Method for health management by detecting perfluoro- and polyfluoroalkyl substances in blood

Abstract

This invention relates to a method for health management based on the detection of perfluorinated and polyfluoroalkyl (PFAL) compounds in blood, belonging to the field of health monitoring technology. The method comprises the following steps: Step 1: Collecting blood samples from the target individual to form a multidimensional time-series raw dataset and generating standardized time-series input samples; Step 2: Constructing a two-dimensional time-series input tensor; Step 3: Constructing an improved sliding window transformer neural network model to output multi-level dynamic health feature representations; Step 4: Automatically adjusting and determining the optimal hyperparameter configuration on the training and validation sets to complete the optimization training of the improved sliding window transformer neural network model; Step 5: Using the results of steps 3 and 4, outputting the perfluorinated and polyfluoroalkyl (PFAL) compound pollution risk score and corresponding health warning level for the target individual; Step 6: Implementing health management based on the health warning level. This invention improves the detection capability of perfluorinated and polyfluoroalkyl (PFAL) compound pollution in blood, providing a basis for health management.

Description

Technical Field

[0001] This invention belongs to the field of health monitoring technology, specifically relating to a method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood. Background Technology

[0002] With the acceleration of industrialization and urbanization, the problem of perfluorinated and polyfluoroalkyl (PFAA) pollution in the environment is becoming increasingly prominent, and the accumulation of PFAA elements in human blood poses a serious threat to health. Currently, the detection of PFAA pollution in blood and the monitoring of health risks mainly rely on periodic manual sampling and traditional laboratory analysis methods. While these methods can quantitatively detect the concentration of PFAA elements in blood, traditional data analysis often uses static statistical methods or simple linear models, making it difficult to capture the complex dynamic characteristics of blood indicators changing over time, and also failing to fully utilize multidimensional time-series information and the nonlinear relationships between health-related biomarkers.

[0003] In recent years, some studies have attempted to introduce deep learning models and temporal neural networks to improve the detection capabilities of abnormal blood indicators. However, most of these studies employ fixed structures, and the selection of model structure parameters and hyperparameters lacks a scientific adaptive tuning mechanism, resulting in limited model generalization ability and prediction accuracy. Traditional methods such as grid search or random search are inefficient for hyperparameter search and model structure optimization, failing to quickly obtain the optimal parameter combination in high-dimensional complex spaces, and thus struggling to meet the real-time and accuracy requirements of actual health monitoring and management. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the prior art by providing a Bayesian-optimized method for detecting perfluorinated and polyfluoroalkyl compounds (PFACs) contamination in blood and for conducting health management based on the detection results. This method improves the detection capability of PFACs in blood, enables early identification and graded warning of PFAC contamination risks, and provides a scientific basis for personalized health management.

[0005] The technical solution is as follows: The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood adopts the following steps: Step 1: Collect blood samples from the tested individuals at different time points, detect the concentration values ​​of various perfluorinated and polyfluoroalkyl compounds and health-related biomarkers in the blood samples, record data timestamps, form a multidimensional time-series raw dataset, preprocess the multidimensional time-series raw dataset, and generate standardized time-series input samples. Step 2: Divide the standardized temporal input samples into several consecutive sliding window subsequences according to the preset window length and sliding step size, and construct a two-dimensional temporal input tensor; Step 3: Construct an improved sliding window transformer neural network model. Input the two-dimensional temporal input tensor into the improved sliding window transformer neural network model and output a multi-level health dynamic feature representation. Step 4: Use the Bayesian optimization algorithm to automatically adjust and determine the optimal hyperparameter configuration on the training and validation sets to complete the optimization training of the improved sliding window transformer neural network model; Step 5: Based on the optimized sliding window transformer neural network model and multi-level health dynamic feature representation, input the multi-dimensional time series data of the newly collected blood sample, and output the perfluorinated and polyfluoroalkyl compound pollution risk score and the corresponding health warning level of the target individual; Step 6: Conduct health management based on the health warning level.

