Regional respiratory system health risk situation awareness method and system

Abstract

The invention relates to a regional respiratory system health risk situation awareness method and system. The method comprises the following steps: firstly, collecting multi-source heterogeneous sensing data of a target area, and constructing a three-dimensional spatio-temporal data cube through spatio-temporal alignment and gridding mapping; fusing the three topologies to generate a comprehensive space situation matrix, and extracting and fusing node features through a multi-modal feature encoder; then generating a future respiratory system health risk situation map by using a deep learning model containing space attention aggregation and time sequence gating circulation; and finally, responding to external parameter correction data, and triggering model reasoning to output a risk evolution trajectory. Compared with the prior art, the method has the advantages of high accuracy and the like.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to a method and system for regional respiratory health risk situation perception. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a chronic respiratory disease characterized by persistent respiratory symptoms and airflow limitation. Its onset is closely related to multiple factors, including smoking, air pollution, and occupational exposure. According to the World Health Organization (WHO), COPD has become the third leading cause of death globally, placing a huge burden on public health systems. The risk of developing COPD is influenced by a combination of factors, including individual factors (such as age, sex, genetic susceptibility, and smoking history), environmental factors (such as air quality and meteorological conditions), and spatiotemporal distribution characteristics.

[0003] Traditional COPD risk assessment methods mainly rely on clinical questionnaires, lung function tests and other means. These methods have the following limitations: (1) they only focus on individual-level risk factors and ignore the dynamic changes of environmental factors; (2) they are difficult to capture the spatial correlation between different geographical regions and the disease transmission patterns; (3) they lack the ability to effectively model time series data and cannot achieve dynamic prediction of risk.

[0004] In recent years, with the development of big data technology and deep learning, researchers have begun to apply machine learning methods to disease risk prediction. For example, invention patent CN121011356A discloses a breast cancer cardiotoxicity risk prediction system based on multimodal data fusion. This system collects multi-source heterogeneous medical data such as electronic health records, dynamic electrocardiogram monitoring, and pathological slides through a distributed architecture, uses a hierarchical compression strategy and a dynamic risk knowledge graph to reveal the path of toxicity, and sets up a feedback loop to adjust the model.

[0005] However, the aforementioned prior art and similar existing technologies still have significant limitations when applied to regional respiratory health risk situational awareness: First, this approach primarily focuses on the integration of individual clinical data (such as electrocardiograms and pathology). While it achieves multimodal analysis, it lacks the ability to integrate macro-environmental factors (such as regional air quality, meteorological changes, and functional zoning). For diseases like COPD, which are heavily influenced by environmental triggers, analyzing only from a clinical perspective cannot comprehensively depict their risk profile. Secondly, while the scheme constructs a knowledge graph, its essence is based on logical relationships and lacks physical modeling of geospatial topology and population flow interactions. In regional health risk perception, pollutant diffusion between adjacent areas (physical adjacency) and cross-regional commuting (functional adjacency) are key pathways for risk transmission, and existing technologies cannot quantify this complex spatial transmission effect. This approach also fails to simulate hypothetical scenarios, focusing instead on static assessments or passive predictions of the current state, lacking the ability to anticipate proactive interventions. Consequently, its reference value in assisting the development of public health strategies is limited.

[0006] Graph Neural Networks (GNNs), as deep learning models for processing graph-structured data, can effectively capture the topological relationships and information propagation patterns between nodes. Spatial-Temporal Graph Neural Networks (ST-GNNs) further extend the capabilities of GNNs, simultaneously modeling spatial dependencies and temporal dynamics, providing a new approach to solving the aforementioned problems.

[0007] Nevertheless, existing COPD risk prediction methods, when combined with spatiotemporal deep learning, still have the following main shortcomings: (1) Insufficient data fusion capability: Existing methods are difficult to effectively integrate data from different sources and modalities (such as electronic health records, environmental monitoring data, meteorological data, population statistics, etc.), and cannot make full use of the complementary information of multi-source data; (2) Insufficient spatial correlation modeling: Traditional methods usually treat each region as an independent unit for analysis, ignoring the spatial correlation between adjacent regions, the risk transmission brought about by population flow, and the spatial diffusion effect of environmental factors; (3) Insufficient capture of time dynamics: Most existing prediction models are static models, which are difficult to capture the time evolution of disease risk with seasonal changes, air quality fluctuations and other factors; (4) Poor interpretability: Deep learning models are often regarded as "black boxes", making it difficult to explain the formation mechanism of prediction results, which is not conducive to clinical decision support and public health policy formulation; (5) Lack of dynamic simulation capability: Existing methods mainly focus on prediction and lack the ability to dynamically simulate the spatiotemporal evolution of disease risk, making it difficult to support "hypothesis-deduction" scenario analysis.

