A human health assessment system and method based on perfluorocompound detection

Abstract

The present application relates to the technical field of electric digital data processing, and particularly relates to a human health evaluation system and method based on perfluorinated compound detection, which comprises a data acquisition unit, a preprocessing unit, a feature fusion unit, a prediction unit and an output unit. Enhanced original data containing multi-dimensional meta-information such as equipment operation state and laboratory environment are collected, noise filtering, missing value completion and dimension unification preprocessing are performed based on the meta-information and data quality mapping knowledge base, the feature distribution of the source domain and the target domain is aligned by using a domain adaptive model with weights, the risk ratio, the confidence interval and the feature reliability score are output in combination with a multi-task learning framework, and finally the health evaluation report is generated through dynamic threshold adjustment and the feature weight is fed back for optimization. Through full-link collaborative optimization, the feature fusion efficiency and the model risk quantification accuracy are improved, and robust support is provided for perfluorinated compound exposure health risk assessment.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a human health assessment system and method based on the detection of perfluorinated compounds. Background Technology

[0002] Human health assessment based on the detection of perfluorinated compounds (PFCs) is a multidisciplinary assessment system that integrates analytical chemistry detection technology with electronic digital data processing. Perfluorinated compounds (PFCs), a class of persistent organic pollutants with long carbon-fluorine bonds, are widely present in environmental media and organisms due to their hydrophobic and oleophobic properties. After accumulating in the human body through the food chain or direct contact, they can accumulate in tissues such as serum and liver, interfering with physiological functions such as the endocrine and immune systems. The assessment process first requires the use of high-sensitivity analytical techniques such as high-performance liquid chromatography and tandem mass spectrometry (HPLC-MS / MS) to qualitatively and quantitatively detect typical PFCs such as perfluorooctane sulfonate (PFOS) and perfluorooctanoic acid (PFOA) in biological samples such as blood and urine, obtaining concentration distribution data of different homologues. Then, electronic digital data processing (EDC) techniques are used to standardize and preprocess the detection data. Combined with an exposure and effect database established from epidemiological cohort studies, statistical learning models such as multiple linear regression and random forests are used to analyze the dose-response relationship between PFC concentrations and specific health endpoints (such as neurobehavioral developmental abnormalities in children and the incidence of metabolic syndrome in adults). Based on statistical measures such as hazard ratio (HR) and confidence interval (CI), the health risk level at different exposure levels is quantified, providing data support for early warning of environmental exposure health risks for individuals or groups.

[0003] The pain points of human health assessment systems and methods based on perfluorinated compound detection in the field of electronic digital data processing are as follows: insufficient feature fusion efficiency of biological sample detection data limits the accuracy of statistical learning models in quantifying health risks. Perfluorinated compound detection data comes from biological samples such as blood and urine generated at different time points and using different detection devices (such as HPLC-MS / MS), and their data formats, noise levels, and feature dimensions exhibit significant heterogeneity. Existing standardized preprocessing methods can only eliminate differences in data dimensions and cannot effectively fuse heterogeneous features, leading to misjudgments of feature importance during the training of statistical learning models such as multiple linear regression and random forests. For example, when integrating PFOS concentration data from Laboratory A (detection limit 0.1 ng / mL) and Laboratory B (detection limit 0.5 ng / mL), Laboratory A outputs a specific value for a 0.2 ng / mL sample, while Laboratory B marks it as "not detected" (missing value). If the mean imputation method is used to handle missing values, the true exposure level of the sample will be underestimated. This will lead to an increase in the error in calculating the hazard ratio (HR) for low-exposure groups when the model predicts the risk of neurobehavioral developmental abnormalities in children, directly affecting the reliability of the quantitative assessment of health risk levels. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a human health assessment system and method based on the detection of perfluorinated compounds, solving the problem that insufficient feature fusion efficiency of biological sample detection data limits the accuracy of statistical learning models in quantifying health risks related to perfluorinated compound exposure.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows: In a first aspect, the human health assessment system based on the detection of perfluorinated compounds provided by the present invention includes a data acquisition unit, a data preprocessing unit, a feature fusion unit, a prediction unit, and an output unit. The data acquisition unit collects biological sample concentration values, missing value markers, and multidimensional information such as equipment operating status and laboratory environment generated by the perfluorinated compound detection device, constructs an enhanced raw dataset, and transmits it to the data preprocessing unit. The data preprocessing unit performs noise filtering, missing value completion, and unit unification processing on the enhanced original dataset based on the metadata and data quality mapping knowledge base, and outputs a structured preprocessed dataset labeled with data quality scores to the feature fusion unit. The feature fusion unit calculates feature quality weights based on the data quality score, aligns the feature distributions of the source and target domains using a weighted domain adaptive model, generates a fused feature set and a feature quality comprehensive index, and transmits them to the prediction unit. The prediction unit uses a multi-task learning framework to output risk ratio and confidence interval based on the fused feature set, and outputs feature reliability score based on feature quality comprehensive index. The output unit receives the risk ratio and confidence interval, combines them with the dynamically adjusted threshold interval to generate a health assessment report, and feeds back the feature reliability score to the feature fusion unit to optimize the weight allocation strategy.

[0006] Furthermore, in the human health assessment system based on perfluorinated compound detection described in this invention, the data acquisition unit acquires biological sample data generated by the perfluorinated compound detection device through a multi-protocol data interface, which supports laboratory information management system protocols and medical information exchange protocols. While acquiring biological sample concentration values ​​and missing value markers, the data acquisition unit simultaneously records device operating status parameters and laboratory environmental parameters, constructing an enhanced original dataset including concentration values, missing value markers, and multidimensional metadata.

[0007] Furthermore, in the human health assessment system based on perfluorinated compound detection described in this invention, the data preprocessing unit stores metadata and a data quality mapping knowledge base, which includes quantitative association rules between device voltage fluctuation values ​​and detection noise levels, and quantitative association rules between temperature and humidity values ​​and sample degradation rates. The data preprocessing unit, based on the quantization rules in the data quality mapping knowledge base, identifies high-noise data by comparing device voltage fluctuation values ​​with preset thresholds, identifies easily degradable samples by comparing temperature and humidity values ​​with standard ranges, and selects the corresponding preprocessing strategy. Median filtering is used for high-noise data, time decay interpolation is used to fill in missing values ​​for easily degradable samples, and quantile standardization is used to unify the units for data from different devices, outputting a structured preprocessed dataset with labeled data quality scores.

[0008] Furthermore, in the human health assessment system based on perfluorinated compound detection described in this invention, the feature fusion unit adopts a weighted domain adaptive model, with the exposure and effect database of epidemiological cohort studies as the source domain and the structured preprocessed dataset as the target domain. The feature fusion unit calculates the quality weights of the target domain features based on the data quality score. In the process of minimizing the maximum mean difference between the feature distributions of the source domain and the target domain, it prioritizes aligning the distribution of high-quality features with high data quality scores, and calculates the feature quality comprehensive index based on the quality weights, outputting the fused feature set and the feature quality comprehensive index.