[0006] Furthermore, the multidimensional time-series raw dataset in step 1 specifically includes multivariate data with timestamps of various perfluorinated and polyfluoroalkyl compounds and health-related biomarkers in blood samples collected from the tested individuals at different time points, including blood glucose, blood lipids, inflammatory factors, liver function indicators, and electrolyte levels.

[0007] Furthermore, the preprocessing of the multidimensional time series raw dataset in step 1 specifically includes normalization, missing value completion, outlier removal, and time alignment, and the multidimensional time series raw dataset is divided into training set and validation set in a 5:5 ratio.

[0008] Furthermore, step 2 involves constructing a two-dimensional temporal input tensor, including the following steps: Step 21: Set the sliding window length and sliding step size, where the sliding window length is the number of consecutive time points contained in each subsequence, and the sliding step size is the interval between the starting time points of adjacent sliding windows; Step 22: Divide the standardized time-series input samples into several sliding window subsequences according to the time order. Each sliding window subsequence contains a preset number of time points and all feature dimensions. Step 23: Number all the obtained sliding window subsequences in chronological order to form a complete set of sliding window subsequences; Step 24: Stack the complete set of sliding window subsequences sequentially to construct a two-dimensional temporal input tensor with the first dimension being the number of sliding window subsequences, the second dimension being the length of the sliding window, and the third dimension being the feature dimension; Step 25: Adjust and standardize the format of the two-dimensional time-series input tensor to make it conform to the input requirements of the sliding window transformer neural network model.

[0009] Furthermore, the improved sliding window transformer neural network model in step 3 includes a window adaptation module, a multi-scale fusion module, a prior attention module, an attention transfer module, and a feature fusion network. The window adaptation module introduces a window adaptation sensitivity threshold, the multi-scale fusion module introduces the number of multi-scale branches, and the prior attention module introduces a prior weight influence coefficient.

[0010] Furthermore, in step 3, the two-dimensional temporal input tensor is fed into the improved sliding window transformer neural network model to output a multi-level health dynamic feature representation, including the following steps: Step 31: Input the two-dimensional time series input tensor into the window adaptation module. The window adaptation module consists of an anomaly evaluation unit and a length adjustment unit. The anomaly evaluation unit performs anomaly sensitivity analysis on the two-dimensional time series input tensor. The length adjustment unit dynamically adjusts the sliding window length according to the window adaptation sensitivity threshold parameter and outputs the adaptive window subsequence feature representation. Step 32: Input the adaptive window subsequence feature representation into the multi-scale fusion module. The multi-scale fusion module consists of multiple parallel window branches. Each branch contains an independent window partitioning unit and a feature extraction unit. The feature extraction unit consists of a self-attention layer and a multi-layer perceptron layer. The number of window branches is controlled by the multi-scale branch number parameter. The features output by each branch are concatenated and weighted through the feature fusion unit to output the multi-scale fusion feature representation. Step 33: Input the multi-scale fusion feature representation into the prior attention module. The prior attention module consists of a knowledge embedding unit and an attention guidance unit. The knowledge embedding unit generates prior information using a trainable weight mask and a medical indicator scoring matrix. The attention guidance unit optimizes the attention weight distribution based on the prior weight influence coefficient parameter and outputs the optimized prior feature representation. Step 34: Input the prior optimized feature representation into the attention transfer module. The attention transfer module consists of a triggering unit and a dynamic adjustment unit. The triggering unit detects abnormalities in key medical indicators, and the dynamic adjustment unit dynamically adjusts the attention distribution between windows to achieve continuous tracking of abnormal features and outputs the attention transfer feature representation. Step 35: Input the interest transfer feature representation into the feature fusion network. The feature fusion network consists of multiple stacked fully connected layers, activation function layers and normalization units. It performs nonlinear transformation and feature integration on the interest transfer feature representation and outputs a multi-level health dynamic feature representation.

[0011] Furthermore, in step 4, the hyperparameters are the sliding window length, window offset parameter, window adaptive sensitivity threshold, number of multi-scale branches, and prior weight influence coefficient.