[0008] In summary, the current field of regional respiratory health risk situational awareness faces several challenges, including the difficulty in unifying and integrating multi-source heterogeneous data, the lack of effective modeling of complex spatial interactions between regions, fragmented spatiotemporal feature extraction, which limits prediction accuracy and lacks the ability to dynamically extrapolate scenarios to support policy formulation. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for regional respiratory health risk situational awareness.

[0010] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for regional respiratory health risk situational awareness is provided, the method comprising the following steps: S1. Collect multi-source heterogeneous sensing data of the target area, and perform spatiotemporal alignment and gridding mapping on the multi-source heterogeneous sensing data to construct a three-dimensional spatiotemporal data cube; S2. Based on multi-source heterogeneous sensing data, construct physical adjacency topology, interactive association topology, and homogeneous association topology, and integrate the above three topologies to generate a comprehensive spatial situation matrix. S3. Based on multi-source heterogeneous sensing data, a multimodal feature encoder is used to extract multimodal features and fuse them to obtain the fused node feature representation; S4. Construct a spatiotemporal deep learning model that includes a spatial attention aggregation mechanism and a temporal gating loop mechanism. In the model, the spatial attention coefficient matrix is ​​dynamically calculated based on the comprehensive spatial situation matrix using the spatial attention aggregation mechanism. The temporal gating loop mechanism is used to perform temporal extrapolation on the fused node feature representation. A future respiratory health risk situation map is generated based on the spatial attention coefficient matrix and the temporal extrapolation results. S5. In response to external input public health strategies or environmental governance parameters, directly correct the corresponding feature layer values ​​in the spatiotemporal data cube; and trigger the spatiotemporal deep learning model to perform forward inference, outputting the future respiratory health risk evolution trajectory under this input.

[0011] As a preferred technical solution, the multi-source heterogeneous sensing data in S1 includes desensitized clinical statistical characteristics, environmental meteorological monitoring data, regional population attribute parameters, and geospatial data. The clinical statistical characteristics of desensitization include: average lung function indicators within the grid area, frequency of respiratory symptom search terms, and density of historical confirmed cases; Environmental meteorological monitoring data include time series of fine particulate matter (PM2.5), inhalable particulate matter (PM10), nitrogen dioxide concentration, and relative humidity. Regional population attribute parameters include: population aging coefficient, residential density index, and regional functional zoning type code; The three-dimensional spatiotemporal data cube is orthogonally composed of a spatial domain, a temporal domain, and a feature domain, and is used to fully digitize the regional health environment status.

[0012] As a preferred technical solution, S2 employs a data-driven approach when constructing physical adjacency topology, interactive association topology, and homogeneous association topology. Physical adjacency topology is constructed based on geographical boundary connections; interactive association topology is constructed based on population flow intensity derived from mobile signaling data; and homogeneous association topology is constructed based on vector similarity of regional socioeconomic indicators. The specific process of S2 includes: Initialize scalar weights for the three matrices: physical adjacency, interaction association, and homogeneous association. During end-to-end training of the model, the scalar weights of the three matrices are updated through backpropagation of the loss function. The three matrices are weighted and superimposed using the updated scalar weights.

[0013] As a preferred technical solution, the specific process of S3 includes: Three parallel feature encoding branches are constructed, and the output dimension of each branch is uniformly mapped to the preset hidden layer dimension; In the first branch, the desensitized clinical statistical features are input into the multilayer perceptron encoder. The multilayer perceptron encoder is configured as a three-layer fully connected structure with the number of neurons decreasing sequentially to obtain the health feature vector. In the second branch, environmental meteorological monitoring data containing 15 features and spanning 4 weeks is input into the temporal convolutional network encoder. After convolutional extraction of the time series, the environmental feature vector is output. In the third branch, the longitude and latitude coordinates of the grid center are input into the position encoder to obtain the spatial feature vector; The health feature vector, environmental feature vector, and spatial feature vector are simultaneously input into the cross-attention layer; The input feature vectors are mapped to query vectors, key vectors, and value vectors respectively. The attention weight coefficients are calculated using the Softmax function, and the weight coefficients are used to perform a weighted summation of each feature vector to output the fused node feature representation.