[0009] Furthermore, in the human health assessment system based on perfluorinated compound detection described in this invention, the prediction unit adopts a multi-task learning framework, where the main task receives the fused feature set and outputs the risk ratio and confidence interval, and the auxiliary task receives the feature quality comprehensive index and outputs the feature reliability score. The prediction unit adjusts the quality weight allocation strategy of the feature fusion unit based on the feature reliability score, and at the same time uses unlabeled new detection data to self-train the model and dynamically update the model parameters to adapt to the feature distribution changes of new devices and new sample types.

[0010] Furthermore, in the human health assessment system based on perfluorinated compound detection described in this invention, the output unit is equipped with a risk threshold dynamic update engine. The risk threshold dynamic update engine analyzes the time series data of historical assessment results and the changing trends of epidemiological data through a sliding window algorithm based on historical assessment results and the latest epidemiological data, and adjusts the risk threshold range accordingly. The output unit maps the risk ratio and confidence interval to the adjusted threshold interval, and combines the feature quality comprehensive index to mark the reliability of the risk level, generating a health assessment report including the risk level, feature reliability score, and the basis for threshold adjustment.

[0011] Secondly, the human health assessment method based on the detection of perfluorinated compounds provided by the present invention, applied to the aforementioned human health assessment system based on the detection of perfluorinated compounds, includes: Collect multidimensional information such as biological sample concentration values, missing value markers, equipment operating status, and laboratory environment generated by perfluorinated compound detection equipment to construct an enhanced original dataset; Based on the metadata and data quality mapping knowledge base, noise filtering, missing value completion and unit unification processing are performed on the enhanced original dataset to output a structured preprocessed dataset with labeled data quality scores. Based on the data quality score, feature quality weights are calculated, and a weighted domain adaptive model is used to align the feature distributions of the source and target domains to generate a fused feature set and a comprehensive feature quality index. Using a multi-task learning framework, the risk ratio and confidence interval are output based on the fused feature set, and the feature reliability score is output based on the feature quality comprehensive index. The system receives the risk ratio and confidence interval, combines them with dynamically adjusted threshold intervals to generate a health assessment report, and feeds back the feature reliability score to the feature fusion unit to optimize the weight allocation strategy.

[0012] Beneficial effects of this invention; This invention improves the feature fusion efficiency and risk quantification accuracy of statistical learning models for biological sample testing data through multi-stage collaborative optimization: In the data acquisition stage, multi-dimensional metadata such as equipment operating status and laboratory environment are recorded simultaneously to construct an enhanced raw dataset including quality-related metadata, solving the information loss problem caused by traditional methods relying solely on concentration values; In the preprocessing stage, based on the association rules between metadata and data quality, targeted strategies such as median filtering for noise reduction, time-decay interpolation to fill missing values, and quantile standardization to unify dimensions are employed, and data quality scores are labeled, improving data reliability and consistency and avoiding the limitations of traditional preprocessing that only eliminates dimensional differences; Feature fusion… The fusion process employs a weighted domain-adaptive model that assigns feature weights based on data quality scores and prioritizes aligning with high-quality feature distributions, reducing the interference of low-quality data on the fusion results and achieving effective alignment of multi-source heterogeneous data. The prediction process uses a multi-task learning framework combined with semi-supervised learning, dynamically optimizing feature weight allocation through feature reliability scores from auxiliary tasks and self-training to adapt to new distributions using unlabeled new data. The output process combines dynamically adjusted risk thresholds and feature quality information to generate reports. This invention enables statistical learning models to more accurately capture the association between perfluorinated compound exposure and health endpoints, improving the robustness and practical application value of health risk quantification. Attached Figure Description

[0013] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0014] Figure 1 A flowchart of an environmental pollution assessment method for volatile environmentally friendly coatings provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of this invention, it will be described in further detail below.

[0016] Firstly, please refer to Figure 1The human health assessment system based on the detection of perfluorinated compounds provided by the present invention includes a data acquisition unit, a data preprocessing unit, a feature fusion unit, a prediction unit, and an output unit. The data acquisition unit collects biological sample concentration values, missing value markers, and multidimensional information such as equipment operating status and laboratory environment generated by the perfluorinated compound detection device, constructs an enhanced raw dataset, and transmits it to the data preprocessing unit. The data preprocessing unit performs noise filtering, missing value completion, and unit unification processing on the enhanced original dataset based on the metadata and data quality mapping knowledge base, and outputs a structured preprocessed dataset labeled with data quality scores to the feature fusion unit. The feature fusion unit calculates feature quality weights based on the data quality score, aligns the feature distributions of the source and target domains using a weighted domain adaptive model, generates a fused feature set and a feature quality comprehensive index, and transmits them to the prediction unit. The prediction unit uses a multi-task learning framework to output risk ratio and confidence interval based on the fused feature set, and outputs feature reliability score based on feature quality comprehensive index. The output unit receives the risk ratio and confidence interval, combines them with the dynamically adjusted threshold interval to generate a health assessment report, and feeds back the feature reliability score to the feature fusion unit to optimize the weight allocation strategy.

[0017] The human health assessment system based on the detection of perfluorinated compounds provided by this invention achieves multi-module collaboration through electro-digital data processing technology. Its technical solution can be decomposed into the following logically coherent steps: The system includes a data acquisition unit that communicates with perfluorinated compound detection equipment (such as HPLC-MS / MS) via a multi-protocol data interface. This interface supports Laboratory Information Management System (LIMS) and Medical Information Exchange (HL7) protocols, ensuring compatibility with the output formats of different detection devices. When acquiring the concentration values ​​of perfluorooctane sulfonate (PFOS), perfluorooctanoic acid (PFOA), and other homologues in biological samples (such as blood and urine), the unit simultaneously records two types of metadata: one is equipment operating status parameters (such as mass spectrometer ion source voltage fluctuations and detector sensitivity), and the other is laboratory environmental parameters (such as sample storage temperature and humidity, and atmospheric pressure). For example, when HPLC-MS / MS detects a PFOS concentration of 0.3 ng / mL in a blood sample, if the ion source voltage fluctuation exceeds 3% during the detection process, the data acquisition unit will store this fluctuation value along with the concentration value and missing values ​​such as "not detected" (e.g., samples below the detection limit are marked as "ND"), forming an enhanced raw dataset that includes the original detection data and multidimensional metadata. This dataset is then transmitted to the data preprocessing unit.