[0012] Furthermore, the optimized training of the improved sliding window transformer neural network model in step 4 includes the following steps: Step 41: Set the sliding window length, window offset parameter, window adaptive sensitivity threshold, number of multi-scale branches and prior weight influence coefficient as hyperparameters to be optimized, and set a reasonable value range for each hyperparameter to form a complete hyperparameter search space; Step 42: In the hyperparameter search space, randomly select several sets of hyperparameter combinations, train the sliding window transformer neural network model on the training set and validation set respectively, record the performance evaluation index corresponding to each set of hyperparameter combinations, and obtain the initial sampling dataset; Step 43: Based on the initial sampled dataset, establish a global Gaussian process surrogate model, and simultaneously cluster the sampled points according to their distribution characteristics. In each cluster subspace, establish a local Gaussian process surrogate model to simultaneously model the global and local objective function distributions. Step 44: For each round of new hyperparameter sampling, calculate the values ​​of the three acquisition functions—expected improvement, confidence upper limit, and probability improvement—on all candidate hyperparameter combinations, and standardize the values ​​of the three acquisition functions respectively. Obtain the joint acquisition score through weighted fusion. The hyperparameter combination with the highest joint acquisition score is used as the sampling point for this round. Step 45: For the hyperparameter combination determined in this round, train the sliding window transformer neural network model on the training set and validation set and evaluate its performance. Add new sampled data to the sampled dataset and update the global and local Gaussian process surrogate model parameters in real time. Step 46: During the training of the surrogate model, regularization terms and Bayesian priors are automatically introduced into the parameters of the global and local Gaussian process surrogate models. The regularization weights are dynamically adjusted according to the sampling density of their respective subspaces, so that the global and local Gaussian process surrogate models maintain high uncertainty in sparse regions and improve fitting ability in dense regions. Step 47: Continuous sampling and updating. When the mean square variance or information entropy of the overall hyperparameter space is lower than the set threshold, or the performance evaluation index converges, the optimization process is terminated, and the best-performing hyperparameter combination among all historical samples is selected as the final optimal hyperparameter configuration for the sliding window transformer neural network model.

[0013] Furthermore, step 6 of the health management process includes the following steps: Step 61: Based on the obtained health warning level, combined with historical health records and individual characteristics, match the preset health intervention rules to generate individualized health intervention recommendations; Step 62: Integrate the risk score, health warning level, and individualized health intervention recommendations to generate a test report. The report includes the concentrations of multiple perfluorinated and polyfluoroalkyl compounds in the blood, major health-related biomarkers, risk assessment results, and corresponding recommendations. Step 63: Based on the risk level in the test report, dynamically adjust the blood sample collection cycle and health monitoring frequency. For high-risk levels, shorten the collection cycle and increase the monitoring frequency; for low-risk levels, extend the collection cycle and reduce the monitoring frequency. Step 64: Simultaneously push the test report and sampling and monitoring adjustment suggestions to the individual health record database, and connect with the medical information system to achieve a closed loop of continuous health management and automatic early warning.

[0014] Beneficial effects: 1) This invention deeply integrates the sliding window transformer neural network model with the Bayesian optimization algorithm. By using Bayesian optimization, it achieves efficient adaptive tuning of model structure parameters and hyperparameters, improves the model's generalization ability and prediction accuracy, effectively solves the problems of arbitrary parameter selection and low model optimization efficiency in traditional methods, and enhances the detection capability of perfluorinated and polyfluoroalkyl compounds in blood.

[0015] 2) By introducing window adaptation, multi-scale feature fusion, and medical prior knowledge-guided attention mechanisms, the model can accurately identify abnormal changes in blood indicators and key risk signals to achieve early identification and graded warning of the risk of perfluorinated and polyfluoroalkyl compounds pollution, providing a scientific basis for personalized health management. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood according to the present invention; Figure 2 This is a schematic diagram of the neural network model of the sliding window transformer. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit it. Terms such as "upper," "lower," "front," "rear," "left," "right," "bottom," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the purpose of describing the invention. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.