[0014] As a preferred technical solution, the temporal convolutional network encoder in the second branch adopts a hierarchical causal convolutional structure, and sets the dilation rate of the convolutional kernel to increase exponentially with the increase of the number of network layers.

[0015] As a preferred technical solution, the specific execution logic of the spatial attention aggregation mechanism in S4 is as follows: Based on the comprehensive spatial situation matrix, the correlation scores between the central grid and its neighboring grids in the feature space are calculated; the correlation scores are then mapped to normalized attention coefficients through nonlinear activation; and the neighborhood features are weighted and summed according to the attention coefficients to obtain the spatial attention coefficient matrix.

[0016] As a preferred technical solution, in S4, the specific execution logic of the time-gated loop mechanism is as follows: receive the fused node features and the hidden state vector output from the previous time step; concatenate the current feature vector with the hidden state vector output from the previous time step, and perform linear transformation and Sigmoid activation operations through the reset gate weight matrix and the update gate weight matrix respectively to generate the reset gate value and the update gate value. The reset gate value is multiplied element-wise with the hidden state vector output from the previous time step. The result of the multiplication is concatenated with the feature vector at the current time step. Then, the current candidate hidden state is generated by processing the candidate state weight matrix and the hyperbolic tangent activation function. Using the update gate value as a weighting coefficient, the hidden state vector output from the previous time step and the current candidate hidden state are weighted and summed to obtain the hidden state vector at the current time step. The generated hidden state vector at the current time step is used as the input for the next time step and iterated until the calculation of the preset time window is completed. The final hidden state vector is then input into the fully connected layer for dimension mapping to obtain the risk level probability value of each grid region, thus obtaining the time series simulation result.

[0017] As a preferred technical solution, the specific process of S5 includes: Based on externally input public health strategies or environmental governance parameters, construct feature correction vectors; The feature correction vector is superimposed onto the corresponding time slice of the spatiotemporal data cube to construct a virtual situation space; By comparing the output results of the model on the original data cube and the virtual situation space, the magnitude of risk reduction and the time delay in the effectiveness of intervention measures are quantified.

[0018] As a preferred technical solution, the method also includes: Extract the spatial attention coefficient matrix generated during model inference; Based on the spatial attention coefficient matrix, high-weight grid areas are connected and mapped onto a geographic map to obtain a risk propagation flow diagram.

[0019] According to another aspect of the present invention, a regional respiratory health risk situational awareness system is provided, the system comprising: The multi-dimensional sensing module is used to collect multi-source heterogeneous sensing data and construct a three-dimensional spatiotemporal data cube; The situation topology construction module is used to calculate and fuse physical, interactive, and homogeneous multi-perspective adjacency relationships; The deep cognitive reasoning module has a built-in spatiotemporal deep learning model that is trained and configured to capture environmental lag effects using dilated convolution, aggregate spatial transmission risks using graph attention mechanisms, and deduce temporal trends using gated recurrent units. The simulation and inference module is used to dynamically adjust the data cube based on the intervention parameters input from the outside, and to infer and generate the future trajectory of respiratory health risk under that input.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs a unified fusion method for multi-source heterogeneous spatiotemporal data, significantly improving the accuracy and comprehensiveness of risk prediction. By proposing a multi-modal feature fusion strategy based on an attention mechanism, this invention effectively integrates multi-source heterogeneous data such as electronic health records, environmental monitoring, meteorology, demographics, and geographic information, achieving complementary advantages across different modalities. Unlike traditional methods that rely solely on a single environmental factor or historical case, this invention utilizes MLP, TCN, and Embedding layers to extract heterogeneous features, and then aligns and fuses them through a cross-attention mechanism. Experimental data show that this method achieves an accuracy of 87.3% in predicting COPD risk, with an AUC-ROC of 0.912, representing improvements of 5.2% and 4.8% respectively compared to the optimal baseline method (XGBoost), and a recall rate of up to 91.5% for identifying high-risk areas.