[0018] The data preprocessing unit incorporates a knowledge base mapping metadata and data quality. This knowledge base stores the association rules between device operating status, environmental parameters, and data quality. For example, the knowledge base records rules such as "when the ion source voltage fluctuation is >3%, the noise level of the detection value increases" and "when the sample storage temperature is >25℃, the PFOS degradation rate of urine samples increases by 10%". Based on these rules, the preprocessing unit first analyzes the metadata in the enhanced raw dataset: if the ion source voltage fluctuation corresponding to a sample is 4%, it is determined to be high-noise data; if the storage temperature of a urine sample is 28℃, it is determined to be an easily degradable sample. For high-noise data, the preprocessing unit uses a median filtering algorithm to reduce noise interference; for "undetected" missing values ​​of easily degradable samples, a time decay interpolation algorithm (dynamically adjusting the interpolation weight according to the sample collection time and degradation rate) is used to complete the missing values; for detection data generated by different devices (e.g., the detection limit of device A is 0.1 ng / mL, and the detection limit of device B is 0.5 ng / mL), a quantile standardization method is used to unify the dimensions. After preprocessing, the unit calculates a data quality score of 0-1 based on indicators such as noise level and missing value imputation error (the higher the score, the stronger the data reliability), outputs the structured preprocessed dataset, and transmits it to the feature fusion unit.

[0019] The feature fusion unit uses an exposure-effect database from an epidemiological cohort study as the source domain (including historical PFC exposure levels and health endpoint correlation data) and a structured preprocessed dataset as the target domain. It employs a weighted domain-adaptive model to align the feature distributions of the two domains. First, the unit calculates the quality weights of the target domain features based on the quality scores of the preprocessed data (higher quality scores result in greater weights). For example, a feature with a quality score of 0.8 has a weight of 0.7, and a feature with a score of 0.5 has a weight of 0.3. In minimizing the maximum mean difference (MMD) between the source and target domain feature distributions, the unit prioritizes aligning the distributions of high-quality weighted features (e.g., prioritizing the PFOS concentration feature with a quality score of 0.8) to reduce the interference of low-quality data on the fusion results. After fusion, the unit calculates a comprehensive feature quality index (e.g., a weighted average of the feature weights) based on the quality weights of each feature, generates a fused feature set and comprehensive index, and transmits them to the prediction unit.

[0020] The prediction unit employs a multi-task learning framework, including a main task and an auxiliary task. The main task takes the fused feature set as input and outputs the hazard ratio (HR) and confidence interval (CI) between PFC exposure and health endpoints (such as neurobehavioral developmental abnormalities in children) through a gradient boosting decision tree (GBDT) or neural network model. The auxiliary task takes the feature quality composite index as input and outputs a feature reliability score (reflecting the credibility of the contribution of fused features to health risk prediction) through a logistic regression model. During training, the loss function of the auxiliary task back-optimizes the quality weight allocation strategy of the feature fusion unit (e.g., if a feature has a low reliability score, its weight in feature fusion is reduced). Simultaneously, the unit introduces a semi-supervised learning mechanism, using unlabeled new detection data (samples without health endpoint labels) for self-training, and dynamically updates model parameters through pseudo-labeling technology, enabling the model to adapt to changes in feature distribution for new devices and sample types.

[0021] The output unit incorporates a dynamic risk threshold update engine. This engine, based on historical assessment results (e.g., actual incidence rate in low-risk groups) and the latest epidemiological data (e.g., baseline incidence rate at health endpoints), analyzes the changing trends of time-series data using a sliding window algorithm (e.g., a 5% increase in the incidence rate in low-risk groups over the past three months), and dynamically adjusts the risk threshold range (e.g., the original low-risk threshold of HR < 1.2 is tightened to HR < 1.15). The unit maps the HR and CI values ​​output by the prediction unit to the adjusted threshold range (e.g., HR = 1.1 corresponds to the new low-risk range), and combines this with the feature quality composite index to indicate the reliability of the risk level (e.g., a composite index above 0.7 is labeled "high reliability"). Finally, a multi-dimensional health assessment report is generated, including the risk level (low / medium / high risk), feature reliability score (e.g., 0.8), and the basis for threshold adjustment (e.g., "a 5% increase in the incidence rate in low-risk groups"). Simultaneously, the unit feeds back the feature reliability score to the feature fusion unit to optimize the quality weight allocation for the next cycle, forming a closed-loop optimization mechanism across the entire chain of "data acquisition-preprocessing-fusion-prediction-output-feedback."

[0022] Specifically, in the human health assessment system based on perfluorinated compound detection described in this invention, the data acquisition unit acquires biological sample data generated by the perfluorinated compound detection device through a multi-protocol data interface, which supports laboratory information management system protocols and medical information exchange protocols. While acquiring biological sample concentration values ​​and missing value markers, the data acquisition unit simultaneously records device operating status parameters and laboratory environmental parameters, constructing an enhanced raw dataset including concentration values, missing value markers, and multidimensional metadata.

[0023] In the human health assessment system based on perfluorinated compound detection described in this invention, the data acquisition unit acquires biological sample data and synchronously records metadata through a multi-protocol data interface. Its specific implementation includes the following logically coherent technical steps: The data acquisition unit is configured with a multi-protocol data interface that supports the Laboratory Information Management System (LIMS) protocol and the Medical Information Exchange (HL7) protocol, enabling communication with perfluorinated compound detection equipment (such as HPLC-MS / MS). The LIMS protocol is used to parse the structured test reports generated by the detection equipment (including fields such as sample number, test item, and concentration value), while the HL7 protocol is used to obtain sample metadata in medical scenarios (such as sample type and collection time). For example, after HPLC-MS / MS completes the detection of perfluorooctane sulfonate (PFOS) in a blood sample, the data acquisition unit reads the output test report from the equipment via the LIMS protocol and extracts the PFOS concentration value; simultaneously, it obtains the sample collection time (e.g., 2024-03-15 09:00) and sample type (e.g., venous blood) via the HL7 protocol, completing the acquisition of basic biological sample data.

[0024] While acquiring biological sample concentration values, the data acquisition unit simultaneously records two types of metadata: one type is equipment operating status parameters, including but not limited to parameters reflecting the equipment's working status such as mass spectrometer ion source voltage fluctuations, detector sensitivity, and column usage time; the other type is laboratory environmental parameters, including environmental parameters affecting sample stability such as sample storage temperature and humidity, laboratory atmospheric pressure, and ventilation system operating status. For example, if the ion source voltage fluctuates during HPLC-MS / MS detection (e.g., briefly rising from 3.5kV to 3.7kV), the data acquisition unit will read the fluctuation value in real time through the equipment interface; if the sample storage environment temperature is 26℃ (exceeding the standard storage temperature by 2-8℃), the unit will simultaneously record the abnormal temperature and humidity values.

[0025] To address potential missing values ​​during the detection process (such as the inability to output a specific value when the sample concentration is below the equipment's detection limit), the data acquisition unit uses standardized labeling rules for recording. For example, if the detection limit of HPLC-MS / MS is 0.1 ng / mL, and the PFOS concentration of a urine sample is found to be below this threshold, the unit will label the PFOS concentration of that sample as "ND" (Not Detected) and associate it with the detection limit value (0.1 ng / mL) and the equipment model (such as AB Sciex 6500+), thus creating a missing value label.

