subset differs from the mean outdoor PM2.5 concentration in the training subset by ≤5%. Define the specific types of nonlinear fitting algorithms, and set multiple combinations of candidate values ​​for the number of variables and trees in various split attempts. Cross-validate and evaluate each combination of candidate values ​​for the number of variables and trees in each split attempt, and calculate the coefficient of determination and the mean absolute error. Repeat the operation of splitting the training subset into K mutually exclusive subsets (K is a positive integer) for 3-10 rounds, preferably 4-9 rounds, more preferably 5-8 rounds, and even more preferably 6-7 rounds; repeat the above operation steps in each round, and obtain the optimal initial model basic parameters of the air pollutant exposure level according to the principle of maximizing the coefficient of determination and minimizing the mean absolute error. If the nonlinear fitting algorithm is chosen as random forest, cross-validation will select the optimal combination of basic parameter values ​​for the number of variables, the number of trees, and the maximum tree depth for the splitting attempt; if the nonlinear fitting algorithm is chosen as lightweight gradient booster, cross-validation will select the optimal combination of basic parameter values ​​for the number of variables, the number of trees, the learning rate, and the tree depth for the splitting attempt.

[0040] In a preferred implementation, in step S1, basic information, housing data, and activity pattern data of the subject population are collected through community surveys and questionnaires. Environmental parameter data and individual air pollutant exposure level monitoring data are monitored through monitoring stations and active sampling devices. Specifically, the subjects carry active sampling devices for three consecutive days, and the active sampling devices are calibrated in real time before sampling. The subject population comes from seven different locations across the country, and each location has a monitoring station closest to the subject population's residence.

[0041] As a preferred embodiment, in step S3, the initial model is trained using the training subset to obtain the air pollutant population exposure level prediction model, which can be implemented using R or Python.

[0042] It should be noted that, see appendix Figure 2 and appendix Figure 3 In step S3, the analysis results of the Bland-Altman method show that: if the air pollutant exposure level prediction model optimized by the Boruta algorithm alone is used, 4 out of 59 paired measurements in the test subset exceed the 95% consensus limit; while the air pollutant exposure level prediction model optimized by the Lasso regression regularization method combined with the SHAP value weighting method and the recursive feature elimination algorithm shows that 2 out of 33 paired measurements in the test subset exceed the 95% consensus limit. Therefore, the overall performance of the Lasso regression regularization method combined with the SHAP value weighting method and the recursive feature elimination algorithm is better than using the Boruta algorithm alone.

[0043] In other words, using the Boruta algorithm alone did not achieve the expected optimization effect on the air pollutant exposure level prediction model. However, the regularization method based on Lasso regression combined with the SHAP value weighting method and the recursive feature elimination algorithm further improved the performance of the air pollutant exposure level prediction model, forming a reliable synergistic mechanism.

[0044] Similarly, while Lasso regression-based regularization methods can reduce dimensionality, they cannot explain the importance of feature variables. The SHAP value weighting method can quantify the marginal contribution of input feature variables to the initial model but cannot select the optimal subset of predictor variables. The recursive feature elimination algorithm can determine the optimal subset of predictor variables but depends on the importance ranking of input feature variables. Therefore, the Lasso regression-based regularization method combined with the SHAP value weighting method and the recursive feature elimination algorithm proposed in this application has significant advantages in terms of the optimization effect of the air pollutant exposure level prediction model.

[0045] For further details, please see the appendix. Figure 4 and appendix Figure 5 Linear regression data of individual air pollutant (PM2.5 concentration) exposure levels predicted and observed using the Boruta algorithm alone showed that R 2 =0.81, with a slope of 1.09. Further analysis using a Lasso regression-based regularization method combined with SHAP value weighting and recursive feature elimination algorithm to predict individual air pollutant (PM2.5 concentration) exposure levels using a predictive model and linear regression data showed that R0 was 0.81. 2 =0.8535, with a slope of 1.109. The data above also shows that the Lasso regression-based regularization method combined with the SHAP value weighting method and the recursive feature elimination algorithm outperforms the Boruta algorithm alone, and can well fit the prediction model for individual air pollutant (PM2.5 concentration) exposure levels.

[0046] In a preferred embodiment, step 4 further includes: Step S4.1: Obtain individual air pollutant exposure level monitoring data and outdoor air pollutant concentration at the same time to obtain the first ratio of individual air pollutant exposure level monitoring data to outdoor air pollutant concentration; Step S4.2: Obtain the individual exposure level data output by the air pollutant population exposure level prediction model after accuracy assessment and the outdoor air pollutant concentration at the same time to obtain the second ratio of individual exposure level data to outdoor air pollutant concentration; Step S4.3: Verify the accuracy of the air pollutant population exposure level prediction model in predicting individual exposure levels based on the first ratio and the second ratio.