[0021] 2. This invention constructs a multi-layered spatiotemporal graph structure, capable of accurately capturing complex regional correlation patterns. The invention designs a three-layered spatial relationship modeling framework encompassing geographical adjacency, functional adjacency (population mobility), and similarity adjacency, comprehensively capturing the physical proximity, functional connections, and characteristic similarities between regions. This multi-perspective topology construction method enables the model to not only identify the cross-border transmission effects of environmental pollution (average impact distance 30-50km) but also discover disease risk transmission paths driven by population mobility. Data shows that between regional pairs with commuter flows exceeding 10,000 people / day, the model captures a risk correlation coefficient as high as r=0.68, effectively solving the problem that traditional geographical adjacency matrices cannot characterize the risk transmission of long-distance population mobility.

[0022] 3. This invention utilizes a spatial attention aggregation mechanism and employs a graph attention mechanism (GAT) to adaptively learn the influence weights of different neighboring nodes, overcoming the limitation of traditional graph convolution that treats all neighboring nodes with equal weight. The model can automatically focus on high-risk input sources based on real-time environmental and population dynamics, thereby accurately identifying the Top-5 key factors affecting COPD risk (PM2.5 concentration contribution 28%, age structure 23%, smoking rate 19%, temperature 15%, humidity 10%), significantly improving the model's robustness in complex urban environments.

[0023] 4. This invention organically combines spatial graph convolution (GAT) with time series modeling (GRU) to achieve joint learning of spatial dependencies and temporal dynamics, avoiding error accumulation in two-stage modeling. This architecture not only boasts high computational efficiency (single prediction time for 120 regions < 2 seconds) but also possesses multi-timescale prediction capabilities, supporting short-term (1 week), medium-term (2-4 weeks), and long-term (1-3 months) risk prediction, meeting diverse needs in emergency response, resource allocation, and strategic planning. Simultaneously, its unique dilated convolution structure effectively captures the seasonal variation patterns of COPD risk (e.g., winter risk is 42% higher than summer risk) and environmental lag effects.

[0024] 5. This invention possesses dynamic scenario simulation capabilities, supporting hypothesis-based risk evolution simulations by modifying environmental factor inputs. This provides a quantitative assessment tool for environmental policy formulation, medical resource allocation, and public health interventions. For example, in a scenario simulating a 30% reduction in PM2.5 concentration, the model shows a 23% reduction in the number of high-risk areas; an intervention that improves smoking cessation rates by 15% predicts a 12% reduction in COPD incidence within 5 years. This simulation capability helps policymakers identify the most contributing interventions (air quality improvement, smoking control), thereby optimizing resource allocation and improving resource utilization efficiency by more than 20%.