[0026] Finally, the data acquisition unit associates and stores biological sample concentration values ​​(e.g., 0.3 ng / mL), missing value markers (e.g., "ND"), and synchronously recorded equipment operating status parameters (e.g., ion source voltage fluctuation value of 0.2 kV) and laboratory environmental parameters (e.g., storage temperature of 26℃). A timestamp alignment mechanism is used during storage to ensure a one-to-one correspondence between each detection data point (e.g., the PFOS concentration value output at 09:05 on 2024-03-15) and the corresponding metadata at that time point (e.g., the ion source voltage fluctuation value at 09:05 on 2024-03-15). This constructs an enhanced raw dataset including the original detection data and multidimensional metadata, and transmits this dataset to the data preprocessing unit via an internal data bus, providing a complete data foundation for subsequent data quality assessment and feature fusion.

[0027] Specifically, in the human health assessment system based on perfluorinated compound detection described in this invention, the data preprocessing unit stores metadata and a data quality mapping knowledge base. The data quality mapping knowledge base includes quantitative association rules between device voltage fluctuation values ​​and noise levels of detection values, and quantitative association rules between temperature and humidity values ​​and sample degradation rates. The data preprocessing unit, based on the quantization rules in the data quality mapping knowledge base, identifies high-noise data by comparing device voltage fluctuation values ​​with preset thresholds, identifies easily degradable samples by comparing temperature and humidity values ​​with standard ranges, and selects the corresponding preprocessing strategy. Median filtering is used for high-noise data, time decay interpolation is used to fill in missing values ​​for easily degradable samples, and quantile standardization is used to unify the units for data from different devices, outputting a structured preprocessed dataset with labeled data quality scores.

[0028] In the human health assessment system based on perfluorinated compound detection described in this invention, the data preprocessing unit achieves quality-aware preprocessing of enhanced raw data through rule-driven mechanisms of a knowledge base mapping metadata to data quality. Its technical implementation includes the following logically coherent steps: The data preprocessing unit's built-in metadata and data quality mapping knowledge base stores two types of quantitative correlation rules: one type correlates equipment operating status with data noise, such as "when the ion source voltage fluctuation of the mass spectrometer exceeds the set threshold, the noise level of the detected value increases"; the other type correlates laboratory environmental parameters with sample stability, such as "when the sample storage temperature and humidity exceed the standard range, the degradation risk of perfluorinated compounds in urine samples increases." These rules are derived from historical detection data and quality verification experiments. For example, by comparing repeated detection data under different ion source voltage fluctuations, a positive correlation between voltage fluctuation values ​​and noise levels is determined; by observing sample degradation through temperature and humidity control experiments, a correlation between temperature and humidity values ​​and degradation risk is established.

[0029] Based on the quantification rules in the knowledge base, the data preprocessing unit first parses the metadata in the enhanced raw dataset: it extracts the device voltage fluctuation value corresponding to each detection data point (e.g., the ion source voltage rises from 3.5kV to 3.7kV during the detection of a sample) and compares it with a preset threshold (e.g., 3.6kV). If the fluctuation value exceeds the threshold, the data is determined to be high-noise data. At the same time, it extracts the sample storage temperature and humidity values ​​(e.g., temperature 26℃, humidity 70%) and compares them with the standard range (e.g., temperature 2-8℃, humidity 40-60%). If they exceed the range, the sample is determined to be a degradable sample. For example, if the detection data of a blood sample corresponds to an ion source voltage fluctuation value of 0.2kV (threshold is 0.15kV), it is marked as high-noise data; if a urine sample is stored at 26℃ (standard range 2-8℃), it is marked as a degradable sample.

[0030] For different types of data, the preprocessing unit selects corresponding preprocessing strategies: for high-noise data (such as detection values ​​where the ion source voltage fluctuation exceeds the standard), a median filtering algorithm is used to reduce noise interference. This algorithm traverses the data sequence through a sliding window, taking the median of the data within the window to replace the original value, thus reducing the impact of random noise on the detection results. For example, a high-noise data sequence is [0.32, 0.28, 0.35, 0.41, 0.30], which is updated to [0.32, 0.35, 0.35, 0.35, 0.30] after 3-point median filtering, reducing the interference of outliers.

[0031] For missing values ​​in easily degradable samples (such as the "not detected" label indicating a concentration below the detection limit due to degradation), a time-decay interpolation algorithm is used to complete the missing values. This algorithm dynamically adjusts the interpolation weights based on the time difference between sample collection and detection. The larger the time difference (the longer the sample has degraded), the smaller the interpolation weight (i.e., a more conservative estimate of the concentration). For example, if a urine sample is collected at time T0 and detected at time T1 (24 hours apart), the algorithm combines the concentration distribution of non-degradable samples in the same batch and calculates the completed value by weighting it according to the time difference, avoiding overestimation of concentration caused by simple mean imputation.

[0032] For data generated by different detection devices (e.g., device A has a lower detection limit, device B has a higher detection limit), quantile standardization is used to unify the dimensions. This method eliminates the influence of device differences on data scale by mapping data from different devices to the same quantile distribution. For example, if device A outputs a PFOS concentration sequence of [0.2, 0.4, 0.6], and device B outputs a sequence of [0.1, 0.3, 0.5], after quantile standardization, the quantile distributions of the two devices are consistent (e.g., the minimum value corresponds to 0.1, the median to 0.3, and the maximum value to 0.5), ensuring the comparability of subsequent feature fusion.

[0033] After preprocessing, the data preprocessing unit calculates a data quality score of 0-1 based on the noise filtering effect (e.g., changes in data variance before and after filtering), missing value completion error (e.g., deviation between completed and true values), and dimensional uniformity (e.g., consistency of data distribution after standardization). (A higher score indicates stronger data reliability.) Finally, the unit associates and stores the processed data (e.g., filtered concentration values, completed missing values, and standardized device data) with the quality score, forming a structured preprocessed dataset. This dataset is then transmitted to the feature fusion unit via a data interface, providing a high-quality data foundation for subsequent feature space alignment and health risk prediction.

[0034] Specifically, in the human health assessment system based on perfluorinated compound detection described in this invention, the feature fusion unit adopts a weighted domain adaptive model, with the exposure and effect database of epidemiological cohort studies as the source domain and the structured preprocessed dataset as the target domain. The feature fusion unit calculates the quality weights of the target domain features based on the data quality score. In the process of minimizing the maximum mean difference between the feature distributions of the source domain and the target domain, it prioritizes aligning the distribution of high-quality features with high data quality scores, and calculates the feature quality comprehensive index based on the quality weights, outputting the fused feature set and the feature quality comprehensive index.