[0047] Taking PM2.5 concentration as an example, it's easy to see from comparing different PM2.5 concentrations that when the ambient PM2.5 concentration is greater than 50 μg / m³... 3 At that time, the ratio of the predicted individual exposure level data output by the air pollutant population exposure level prediction model to the outdoor air pollutant concentration value was less than 1:1, and this ratio would increase as the ambient PM2.5 concentration decreased, as detailed in the attached figure. Figure 6 As shown.

[0048] Among them, the appendix Figure 6 The black dots represent the ratio of an individual's PM2.5 exposure level monitoring data to the outdoor PM2.5 concentration, while the red asterisks represent the ratio of an individual's predicted PM2.5 exposure level to the outdoor PM2.5 concentration. It can be seen that as the environmental concentration increases, the changes in monitoring data and predicted PM2.5 exposure levels generally follow the same pattern. (Appendix) Figure 6 Further optimization of variable selection strategy was adopted, and similar monitoring data and predicted PM2.5 exposure level change patterns were observed.

[0049] In one specific embodiment of the present invention, data information of a target population that is different from the subject population is imported through a level to obtain the individual exposure level assessment results of the target population, thereby achieving accurate prediction of the exposure level of unsampled individuals.

[0050] The study found that outdoor PM2.5 concentration had the greatest impact on individual exposure levels, indicating that outdoor particulate matter concentration has a significant influence on individual exposure levels. Indoor passive smoking also made a significant contribution to individual exposure levels; the indoor PM2.5 concentration for smokers could be 2.5 times higher than that for non-smokers, indicating that smoking is a significant source of particulate matter. Temperature ranked high in the feature importance ranking, and its contribution may stem from the effect of temperature on indoor and outdoor air exchange rates. Daily activities, including commuting time, sleep time, and indoor time, ranked high in feature importance, indicating that the duration of time spent in different microenvironments has a significant impact on individual particulate matter exposure, suggesting that considering the duration of time spent in different microenvironments is important when estimating individual exposure levels in epidemiological studies.

[0051] The air pollutant exposure level prediction model for accurate assessment provided by this invention can well fit nonlinear relationships. Compared with linear regression models of the same study design, this model shows better performance. Furthermore, the model can be applied to various scenarios. The model establishment process, including model training, optimization, and sensitivity analysis, demonstrates its reliability in assessing the spatiotemporal distribution of pollutants, solving current problems such as inaccurate pollutant concentration levels, difficulty in obtaining parameters for large sample populations, and insufficient simulation accuracy. Further evaluation and verification of the model's performance on a cohort dataset with identical variable information showed that the predicted individual exposure concentration and the ratio of the monitored individual exposure concentration to the monitored outdoor exposure concentration have similar distributions, strongly supporting the accuracy of the assessment model.

[0052] Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or practice the method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0053] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing description of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for predicting an individual's air pollutant exposure level, characterized in that, include: Step S1: Collect raw sample data of the subject population and organize the raw sample data into a raw dataset. Divide the raw dataset into a training subset and a test subset according to the random principle. Both the training subset and the test subset include input feature variables and individual air pollutant exposure level monitoring data. Step S2: Based on the training subset, construct an initial model of the population exposure level to air pollutants using a nonlinear fitting algorithm; Step S3: Train the initial model using the training subset to obtain the air pollutant exposure level prediction model, and optimize the air pollutant exposure level prediction model based on the combination of Lasso regression regularization method, SHAP value weighting method and recursive feature elimination algorithm. Step S4: Input the input feature variables from the test subset into the optimized air pollutant exposure level prediction model, predict the individual air pollutant exposure level of the subject population, obtain the model prediction data, compare the individual air pollutant exposure level monitoring data from the test subset with the model prediction data, and evaluate the accuracy of the prediction model. Step S5: If the accuracy evaluation of the prediction model passes, proceed to step S6; otherwise, repeat steps S2 to S4 until the accuracy evaluation of the prediction model passes. Step S6: Collect input characteristic variables from a target population that is different from the subject population, and substitute the input characteristic variables of the target population into the prediction model that has been evaluated for accuracy, so as to predict the individual exposure level of air pollutants for the target population.

2. The method for predicting individual air pollutant exposure levels according to claim 1, characterized in that, In step S1, the original sample data information includes: basic information of the subject, housing information, activity pattern data, environmental parameter data, and individual air pollutant exposure level monitoring data.