[0018] like Figures 1-2The method shown for detecting perfluorinated and polyfluoroalkyl compounds in blood for health management employs the following steps: Step 1: Collect blood samples from the tested individuals at different time points, detect the concentrations of various perfluorinated and polyfluoroalkyl compounds and health-related biomarkers in the blood samples, record the data timestamps, and form a multidimensional time-series raw dataset. The multidimensional time-series raw dataset specifically includes time-stamped multivariate data on the concentrations of various perfluorinated and polyfluoroalkyl compounds and health-related biomarkers in blood samples collected from the tested individuals at different time points, including blood glucose, blood lipids, inflammatory factors, liver function indicators, and electrolyte levels. Preprocess the multidimensional time-series raw dataset, specifically including normalization, missing value completion, outlier removal, and time alignment, and divide the multidimensional time-series raw dataset into a 5:5 ratio for training and validation sets; generate standardized time-series input samples. Step 2: Divide the standardized temporal input samples into several consecutive sliding window subsequences according to the preset window length and sliding step size, and construct a two-dimensional temporal input tensor; Step 3: Construct an improved sliding window transformer neural network model. Input the two-dimensional temporal input tensor into the improved sliding window transformer neural network model and output a multi-level health dynamic feature representation. Step 4: Use the Bayesian optimization algorithm to automatically adjust and determine the optimal hyperparameter configuration on the training and validation sets to complete the optimization training of the improved sliding window transformer neural network model; Step 5: Based on the optimized sliding window transformer neural network model and multi-level health dynamic feature representation, input the multi-dimensional time series data of the newly collected blood sample, and output the perfluorinated and polyfluoroalkyl compound pollution risk score and the corresponding health warning level of the target individual; Step 6: Conduct health management based on the health warning level.

[0019] Step 2 involves constructing a two-dimensional temporal input tensor, including the following steps: Step 21: Set the sliding window length and sliding step size, where the sliding window length is the number of consecutive time points contained in each subsequence, and the sliding step size is the interval between the starting time points of adjacent sliding windows; Step 22: Divide the standardized time-series input samples into several sliding window subsequences according to the time order. Each sliding window subsequence contains a preset number of time points and all feature dimensions. Step 23: Number all the obtained sliding window subsequences in chronological order to form a complete set of sliding window subsequences; Step 24: Stack the complete set of sliding window subsequences sequentially to construct a two-dimensional temporal input tensor with the first dimension being the number of sliding window subsequences, the second dimension being the length of the sliding window, and the third dimension being the feature dimension; Step 25: Adjust and standardize the format of the two-dimensional time-series input tensor to make it conform to the input requirements of the sliding window transformer neural network model.

[0020] The improved sliding window transformer neural network model in step 3 includes a window adaptation module, a multi-scale fusion module, a prior attention module, an attention transfer module, and a feature fusion network. The window adaptation module introduces a window adaptation sensitivity threshold, the multi-scale fusion module introduces the number of multi-scale branches, and the prior attention module introduces prior weight influence coefficients. The two-dimensional temporal input tensor is fed into the improved sliding window transformer neural network model, outputting a multi-level health dynamic feature representation, including the following steps: Step 31: Input the two-dimensional time series input tensor into the window adaptation module. The window adaptation module consists of an anomaly evaluation unit and a length adjustment unit. The anomaly evaluation unit performs anomaly sensitivity analysis on the two-dimensional time series input tensor. The length adjustment unit dynamically adjusts the sliding window length according to the window adaptation sensitivity threshold parameter and outputs the adaptive window subsequence feature representation. Step 32: Input the adaptive window subsequence feature representation into the multi-scale fusion module. The multi-scale fusion module consists of multiple parallel window branches. Each branch contains an independent window partitioning unit and a feature extraction unit. The feature extraction unit consists of a self-attention layer and a multi-layer perceptron layer. The number of window branches is controlled by the multi-scale branch number parameter. The features output by each branch are concatenated and weighted through the feature fusion unit to output the multi-scale fusion feature representation. Step 33: Input the multi-scale fusion feature representation into the prior attention module. The prior attention module consists of a knowledge embedding unit and an attention guidance unit. The knowledge embedding unit generates prior information using a trainable weight mask and a medical indicator scoring matrix. The attention guidance unit optimizes the attention weight distribution based on the prior weight influence coefficient parameter and outputs the optimized prior feature representation. Step 34: Input the prior optimized feature representation into the attention transfer module. The attention transfer module consists of a triggering unit and a dynamic adjustment unit. The triggering unit detects abnormalities in key medical indicators, and the dynamic adjustment unit dynamically adjusts the attention distribution between windows to achieve continuous tracking of abnormal features and outputs the attention transfer feature representation. Step 35: Input the interest transfer feature representation into the feature fusion network. The feature fusion network consists of multiple stacked fully connected layers, activation function layers and normalization units. It performs nonlinear transformation and feature integration on the interest transfer feature representation and outputs a multi-level health dynamic feature representation.