[0035] In the human health assessment system based on perfluorinated compound detection described in this invention, the feature fusion unit achieves feature space alignment of multi-source data through a weighted domain adaptive model. Its technical implementation includes the following logically coherent steps: The feature fusion unit takes as input the structured preprocessed dataset (target domain) output by the data preprocessing unit and the exposure and effect database (source domain) from epidemiological cohort studies. The source domain data includes historically accumulated correlation data between perfluorinated compound exposure levels (e.g., PFOS concentration) and health endpoints (e.g., neurobehavioral developmental abnormalities in children), the reliability of which has been validated through long-term epidemiological studies. The target domain data consists of currently detected structured preprocessed data (e.g., test values ​​and quality scores of recently collected blood and urine samples). The core objective of this unit is to map the target domain features to a feature representation space consistent with the source domain, resolving the issue of inconsistent feature distribution caused by differences in testing equipment and sample types.

[0036] To differentiate the reliability of different features within the target domain, the feature fusion unit first calculates the quality weights of the target domain features based on the quality scores of the structured preprocessed data. The quality scores (0-1 points, with higher scores indicating more reliable data) are generated by the data preprocessing unit based on indicators such as noise level and missing value imputation error. The quality weights are positively correlated with the quality scores. For example, if a PFOS concentration feature has a quality score of 0.8, its quality weight is 0.7; if a PFOA concentration feature has a quality score of 0.5, its quality weight is 0.3. This weighting mechanism ensures that the subsequent feature alignment process focuses more on high-reliability features, reducing the interference of low-quality data on the fusion results.

[0037] During the feature distribution alignment stage, the unit employs a weighted domain adaptive model to achieve feature space alignment by minimizing the maximum mean difference (MMD) between the feature distributions of the source and target domains. MMD is a metric that measures the difference between two distributions. It is calculated using a kernel function to measure the distance between the features of the source and target domains in the regeneration kernel Hilbert space; a smaller distance indicates closer distributions. When calculating MMD, the unit applies quality weights to the target domain features: features with high quality weights (such as PFOS concentration with a quality score of 0.8) contribute more weight in the MMD calculation, preferentially pushing their distribution towards the source domain alignment; features with low quality weights (such as PFOA concentration with a quality score of 0.5) contribute less weight, avoiding misalignment due to unreliable data. For example, if the PFOS concentration distribution in the source domain is concentrated in the 0.2-0.5 ng / mL range, while the distribution of a high-quality PFOS feature in the target domain is concentrated in the 0.3-0.6 ng / mL range, the model will preferentially adjust the representation of that feature, making its distribution closer to the 0.2-0.5 ng / mL range of the source domain.

[0038] After feature alignment, the unit calculates a comprehensive feature quality index based on the quality weights of each feature. This index comprehensively reflects the overall reliability of the fused features through weighted summation or weighted averaging. For example, if the fused features include two sub-features, PFOS (weight 0.7) and PFOA (weight 0.3), its comprehensive index can be expressed as 0.7 × 0.8 (PFOS quality score) + 0.3 × 0.5 (PFOA quality score) = 0.71. A higher value indicates a stronger overall reliability of the fused features.

[0039] Ultimately, the feature fusion unit outputs two parts: a fused feature set (a new feature representation obtained by mapping the target domain features to a feature space consistent with the source domain) and a feature quality index (reflecting the reliability of the fused features). The fused feature set provides unified input features for subsequent health risk prediction, while the feature quality index is used to evaluate the reliability of the prediction results and is fed back to the prediction unit to assist model training. Through this process, the unit achieves effective fusion of multi-source heterogeneous data, laying a feature foundation for improving the accuracy of health risk prediction.

[0040] Specifically, in the human health assessment system based on perfluorinated compound detection described in this invention, the prediction unit adopts a multi-task learning framework. The main task receives the fused feature set and outputs the risk ratio and confidence interval, while the auxiliary task receives the feature quality comprehensive index and outputs the feature reliability score. The prediction unit adjusts the quality weight allocation strategy of the feature fusion unit based on the feature reliability score, and at the same time uses unlabeled new detection data to self-train the model and dynamically update the model parameters to adapt to the feature distribution changes of new devices and new sample types.

[0041] In the human health assessment system based on perfluorinated compound detection described in this invention, the prediction unit achieves synergy between health risk prediction and feature reliability verification through a multi-task learning framework, while dynamically adapting to changes in data distribution using semi-supervised learning. Its technical implementation includes the following logically coherent steps: The prediction unit employs a multi-task learning framework comprising two sub-tasks: a main task and an auxiliary task. The main task takes the fused feature set output by the feature fusion unit as input and aims to predict the association risk between perfluorinated compound (PFOC) exposure and specific health endpoints (such as neurobehavioral developmental disorders in children and metabolic syndrome in adults). The main task model can utilize supervised learning models such as Gradient Boosting Decision Tree (GBDT) or Deep Neural Network (DNN), trained using historical labeled data (including PFOC exposure levels and health endpoint results), and outputs a hazard ratio (HR) and confidence interval (CI). HR represents the risk multiple between the exposed and unexposed groups, and CI represents the confidence range of the HR. For example, if the main task inputs the fused features of a sample (including PFOS concentration, PFOA concentration, etc.), the model might output an HR of 1.3 (95% CI: 1.1-1.5), indicating that the health risk at that exposure level is 30% higher than that of the unexposed group.

[0042] The auxiliary task takes the feature quality composite index output by the feature fusion unit as input and aims to evaluate the reliability of the fused features in predicting health risks. The auxiliary task model can use logistic regression or a lightweight neural network, trained using correlation data between the feature quality composite index (0-1 points, higher values ​​indicating greater reliability) and historical prediction errors, outputting a feature reliability score of 0-1. For example, if the composite index of a fused feature is 0.8, the auxiliary task might output a reliability score of 0.75, indicating a high degree of confidence in the feature's contribution to the prediction result; if the composite index is 0.5, the reliability score might be 0.4, indicating a lower degree of confidence in the feature's predictive contribution.

[0043] The prediction unit utilizes a multi-task learning framework to collaboratively optimize the primary and auxiliary tasks. During training, the loss function of the primary task (e.g., cross-entropy loss) optimizes the accuracy of health risk prediction, while the loss function of the auxiliary task (e.g., mean squared error loss) optimizes the accuracy of feature reliability scoring. The two loss functions are jointly optimized by weighted summation, enabling the model to improve prediction accuracy while more precisely identifying reliable features. For example, if the primary task's prediction error is large, the model will prioritize adjusting the primary task parameters to reduce the error; conversely, if the auxiliary task's prediction error for reliability scoring is large, the model will adjust the auxiliary task parameters to improve scoring accuracy.