[0025] 6. This invention provides a multi-level interpretability analysis method, enhancing the model's credibility and clinical application value. By extracting the attention weight matrix, it achieves attention weight visualization, feature importance analysis (SHAP value), and risk propagation path identification, improving the interpretability and credibility of the model's prediction results and facilitating clinical understanding and policy application. The system can generate a visualized risk propagation flow diagram, intuitively displaying the risk source and its diffusion direction. Based on this accurate identification, it can guide the dynamic allocation of emergency beds and respiratory therapy equipment. Compared with traditional large-scale blind screening, screening costs can be reduced by approximately 40%, making it suitable for large-scale screening and improving its practicality. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the steps in a regional respiratory system health risk situation perception method according to the present invention; Figure 2 This is a schematic diagram of the multi-source data acquisition and preprocessing module in the embodiment; Figure 3a This is a schematic diagram of the data feature distribution in the embodiment; Figure 3b This is a schematic diagram of the data time resolution in the embodiment; Figure 3c This is a schematic diagram illustrating data quality in the embodiment; Figure 4a This is a spatiotemporal structure diagram of the geographic adjacency matrix in the embodiment; Figure 4b This is a geographic association topology diagram of the nodes in the embodiment; Figure 5a This is a matrix representation of the geographical adjacency graph in the embodiment; Figure 5b This is a schematic diagram of the matrix representation of the corresponding functional adjacency graph in the embodiment; Figure 6a The topological graph showing the feature similarity of nodes in the embodiment. Figure 6b This is a schematic diagram illustrating the matrix representation of corresponding feature similarity in the embodiment; Figure 7 This is a flowchart of the multimodal feature fusion process in the embodiment; Figure 8 This is a schematic diagram of the spatiotemporal graph neural network model architecture in the embodiment; Figure 9a This is a schematic diagram of attention calculation node i and its neighboring nodes in the embodiment. Figure 9b This is a schematic diagram of multi-head attention in the embodiment; Figure 9c This is a schematic diagram illustrating the adjustment of attention weights for each neighboring node in the embodiment. Figure 9d This is a schematic diagram of the multi-timescale risk prediction process in the embodiment; Figure 10a This is a flowchart illustrating the dynamic scenario simulation in the embodiment. Figure 10b This is a schematic diagram illustrating the risk evolution under different scenarios in the embodiments; Figure 11a This is a visualization heatmap of the spatiotemporal distribution of COPD risk in spring, as shown in the example. Figure 11b This is a visualization heatmap of the spatiotemporal distribution of COPD risk in summer, as shown in the example. Figure 11c This is a visualization heatmap of the spatiotemporal distribution of COPD risk in autumn, as shown in the example. Figure 11d This is a visualization heatmap of the spatiotemporal distribution of COPD risk in winter, as shown in the example. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] Example 1 In this embodiment, a regional respiratory health risk situational awareness method is adopted, and the method steps are as follows: Figure 1 As shown, it specifically includes: S1. Collect multi-source heterogeneous sensing data of the target area, and perform spatiotemporal alignment and gridding mapping on the multi-source heterogeneous sensing data to construct a three-dimensional spatiotemporal data cube; S2. Based on multi-source heterogeneous sensing data, construct physical adjacency topology, interactive association topology, and homogeneous association topology, and integrate the above three topologies to generate a comprehensive spatial situation matrix. S3. Based on multi-source heterogeneous sensing data, a multimodal feature encoder is used to extract multimodal features and fuse them to obtain the fused node feature representation; S4. Construct a spatiotemporal deep learning model that includes a spatial attention aggregation mechanism and a temporal gating loop mechanism. In the model, the spatial attention coefficient matrix is ​​dynamically calculated based on the comprehensive spatial situation matrix using the spatial attention aggregation mechanism. The temporal gating loop mechanism is used to perform temporal extrapolation on the fused node feature representation. A future respiratory health risk situation map is generated based on the spatial attention coefficient matrix and the temporal extrapolation results. S5. In response to external input public health strategies or environmental governance parameters, directly correct the corresponding feature layer values ​​in the spatiotemporal data cube; and trigger the spatiotemporal deep learning model to perform forward inference, outputting the future respiratory health risk evolution trajectory under this input.

[0029] The specific implementation of this method includes the following steps: Multi-source spatiotemporal data acquisition and preprocessing: Collect multi-source heterogeneous data, including: electronic health record (EHR) data, environmental monitoring data, meteorological data, demographic data, and geographic information data; clean, standardize, and spatiotemporally align the collected data to form a unified spatiotemporal data cube; The electronic health record data includes: individual health information such as patient age, gender, smoking history, past medical history, lung function test results, and medical records; environmental monitoring data includes: PM2.5. 2.5 PM 10 Concentrations of air pollutants such as SO2, NO2, O3, and CO; meteorological data including climate factors such as temperature, humidity, air pressure, wind speed, and precipitation; demographic data including socio-economic indicators such as population density, age structure, and urbanization rate; and geographic information data including spatial information such as administrative boundaries, road networks, and distribution of medical institutions.

[0030] Spatiotemporal graph structure construction Constructing a multi-level spatiotemporal map structure based on geospatial relationships and functional connections: Definition of a spatial diagram: Construct an undirected weighted graph in: V: Set of nodes, |V| = N, where each node represents a geographical region (e.g., a district or county). E: Edge set, representing the spatial relationships between regions. A: Adjacency matrix, elements Multi-level adjacency matrix construction: (1) Geographic adjacency matrix A_geo: Based on geospatial adjacency relationships: in (Adjustable) (2) Functional adjacency matrix A_func: Based on population mobility intensity, this structure reflects actual connections: in: The average daily population flow from region i to region j is represented by data such as mobile phone signaling data, transportation card swipe data, and navigation data, which are normalized to the [0,1] interval.

[0031] (3) Similarity adjacency matrix A_sim: Constructed based on region feature similarity: in Feature vectors of regions i and j (population structure, economic development, medical resources, etc.) Comprehensive adjacency matrix: By weighted summation and merging of multi-level spatial relationships: Among them, weight , , satisfy: + + =1 Normalization process: Symmetric normalization of the adjacency matrix facilitates graph convolution calculations. Where D is the degree matrix.