[0021] In step 4, the hyperparameters are the sliding window length, window offset parameter, window adaptive sensitivity threshold, number of multi-scale branches, and prior weight influence coefficient. The optimization training of the improved sliding window transformer neural network model includes the following steps: Step 41: Set the sliding window length, window offset parameter, window adaptive sensitivity threshold, number of multi-scale branches and prior weight influence coefficient as hyperparameters to be optimized, and set a reasonable value range for each hyperparameter to form a complete hyperparameter search space; Step 42: In the hyperparameter search space, randomly select several sets of hyperparameter combinations, train the sliding window transformer neural network model on the training set and validation set respectively, record the performance evaluation index corresponding to each set of hyperparameter combinations, and obtain the initial sampling dataset; Step 43: Based on the initial sampled dataset, establish a global Gaussian process surrogate model, and simultaneously cluster the sampled points according to their distribution characteristics. In each cluster subspace, establish a local Gaussian process surrogate model to simultaneously model the global and local objective function distributions. Step 44: For each round of new hyperparameter sampling, calculate the values ​​of the three acquisition functions—expected improvement, confidence upper limit, and probability improvement—on all candidate hyperparameter combinations, and standardize the values ​​of the three acquisition functions respectively. Obtain the joint acquisition score through weighted fusion. The hyperparameter combination with the highest joint acquisition score is used as the sampling point for this round. Step 45: For the hyperparameter combination determined in this round, train the sliding window transformer neural network model on the training set and validation set and evaluate its performance. Add new sampled data to the sampled dataset and update the global and local Gaussian process surrogate model parameters in real time. Step 46: During the training of the surrogate model, regularization terms and Bayesian priors are automatically introduced into the parameters of the global and local Gaussian process surrogate models. The regularization weights are dynamically adjusted according to the sampling density of their respective subspaces, so that the global and local Gaussian process surrogate models maintain high uncertainty in sparse regions and improve fitting ability in dense regions. Step 47: Continuous sampling and updating. When the mean square variance or information entropy of the overall hyperparameter space is lower than the set threshold, or the performance evaluation index converges, the optimization process is terminated, and the best-performing hyperparameter combination among all historical samples is selected as the final optimal hyperparameter configuration for the sliding window transformer neural network model.

[0022] Step 6, health management, includes the following steps: Step 61: Based on the obtained health warning level, combined with historical health records and individual characteristics, match the preset health intervention rules to generate individualized health intervention recommendations; Step 62: Integrate the risk score, health warning level, and individualized health intervention recommendations to generate a test report. The report includes the concentrations of multiple perfluorinated and polyfluoroalkyl compounds in the blood, major health-related biomarkers, risk assessment results, and corresponding recommendations. Step 63: Based on the risk level in the test report, dynamically adjust the blood sample collection cycle and health monitoring frequency. For high-risk levels, shorten the collection cycle and increase the monitoring frequency; for low-risk levels, extend the collection cycle and reduce the monitoring frequency. Step 64: Simultaneously push the test report and sampling and monitoring adjustment suggestions to the individual health record database, and connect with the medical information system to achieve a closed loop of continuous health management and automatic early warning.