[0044] Based on the feature reliability scores output by the auxiliary task, the prediction unit dynamically adjusts the quality weight allocation strategy of the feature fusion unit. Specifically, if a feature has a low reliability score (e.g., 0.4), it indicates that the feature's contribution to health risk prediction is unreliable, and the unit sends an adjustment instruction to the feature fusion unit to reduce the weight of that feature in the target domain feature quality weight calculation (e.g., from 0.3 to 0.2); if a feature has a high reliability score (e.g., 0.75), its weight is increased (e.g., from 0.5 to 0.6). Through this feedback mechanism, the feature fusion unit can optimize the feature alignment strategy according to the actual prediction effect, improving the reliability of subsequent fused features.

[0045] To adapt to changes in feature distribution brought about by new equipment and sample types, the prediction unit introduces a semi-supervised learning mechanism, using unlabeled new detection data (samples without health endpoint labels) for self-training. During self-training, the model first uses the trained main task to generate pseudo-labels for the unlabeled data (such as predicting its health risk level), and then mixes the pseudo-labeled data with the original labeled data to retrain the model parameters. For example, when a new model of HPLC-MS / MS equipment is connected to the system, the generated detection data may have a new feature distribution (such as an expanded detection range for PFOS concentration). The model learns the feature representation under the new distribution through self-training and updates the parameters to maintain prediction accuracy.

[0046] Through collaborative optimization of multi-task learning, feedback adjustment of feature reliability scores, and dynamic updates of semi-supervised learning, the prediction unit achieves end-to-end optimization from feature input to risk output. This not only improves the accuracy of health risk prediction but also enhances the system's adaptability to multi-source heterogeneous data, providing robust technical support for the quantitative assessment of health risks related to perfluorinated compound exposure.

[0047] Specifically, in the human health assessment system based on perfluorinated compound detection described in this invention, the output unit is equipped with a risk threshold dynamic update engine. The risk threshold dynamic update engine analyzes the time series data of historical assessment results and the changing trends of epidemiological data through a sliding window algorithm based on historical assessment results and the latest epidemiological data, and adjusts the risk threshold range accordingly. The output unit maps the risk ratio and confidence interval to the adjusted threshold interval, and combines the feature quality comprehensive index to mark the reliability of the risk level, generating a health assessment report including the risk level, feature reliability score, and the basis for threshold adjustment.

[0048] In the human health assessment system based on perfluorinated compound detection described in this invention, the output unit adaptively adjusts the risk threshold through a dynamic risk threshold update engine and generates a multi-dimensional health assessment report by combining feature quality information. Its technical implementation includes the following logically coherent steps: The output unit's built-in dynamic risk threshold update engine takes historical assessment results and the latest epidemiological data as input. Historical assessment results include health assessment reports previously generated by the system, covering the actual incidence of disease in different risk levels (e.g., low, medium, and high risk groups) (e.g., the number of samples with health endpoint events in low-risk groups); the latest epidemiological data comes from baseline statistics on health endpoints published by authoritative institutions (e.g., the annual incidence of neurobehavioral developmental disorders in children and the regional prevalence of metabolic syndrome in adults). The engine regularly obtains the above data through a data interface (e.g., updating epidemiological data monthly and synchronizing historical assessment results daily) to ensure the timeliness and accuracy of the input information.

[0049] To analyze data trends, the engine employs a sliding window algorithm to process time-series data. This algorithm extracts historical assessment results and epidemiological data within a fixed-length time window (e.g., the most recent 12 months) and calculates the changing trends of key indicators within that window. For example, if the actual incidence rate in the low-risk group rises from 0.5% to 0.7% within the window, the engine identifies a trend of "increased risk of disease in the low-risk group"; if the latest epidemiological data shows that the baseline incidence rate of a health endpoint has increased compared to the previous year, the engine records a trend of "increased baseline risk." These trend analyses provide data support for threshold adjustments.

[0050] Based on trend analysis results, the engine dynamically adjusts the risk threshold range. The adjustment strategy for the threshold range is related to the trend direction: if the incidence rate in low-risk groups increases, the engine will tighten the low-risk threshold (e.g., the original low-risk threshold of HR < 1.2 will be adjusted to HR < 1.15), causing more samples to be classified as medium-risk, thus providing a more stringent warning of potential risks; if the baseline incidence rate decreases, the engine may relax the high-risk threshold (e.g., the original high-risk threshold of HR ≥ 1.5 will be adjusted to HR ≥ 1.6) to avoid over-warning. The adjusted threshold range is stored in the system configuration file and synchronized to the prediction unit to ensure the consistency of subsequent risk mapping.

[0051] The output unit maps the hazard ratio (HR) and confidence interval (CI) output by the prediction unit to the adjusted threshold intervals to determine the risk level of a sample. For example, if the adjusted threshold intervals are low risk (HR < 1.15), medium risk (1.15 ≤ HR < 1.3), and high risk (HR ≥ 1.3), a sample with an HR of 1.2 is mapped to a medium risk level. During the mapping process, the unit also considers the coverage of the confidence interval (e.g., if the CI includes the threshold boundary, it may be marked as "boundary risk"), improving the rigor of the risk level determination.

[0052] To indicate the reliability of risk levels, the output unit combines the feature quality composite index (0-1 points, with higher values ​​indicating more reliable features) provided by the feature fusion unit. For example, a composite feature quality index of 0.8 indicates a risk level of "high reliability"; 0.6 indicates "medium reliability"; and 0.4 indicates "low reliability". This labeling reflects the credibility of the feature data upon which the risk prediction results rely, helping users determine the reference value of the assessment results.

[0053] Ultimately, the output unit generates a health assessment report that includes a risk level (e.g., "medium risk"), a feature reliability score (e.g., "0.6 (medium reliable)"), and the basis for threshold adjustments (e.g., "Incidence rate has increased in low-risk groups in the past three months"). The report displays key information through a visual interface (e.g., tables, charts) and includes a summary of trend analysis for threshold adjustments (e.g., "Based on incidence data from low-risk groups in January-March 2024, the incidence rate increased from 0.5% to 0.7%)), providing interpretable health risk assessment results for users (e.g., public health administrators, clinicians). Through this process, the output unit achieves dynamic adaptation of risk thresholds and transparent display of assessment results, enhancing the system's decision support value in practical applications.

[0054] Secondly, the human health assessment method based on the detection of perfluorinated compounds provided by the present invention, applied to the aforementioned human health assessment system based on the detection of perfluorinated compounds, includes: Collect multidimensional information such as biological sample concentration values, missing value markers, equipment operating status, and laboratory environment generated by perfluorinated compound detection equipment to construct an enhanced original dataset; Based on the metadata and data quality mapping knowledge base, noise filtering, missing value completion and unit unification processing are performed on the enhanced original dataset to output a structured preprocessed dataset with labeled data quality scores. Based on the data quality score, feature quality weights are calculated, and a weighted domain adaptive model is used to align the feature distributions of the source and target domains to generate a fused feature set and a comprehensive feature quality index. Using a multi-task learning framework, the risk ratio and confidence interval are output based on the fused feature set, and the feature reliability score is output based on the feature quality comprehensive index. The system receives the risk ratio and confidence interval, combines them with dynamically adjusted threshold intervals to generate a health assessment report, and feeds back the feature reliability score to the feature fusion unit to optimize the weight allocation strategy.