3. The method for predicting individual air pollutant exposure levels according to claim 2, characterized in that, In step S1, basic information, housing conditions, and activity patterns of the subject population are collected through community surveys and questionnaires. Environmental parameter data and individual air pollutant exposure level monitoring data are monitored through monitoring stations and active sampling equipment, respectively.

4. The method for predicting individual air pollutant exposure levels according to claim 3, characterized in that, The process of organizing the original sample data information into an original dataset includes: Step S1.1: Clean the raw sample data of the collected subject population and remove missing values ​​and outliers; Step S1.2: Transform the unstructured data in the original sample data into structured data to form an original dataset that can be directly used for modeling.

5. The method for predicting individual air pollutant exposure levels according to claim 4, characterized in that, Step S1 further includes: performing multiple rounds of cross-validation on the data in the training subset using a grid search method. Each round of cross-validation includes the number of variables and trees that attempt to split the data, thereby selecting the initial basic parameters of the final air pollutant exposure level model.

6. The method for predicting individual air pollutant exposure levels according to claim 5, characterized in that, In step S2, the nonlinear fitting algorithm includes an ensemble learning algorithm based on decision trees, wherein the ensemble learning algorithm is selected from any one of random forest, gradient boosting decision tree, lightweight gradient boosting machine, extreme gradient boosting tree, and categorical feature boosting.

7. The method for predicting individual air pollutant exposure levels according to claim 6, characterized in that, In step S3, the air pollutant population exposure level prediction model is optimized by combining the Lasso regression regularization method, the SHAP value weighting method, and the recursive feature elimination algorithm, including: Step S3.1: Using the regularization method of Lasso regression, an objective function with an L1 norm penalty term is introduced to compress the regression coefficients of the input feature variables in the initial model. By adjusting the penalty coefficient, the automatic dimensionality reduction of the input feature variable space is achieved, resulting in input feature variables that can explain the prediction of the population exposure level of air pollutants. Step S3.2: Use the SHAP value weighting method to quantify the marginal contribution of the input feature variables that have explanatory power for predicting the population exposure level of air pollutants to the initial model, and establish an importance ranking matrix of the input feature variables; Step S3.3: Use the recursive feature elimination algorithm to carry out iterative optimization. Use SHAP weights as the basis for feature removal. Evaluate the predictive performance of each input feature variable in the importance ranking matrix through cross-validation. Eliminate the input feature variables with the lowest contribution in each round until the optimal subset of input feature variables for prediction is determined. Step S3.4: Reconstruct the air pollutant exposure level prediction model based on the optimal subset of predictor variables, and output the ranking results of the importance of the input feature variables.

8. The method for predicting individual air pollutant exposure levels according to claim 7, characterized in that, Step S4 further includes: Step S4.1: Obtain individual air pollutant exposure level monitoring data and outdoor air pollutant concentration at the same time to obtain the first ratio of individual air pollutant exposure level monitoring data to outdoor air pollutant concentration; Step S4.2: Obtain the individual exposure level data output by the air pollutant population exposure level prediction model after accuracy assessment and the outdoor air pollutant concentration at the same time to obtain the second ratio of individual exposure level data to outdoor air pollutant concentration; Step S4.3: Verify the accuracy of the air pollutant population exposure level prediction model in predicting individual exposure levels based on the first ratio and the second ratio.

9. A system for predicting an individual's exposure level to air pollutants, characterized in that, include: The data acquisition and preprocessing module collects raw sample data from the subject population and organizes the raw sample data into a raw dataset. Based on a random principle, the raw dataset is divided into a training subset and a test subset. Both the training subset and the test subset include input feature variables and individual air pollutant exposure level monitoring data. An initial model building module, which constructs an initial model of the population exposure level to air pollutants based on the training subset using a nonlinear fitting algorithm; An initial model training module trains the initial model using the training subset to obtain an air pollutant exposure level prediction model, and optimizes the air pollutant exposure level prediction model based on a combination of Lasso regression regularization, SHAP value weighting, and recursive feature elimination algorithm. The testing module inputs the input feature variables from the test subset into the optimized air pollutant exposure level prediction model, predicts the individual air pollutant exposure level of the subject population, obtains model prediction data, and compares the individual air pollutant exposure level monitoring data from the test subset with the model prediction data to evaluate the accuracy of the prediction model. The prediction module collects input characteristic variables from a target population that is different from the subject population, and substitutes the input characteristic variables of the target population into a prediction model that has been evaluated for accuracy, so as to predict the individual exposure level of air pollutants for the target population.

Images