[0023] Example: A dynamic monitoring project on the health risks of perfluorinated and polyfluoroalkyl substances (PFAS) in residents' blood was conducted by a municipal disease control and prevention center in conjunction with community health service centers. In this project, the team followed up with 900 permanent residents aged 18-60 in the area for one year, collecting blood samples quarterly to test the concentrations of various PFAS, including perfluorooctanoic acid (PFOA), perfluorooctane sulfonate (PFOS), perfluorohexane sulfonate (PFHxS), and perfluorononane sulfonate (PFNA), while simultaneously testing serum health-related biomarkers. In addition, the physical activity level of each participant was recorded in detail during the collection process, including average daily steps, exercise duration, and exercise intensity (METs), obtained through both wearable health devices and activity questionnaires. The data was annotated with a unified timestamp and uploaded to a cloud platform.

[0024] All raw blood and physical activity data were normalized, outlier removed, missing value imputed, and time-aligned to form standardized multidimensional time-series input samples. The system constructs a two-dimensional time-series input tensor based on quarterly units, combined with physical activity level data, and using optimized window lengths and step sizes. A sliding window transformer neural network model is used to fully extract the dynamic interaction features between blood PFAS indicators, various biomarker parameters, and physical activity levels. The model structure and key hyperparameters were all automatically selected through Bayesian optimization, ensuring both high sensitivity and good generalization ability.

[0025] During the project, the platform dynamically outputs a blood PFAS contamination risk score and a brain health warning level for each resident, while automatically analyzing their recent physical activity trends. When the model detects a high PFAS risk warning in a given quarter, the platform immediately generates personalized intervention recommendations for that individual, including increasing the frequency of aerobic exercise, strengthening walking, and encouraging participation in community sports programs. All high-risk residents subsequently receive 12 weeks of exercise intervention tracking during the health management period, regularly uploading exercise data and blood test data.

[0026] Table 1. Comparison of the effects of blood PFAS risk monitoring and intervention based on physical activity intervention. Table 1 clearly shows that different monitoring and intervention methods exhibit significant differences in their effectiveness in identifying and assessing the health risks of per- and polyfluoroalkyl substances (PFAS) in blood. The traditional linear statistical model has an AUC of only 0.79, a recall of 0.65, and a specificity of 0.73, with an average early warning period of only 1.2 weeks and a high false alarm rate of 0.26. In terms of actual improvement in high-risk populations, the risk level improvement rate after intervention was only 36%, the PFAS reduction rate was 4.1%, and the interpretation time per case was 18 seconds. This indicates that this method has limited support for early detection and scientific intervention in high-risk individuals, and its intelligence and accuracy are insufficient.

[0027] LSTM temporal neural networks have shown significant improvement as a deep learning method, achieving an AUC of 0.85, recall and specificity increased to 0.76 and 0.80 respectively, the average early warning period was extended to 2.0 weeks, and the false alarm rate decreased to 0.18. The improvement rate and post-intervention PFAS reduction rate for high-risk groups were 53% and 8.3% respectively, although the interpretation efficiency decreased slightly, requiring 25 seconds per case. While the LSTM model improves the ability to identify risk signals and enhances individual intervention effectiveness, it still has shortcomings in intelligent closed-loop processing and the fusion analysis of complex multidimensional dynamic factors in practical applications.

[0028] The method of this invention exhibits superior performance across all key metrics. It achieves an AUC of 0.93, recall of 0.88, and specificity of 0.86, significantly extending the average early warning period to 4.3 weeks, with a false alarm rate of only 0.08. After intervention, the risk level improvement rate for high-risk residents reached 73%, and the PFAS concentration reduction rate reached 17.2%, while the interpretation time per case was only 8 seconds. Compared to the previous two methods, this invention not only significantly improves early detection and accurate warning capabilities but also fully integrates intervention methods such as physical activity to achieve dynamic risk assessment and individualized improvement.

[0029] Overall, this invention provides a more efficient, scientific, and practical solution for the intelligent monitoring and intervention quantification of health risks associated with blood PFAS contamination.