[0055] The human health assessment method based on the detection of perfluorinated compounds provided by this invention achieves multi-stage coordination through electro-digital data processing technology. Its technical process can be decomposed into the following logically coherent steps: The first step of the method is the acquisition of biological sample data and metadata. The system establishes communication with perfluorinated compound detection equipment (such as high-performance liquid chromatography-tandem mass spectrometry-mass spectrometry / analytical mass spectrometry) through a multi-protocol data interface. This interface supports the Laboratory Information Management System (LIMS) protocol and the Medical Information Exchange (HL7) protocol, which are used to parse the structured detection report output by the detection equipment (including sample number, perfluorooctane sulfonate (PFOS) concentration value, etc.) and sample metadata in a medical setting (such as sample type, acquisition time). While acquiring the biological sample concentration value (e.g., 0.3 ng / mL), the system simultaneously records two types of metadata: one is the equipment operating status parameters (such as mass spectrometer ion source voltage fluctuation value, detector sensitivity), and the other is the laboratory environmental parameters (such as sample storage temperature and humidity, atmospheric pressure). For example, when HPLC-MS / MS detects that the PFOS concentration in a urine sample is below the detection limit, the system will mark the value as "ND" (not detected) and associate it with the ion source voltage fluctuation value (e.g., 0.2kV) and the sample storage temperature (e.g., 26℃). Finally, the concentration value, missing value label, and multidimensional information are associated and stored to construct an enhanced raw dataset, providing a complete data foundation for subsequent data processing.

[0056] The second step is heterogeneous data preprocessing based on quality awareness. The system has a built-in knowledge base mapping metadata and data quality, storing the association rules between device operating status, environmental parameters, and data quality (such as "the noise of the detection value increases when the ion source voltage fluctuation exceeds the set threshold" and "the degradation risk increases when the sample storage temperature and humidity exceed the standard range"). The preprocessing stage first analyzes the metadata in the enhanced original dataset: if the ion source voltage fluctuation exceeds the threshold (e.g., 0.15kV) during the detection of a sample, it is judged as high-noise data; if the sample storage temperature exceeds the standard range (e.g., 2-8℃), it is judged as an easily degradable sample. For high-noise data, the system uses a median filtering algorithm to reduce noise interference (e.g., taking the median through a sliding window to replace the original value); for the "undetected" missing values ​​of easily degradable samples, a time decay interpolation algorithm (dynamically adjusting the interpolation weight according to the time difference between sample collection and detection) is used to complete the missing values; for data generated by different devices (e.g., device A has a lower detection limit, device B has a higher detection limit), quantile standardization is used to unify the dimensions (mapping the data to the same quantile distribution). After preprocessing, the system calculates a data quality score of 0-1 based on indicators such as noise filtering effect and missing value imputation error (the higher the score, the more reliable the data), and outputs a structured preprocessed dataset to provide high-quality input for feature fusion.

[0057] The third step is weighted feature space fusion. The system uses an exposure and effect database from epidemiological cohort studies (source domain, including historical PFC exposure and health endpoint correlation data) as a benchmark and a structured preprocessed dataset (target domain, current detection data) as the object. A weighted domain adaptive model aligns the feature distributions of the two domains. First, the quality weights of the target domain features are calculated based on the quality scores of the preprocessed data (higher quality scores result in greater weights; for example, a feature with a quality score of 0.8 has a weight of 0.7). In minimizing the maximum mean difference (MMD) between the source and target domain feature distributions, the distribution of high-quality weighted features is prioritized for alignment (e.g., adjusting the representation of PFOS concentration features first) to reduce interference from low-quality data. After fusion, the system calculates a comprehensive feature quality index based on the quality weights of each feature (e.g., a weighted average of the feature weights), generating a fused feature set and a comprehensive index, providing unified input features for health risk prediction.

[0058] The fourth step is multi-task collaborative health risk prediction. The system adopts a multi-task learning framework, including a main task and an auxiliary task: the main task takes a fused feature set as input and outputs the hazard ratio (HR) and confidence interval (CI) between PFC exposure and health endpoints (such as neurobehavioral developmental abnormalities in children) through a gradient boosting decision tree (GBDT) or neural network model, reflecting the risk multiple between the exposed and unexposed groups; the auxiliary task takes a feature quality comprehensive index as input and outputs a feature reliability score (0-1 points, with higher values ​​indicating more reliable contributions to the prediction) through a logistic regression model. During training, the main task optimizes prediction accuracy, and the auxiliary task optimizes the accuracy of the reliability score. The two work together to adjust the model parameters through a joint loss function. At the same time, the system introduces a semi-supervised learning mechanism, using unlabeled new detection data (samples without health endpoint labels) to generate pseudo-labels and retrain the model, dynamically updating parameters to adapt to changes in feature distribution for new devices and new sample types.

[0059] The fifth step involves generating and feeding back risk reports using dynamic thresholds. The system has a built-in dynamic risk threshold update engine. Based on historical assessment results (such as actual incidence rates in low-risk groups) and the latest epidemiological data (such as baseline incidence rates at health endpoints), it analyzes the changing trends of time-series data using a sliding window algorithm (such as an increase in incidence rates in low-risk groups over the past three months) and dynamically adjusts the risk threshold range (e.g., tightening the original low-risk threshold HR<1.2 to HR<1.15). The system maps the predicted HR and CI values ​​to the adjusted threshold range (e.g., HR=1.2 corresponds to medium risk), and combines the reliability of the risk level labeling using the feature quality composite index (e.g., a composite index of 0.8 is labeled "high reliability") to generate a health assessment report including the risk level (low / medium / high risk), feature reliability score (e.g., 0.75), and the basis for threshold adjustment (e.g., "increased incidence rates in low-risk groups"). Simultaneously, the system feeds back the feature reliability score to the feature fusion unit to optimize the feature quality weight allocation for the next cycle.

[0060] This invention improves the feature fusion efficiency of biological sample detection data through multi-stage collaborative design, providing high-quality input for statistical learning models. The data acquisition stage not only obtains basic detection data such as perfluorinated compound concentration values ​​and missing value labels, but also simultaneously records multi-dimensional metadata such as equipment operating status (e.g., mass spectrometer ion source voltage fluctuations) and laboratory environmental parameters (e.g., sample storage temperature and humidity), constructing an enhanced dataset including raw data and quality-related metadata. The preprocessing stage, based on the association rules between metadata and data quality (e.g., the positive correlation between voltage fluctuations and noise levels, and the correlation between temperature and humidity and sample degradation risk), employs median filtering for noise reduction of high-noise data, time-decay interpolation for missing values ​​in easily degradable samples, and quantile standardization for data from different devices to unify dimensions, assigning a data quality score of 0-1. These steps eliminate data heterogeneity, improve data reliability, and provide a structured, high-quality input foundation for feature fusion.