[0032] Spatiotemporal diagram sequence: Introducing the time dimension into the graph structure forms a spatiotemporal graph sequence: The graph for each time step: The graph's topology (V, E) remains constant, while the node characteristics change dynamically over time. Multimodal feature fusion: Design a multimodal feature encoder to extract features from different types of data: For health record data, use a multilayer perceptron (MLP) encoder to extract individual risk features. Temporal convolutional network (TCN) is used to extract time series features from environmental-meteorological data. Location encoding features are extracted from geospatial data using a spatial embedding layer; an attention mechanism is employed for cross-modal feature fusion, calculating the importance weights of different modal features to generate fused node feature representations. Spatiotemporal graph neural network modeling: Construct a spatiotemporal graph neural network (ST-GNN) model that simultaneously models spatial dependencies and temporal dynamics: Model architecture: ST-GNN consists of a spatial graph convolution module and a time series modeling module: (1) Spatial Graph Convolution Module: A Graph Attention Network (GAT) layer is used to adaptively learn the importance of neighboring nodes. Single-layer GAT calculation: For node i, the update is represented as: (2) Time series modeling module: Time-dependent capture is achieved using a gated recurrent unit (GRU) or a Transformer encoder: GRU scheme: Final spatiotemporal characteristics: This scheme uses a 2-layer GRU with a hidden dimension d_hidden=128.

[0033] Disease risk prediction and dynamic simulation Risk prediction and scenario simulation based on the output of spatiotemporal graph neural networks: (1) Risk level prediction: Spatiotemporal features are mapped to risk categories using fully connected layers and Softmax: Risk level definition: Low risk: COPD incidence rate < threshold_1 (e.g., < 2%) Medium risk: Threshold 1 ≤ Incidence rate < Threshold 2 (e.g., 2%-5%) High risk: Incidence rate ≥ threshold_2 (e.g., ≥ 5%) (2) Multi-timescale prediction: Supports different forecast time spans: Short-term forecast: next week (emergency response); Medium-term forecast: next 2-4 weeks (resource allocation); Long-term forecast: next 1-3 months (strategic planning). This is achieved by adjusting the output steps of the GRU or through multi-task learning. Where τ ∈ {1, 2, 4} (3) Dynamic scenario simulation: Design a scenario simulation interface to support "hypothesis-deduction" style analysis: Scenario definition: Supported scenario types: Air quality improvement scenario: Simulated pollution control effect example: PM2.5 concentration reduced by 30%, NO2 concentration reduced by 25%; Extreme weather scenarios: assess the impact of climate change, e.g., sustained low temperatures (< 0°C) + high humidity (> 80%) for 2 weeks; Public health intervention scenarios: assess policy effectiveness, e.g., a 15% increase in smoking cessation rates and an increase in lung function screening coverage to 60%.

[0034] Interpretability analysis and visualization: Provides multi-dimensional model interpretability analysis to enhance clinical usability: (1) Visualization of attention weights: Spatial attention analysis: Extracting attention coefficients from the GAT layer Generate an inter-regional influence relationship diagram: Node: Geographic region Directed edge: attention weight (i→j) Edge weight visualization: line thickness / color depth is proportional to Identify key propagation paths: Cross-modal attention analysis: Extract the attention weights of the multimodal fusion layer and evaluate the relative importance of different data sources: (2) Feature importance analysis: Feature contributions are quantified using integrated gradients or SHAP values: (3) Visualization of the spatiotemporal distribution of risk: Spatial distribution heat map: Overlay risk levels on a GIS map, using color coding: Green: Low risk; Yellow: Medium risk; Red: High risk (4) Risk propagation path analysis: Based on graph attention weights and risk levels, identify propagation paths: Risk transmission intensity: Identify key areas requiring focused monitoring and intervention. (5) Automatic generation of decision support reports: Generate a structured report, including: Current risk assessment: risk level distribution in various regions; trend prediction: risk evolution in the next 2-4 weeks; key influencing factors: top 5 risk factors; high-risk area identification: top-10 risk level areas; intervention recommendations: environmental governance: priority control of pollutants; population screening: high-risk population identification; resource allocation: priority allocation of medical resources to key areas.

[0035] Model evaluation metrics: Classification indicators: Accuracy, Precision, Recall, F1 Score, AUC-ROC Regression Indicator (Continuous Risk Value): MAE, RMSE, MAPE Spatiotemporal indicators: Spatial Moran's I: Evaluating the Spatial Autocorrelation of Predictions Time autocorrelation coefficient: assesses the time smoothness of the prediction results. In summary, the following improvements have been made to this solution: A unified fusion method for multi-source heterogeneous spatiotemporal data: This paper proposes a multi-modal feature fusion strategy based on attention mechanism to effectively integrate multi-source heterogeneous data such as electronic health records, environmental monitoring, meteorology, population statistics, and geographic information, and achieve complementary advantages of different modalities.