[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the principles and spirit of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood, characterized in that... The following steps are used: Step 1: Collect blood samples from the tested individuals at different time points, detect the concentration values ​​of various perfluorinated and polyfluoroalkyl compounds and health-related biomarkers in the blood samples, record data timestamps, form a multidimensional time-series raw dataset, preprocess the multidimensional time-series raw dataset, and generate standardized time-series input samples. Step 2: Divide the standardized temporal input samples into several consecutive sliding window subsequences according to the preset window length and sliding step size, and construct a two-dimensional temporal input tensor; Step 3: Construct an improved sliding window transformer neural network model. Input the two-dimensional temporal input tensor into the improved sliding window transformer neural network model and output a multi-level health dynamic feature representation. Step 4: Use the Bayesian optimization algorithm to automatically adjust and determine the optimal hyperparameter configuration on the training and validation sets to complete the optimization training of the improved sliding window transformer neural network model; Step 5: Based on the optimized sliding window transformer neural network model and multi-level health dynamic feature representation, input the multi-dimensional time series data of the newly collected blood sample, and output the perfluorinated and polyfluoroalkyl compound pollution risk score and the corresponding health warning level of the target individual; Step 6: Conduct health management based on the health warning level.

2. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 1, characterized in that: The multidimensional time-series raw dataset mentioned in step 1 specifically includes multivariate data with timestamps of various perfluorinated and polyfluoroalkyl compounds and health-related biomarkers in blood samples collected from the tested individuals at different time points, including blood glucose, blood lipids, inflammatory factors, liver function indicators, and electrolyte levels.

3. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 1, characterized in that: The preprocessing of the multidimensional time series raw dataset described in step 1 specifically includes normalization, missing value completion, outlier removal, and time alignment, and the multidimensional time series raw dataset is divided into training set and validation set in a 5:5 ratio.

4. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 1, characterized in that: The construction of the two-dimensional temporal input tensor in step 2 includes the following steps: Step 21: Set the sliding window length and sliding step size, where the sliding window length is the number of consecutive time points contained in each subsequence, and the sliding step size is the interval between the starting time points of adjacent sliding windows; Step 22: Divide the standardized time-series input samples into several sliding window subsequences according to the time order. Each sliding window subsequence contains a preset number of time points and all feature dimensions. Step 23: Number all the obtained sliding window subsequences in chronological order to form a complete set of sliding window subsequences; Step 24: Stack the complete set of sliding window subsequences sequentially to construct a two-dimensional temporal input tensor with the first dimension being the number of sliding window subsequences, the second dimension being the length of the sliding window, and the third dimension being the feature dimension; Step 25: Adjust and standardize the format of the two-dimensional time-series input tensor to make it conform to the input requirements of the sliding window transformer neural network model.

5. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 1, characterized in that: The improved sliding window transformer neural network model described in step 3 includes a window adaptation module, a multi-scale fusion module, a prior attention module, an attention transfer module, and a feature fusion network. The window adaptation module introduces a window adaptation sensitivity threshold, the multi-scale fusion module introduces the number of multi-scale branches, and the prior attention module introduces a prior weight influence coefficient.

6. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 5, characterized in that: Step 3, which involves inputting the two-dimensional temporal input tensor into the improved sliding window transformer neural network model and outputting a multi-level health dynamic feature representation, includes the following steps: Step 31: Input the two-dimensional time series input tensor into the window adaptation module. The window adaptation module consists of an anomaly evaluation unit and a length adjustment unit. The anomaly evaluation unit performs anomaly sensitivity analysis on the two-dimensional time series input tensor. The length adjustment unit dynamically adjusts the sliding window length according to the window adaptation sensitivity threshold parameter and outputs the adaptive window subsequence feature representation. Step 32: Input the adaptive window subsequence feature representation into the multi-scale fusion module. The multi-scale fusion module consists of multiple parallel window branches. Each branch contains an independent window partitioning unit and a feature extraction unit. The feature extraction unit consists of a self-attention layer and a multi-layer perceptron layer. The number of window branches is controlled by the multi-scale branch number parameter. The features output by each branch are concatenated and weighted through the feature fusion unit to output the multi-scale fusion feature representation. Step 33: Input the multi-scale fusion feature representation into the prior attention module. The prior attention module consists of a knowledge embedding unit and an attention guidance unit. The knowledge embedding unit generates prior information using a trainable weight mask and a medical indicator scoring matrix. The attention guidance unit optimizes the attention weight distribution based on the prior weight influence coefficient parameter and outputs the optimized prior feature representation. Step 34: Input the prior optimized feature representation into the attention transfer module. The attention transfer module consists of a triggering unit and a dynamic adjustment unit. The triggering unit detects abnormalities in key medical indicators, and the dynamic adjustment unit dynamically adjusts the attention distribution between windows to achieve continuous tracking of abnormal features and outputs the attention transfer feature representation. Step 35: Input the interest transfer feature representation into the feature fusion network. The feature fusion network consists of multiple stacked fully connected layers, activation function layers and normalization units. It performs nonlinear transformation and feature integration on the interest transfer feature representation and outputs a multi-level health dynamic feature representation.

7. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 1, characterized in that: The hyperparameters mentioned in step 4 are the sliding window length, window offset parameter, window adaptive sensitivity threshold, number of multi-scale branches, and prior weight influence coefficient.

8. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 7, characterized in that: The optimization training of the improved sliding window transformer neural network model described in step 4 includes the following steps: Step 41: Set the sliding window length, window offset parameter, window adaptive sensitivity threshold, number of multi-scale branches and prior weight influence coefficient as hyperparameters to be optimized, and set a reasonable value range for each hyperparameter to form a complete hyperparameter search space; Step 42: In the hyperparameter search space, randomly select several sets of hyperparameter combinations, train the sliding window transformer neural network model on the training set and validation set respectively, record the performance evaluation index corresponding to each set of hyperparameter combinations, and obtain the initial sampling dataset; Step 43: Based on the initial sampled dataset, establish a global Gaussian process surrogate model, and simultaneously cluster the sampled points according to their distribution characteristics. In each cluster subspace, establish a local Gaussian process surrogate model to simultaneously model the global and local objective function distributions. Step 44: For each round of new hyperparameter sampling, calculate the values ​​of the three acquisition functions—expected improvement, confidence upper limit, and probability improvement—on all candidate hyperparameter combinations, and standardize the values ​​of the three acquisition functions respectively. Obtain the joint acquisition score through weighted fusion. The hyperparameter combination with the highest joint acquisition score is used as the sampling point for this round. Step 45: For the hyperparameter combination determined in this round, train the sliding window transformer neural network model on the training set and validation set and evaluate its performance. Add new sampled data to the sampled dataset and update the global and local Gaussian process surrogate model parameters in real time. Step 46: During the training of the surrogate model, regularization terms and Bayesian priors are automatically introduced into the parameters of the global and local Gaussian process surrogate models. The regularization weights are dynamically adjusted according to the sampling density of their respective subspaces, so that the global and local Gaussian process surrogate models maintain high uncertainty in sparse regions and improve fitting ability in dense regions. Step 47: Continuous sampling and updating. When the mean square variance or information entropy of the overall hyperparameter space is lower than the set threshold, or the performance evaluation index converges, the optimization process is terminated, and the best-performing hyperparameter combination among all historical samples is selected as the final optimal hyperparameter configuration for the sliding window transformer neural network model.

9. The method for health management by detecting perfluorinated and polyfluoroalkyl compounds in blood as described in claim 1, characterized in that: The health management described in step 6 includes the following steps: Step 61: Based on the obtained health warning level, combined with historical health records and individual characteristics, match the preset health intervention rules to generate individualized health intervention recommendations; Step 62: Integrate the risk score, health warning level, and individualized health intervention recommendations to generate a test report. The report includes the concentrations of multiple perfluorinated and polyfluoroalkyl compounds in the blood, major health-related biomarkers, risk assessment results, and corresponding recommendations. Step 63: Based on the risk level in the test report, dynamically adjust the blood sample collection cycle and health monitoring frequency. For high-risk levels, shorten the collection cycle and increase the monitoring frequency; for low-risk levels, extend the collection cycle and reduce the monitoring frequency. Step 64: Simultaneously push the test report and sampling and monitoring adjustment suggestions to the individual health record database, and connect with the medical information system to achieve a closed loop of continuous health management and automatic early warning.

Images