[0061] The feature fusion stage addresses the issue of multi-source data distribution discrepancies through a weighted domain-adaptive model. Using an exposure-effect database from an epidemiological cohort study as the source domain (with validated and reliable feature distributions) and preprocessed current detection data as the target domain, quality weights for target domain features are calculated based on data quality scores (higher quality scores result in greater weights). In minimizing the maximum mean difference between the feature distributions of the source and target domains, the distribution of high-quality weighted features is prioritized for alignment (e.g., prioritizing the adjustment of PFOS concentration features with high quality scores), reducing the interference of low-quality data on the fusion results. This mechanism maps heterogeneous data to a unified feature space, improving the efficiency and reliability of feature fusion and enabling statistical learning models to more accurately capture the association between perfluorinated compound exposure and health endpoints.

[0062] The prediction and feedback stages further enhance model accuracy through multi-task learning and dynamic optimization. The prediction unit employs a multi-task learning framework: the primary task outputs risk ratios and confidence intervals based on fused features, while the auxiliary task outputs feature reliability scores based on a comprehensive feature quality index. The reliability scores from the auxiliary tasks are fed back to the feature fusion unit, dynamically adjusting the feature quality weight allocation. Simultaneously, a semi-supervised learning mechanism is introduced, utilizing unlabeled new detection data to self-train and update model parameters, adapting to changes in feature distribution for new devices and sample types. The output unit combines dynamically adjusted risk thresholds and feature quality information to generate reports, ensuring continuous model learning and optimization, ultimately improving the quantitative accuracy of health risks related to perfluorinated compound exposure.

Claims

1. A human health assessment system based on the detection of perfluorinated compounds, characterized in that, It includes a data acquisition unit, a data preprocessing unit, a feature fusion unit, a prediction unit, and an output unit; The data acquisition unit collects biological sample concentration values, missing value markers, and multidimensional information such as equipment operating status and laboratory environment generated by the perfluorinated compound detection device, constructs an enhanced raw dataset, and transmits it to the data preprocessing unit. The data preprocessing unit performs noise filtering, missing value completion, and unit unification processing on the enhanced original dataset based on the metadata and data quality mapping knowledge base, and outputs a structured preprocessed dataset labeled with data quality scores to the feature fusion unit. The feature fusion unit calculates feature quality weights based on the data quality score, aligns the feature distributions of the source and target domains using a weighted domain adaptive model, generates a fused feature set and a feature quality comprehensive index, and transmits them to the prediction unit. The prediction unit uses a multi-task learning framework to output risk ratio and confidence interval based on the fused feature set, and outputs feature reliability score based on feature quality comprehensive index. The output unit receives the risk ratio and confidence interval, combines them with the dynamically adjusted threshold interval to generate a health assessment report, and feeds back the feature reliability score to the feature fusion unit to optimize the weight allocation strategy. The data preprocessing unit stores metadata and a data quality mapping knowledge base. The data quality mapping knowledge base includes quantitative association rules between device voltage fluctuation values ​​and noise levels of detection values, and quantitative association rules between temperature and humidity values ​​and sample degradation rates. The data preprocessing unit, based on the quantization rules in the data quality mapping knowledge base, identifies high-noise data by comparing device voltage fluctuation values ​​with preset thresholds, identifies easily degradable samples by comparing temperature and humidity values ​​with standard ranges, and selects the corresponding preprocessing strategy. Median filtering is used for high-noise data, time decay interpolation is used to fill in missing values ​​for easily degradable samples, and quantile standardization is used to unify the units for data from different devices, outputting a structured preprocessed dataset with labeled data quality scores. The feature fusion unit adopts a weighted domain adaptive model, using the exposure and effect database of epidemiological cohort studies as the source domain and the structured preprocessed dataset as the target domain. The feature fusion unit calculates the quality weight of the target domain features based on the data quality score. In the process of minimizing the maximum mean difference between the feature distributions of the source domain and the target domain, it prioritizes aligning the distribution of high-quality features with high data quality scores, and calculates the feature quality comprehensive index based on the quality weight, outputting the fused feature set and the feature quality comprehensive index. The prediction unit adopts a multi-task learning framework. The main task receives the fused feature set and outputs the risk ratio and confidence interval, while the auxiliary task receives the feature quality comprehensive index and outputs the feature reliability score. The prediction unit adjusts the quality weight allocation strategy of the feature fusion unit based on the feature reliability score, and at the same time uses unlabeled new detection data to self-train the model and dynamically update the model parameters to adapt to the feature distribution changes of new devices and new sample types.

2. The human health assessment system based on the detection of perfluorinated compounds according to claim 1, characterized in that, The data acquisition unit acquires biological sample data generated by the perfluorinated compound detection device through a multi-protocol data interface, which supports the laboratory information management system protocol and the medical information exchange protocol. While acquiring biological sample concentration values ​​and missing value markers, the data acquisition unit simultaneously records device operating status parameters and laboratory environmental parameters to construct an enhanced raw dataset including concentration values, missing value markers, and multidimensional metadata.

3. The human health assessment system based on perfluorinated compound detection according to claim 2, characterized in that, The output unit is equipped with a risk threshold dynamic update engine. The risk threshold dynamic update engine analyzes the time series data of historical assessment results and the changing trends of epidemiological data through a sliding window algorithm based on historical assessment results and the latest epidemiological data, and adjusts the risk threshold range accordingly. The output unit maps the risk ratio and confidence interval to the adjusted threshold interval, and combines the feature quality comprehensive index to mark the reliability of the risk level, generating a health assessment report including the risk level, feature reliability score, and the basis for threshold adjustment.

4. A method for human health assessment based on the detection of perfluorinated compounds, applied to the human health assessment system based on the detection of perfluorinated compounds as described in any one of claims 1-3, characterized in that, include: Step 1: Collect multidimensional information such as biological sample concentration values, missing value markers, equipment operating status, and laboratory environment generated by the perfluorinated compound detection device to construct an enhanced original dataset; Step 2: Based on the metadata and data quality mapping knowledge base, perform noise filtering, missing value completion and unit unification processing on the enhanced original dataset to output a structured preprocessed dataset with labeled data quality scores. Step 3: Calculate feature quality weights based on data quality scores, align feature distributions between the source and target domains using a weighted domain adaptive model, and generate a fused feature set and a comprehensive feature quality index. Step 4: Using a multi-task learning framework, output the risk ratio and confidence interval based on the fused feature set, and output the feature reliability score based on the feature quality comprehensive index. Step 5: Receive the risk ratio and confidence interval, combine them with the dynamically adjusted threshold interval to generate a health assessment report, and feed back the feature reliability score to the feature fusion unit to optimize the weight allocation strategy.

Images