[0036] Multi-level spatiotemporal graph structure construction method: Design a three-level spatial relationship modeling framework that includes geographical adjacency, functional adjacency (population flow) and similarity adjacency, to comprehensively capture the physical proximity, functional connection and feature similarity between regions.

[0037] Adaptive Spatial Dependency Modeling Based on Graph Attention Networks: The graph attention mechanism (GAT) is used to adaptively learn the influence weights of different neighboring nodes, overcoming the limitation of traditional graph convolution that treats all neighboring nodes with equal weight.

[0038] An end-to-end deep learning framework for spatiotemporal joint modeling: It organically combines spatial graph convolution (GAT) with time series modeling (GRU / Transformer) to achieve joint learning of spatial dependencies and temporal dynamics, avoiding error accumulation in two-stage modeling.

[0039] Dynamic scenario simulation capability: It supports "hypothesis-deduction" risk evolution simulation by modifying environmental factor inputs, providing quantitative assessment tools for environmental policy formulation, medical resource allocation, and public health intervention.

[0040] Multi-level interpretability analysis method: By visualizing attention weights, analyzing feature importance (SHAP value), and identifying risk propagation paths, the interpretability and credibility of model prediction results are improved, making them easier for clinical understanding and policy application.

[0041] Multi-timescale forecasting capability: Supports short-term (1 week), medium-term (2-4 weeks), and long-term (1-3 months) risk forecasting to meet different needs of emergency response, resource allocation, and strategic planning.

[0042] Example 2 In this embodiment, a respiratory system health risk situational awareness system is used, the system comprising: The multi-dimensional sensing module is used to collect multi-source heterogeneous sensing data and construct a three-dimensional spatiotemporal data cube; The situation topology construction module is used to calculate and fuse physical, interactive, and homogeneous multi-perspective adjacency relationships; The deep cognitive reasoning module has a built-in spatiotemporal deep learning model that is trained and configured to capture environmental lag effects using dilated convolution, aggregate spatial transmission risks using graph attention mechanisms, and deduce temporal trends using gated recurrent units. The simulation and inference module is used to dynamically adjust the data cube based on the intervention parameters input from the outside, and to infer and generate the future trajectory of respiratory health risk under that input.

[0043] In this embodiment, the specific implementation steps are as follows: Step 1: Multi-source spatiotemporal data acquisition and preprocessing Multi-source spatiotemporal data acquisition and preprocessing workflow as follows Figure 2 As shown, this embodiment uses a provincial-level administrative region as the research area and collects multi-source data from 2018 to 2023: the characteristic distribution of this data is as follows. Figure 3a As shown, the time resolution is as follows Figure 3bAs shown, the data quality is as follows Figure 3c As shown.

[0044] (1) Electronic health record data: obtained from the regional health information platform, including basic information (age, gender, ID number), health records (smoking history, BMI, past medical history), and medical records (outpatient diagnosis, inpatient diagnosis, pulmonary function test results) of approximately 5 million residents. After data anonymization, the data was aggregated and statistically analyzed by community / street to calculate indicators such as COPD incidence rate and proportion of high-risk groups in each region.

[0045] (2) Environmental monitoring data: Hourly data from air quality monitoring stations obtained from the ecological and environmental departments, including the concentrations of six pollutants: PM2.5, PM10, SO2, NO2, O3, and CO, totaling 120 monitoring stations. The Kriging spatial interpolation method was used to interpolate the point data to a 1km×1km grid.

[0046] (3) Meteorological data: Daily meteorological observation data were obtained from the meteorological department, including variables such as average temperature, maximum / minimum temperature, relative humidity, air pressure, wind speed, and precipitation, totaling 85 meteorological stations. The same spatial interpolation method was used to generate gridded data.

[0047] (4) Population statistics: Obtain population census data for each district and county from the statistics department, including indicators such as permanent residents, age structure (aging rate), urbanization rate, and GDP per capita. At the same time, obtain the population flow matrix after processing mobile phone signaling data.

[0048] (5) Geographic information data: Obtain data such as administrative division vector boundaries, road networks, and POIs of medical institutions from geographic information departments.

[0049] The data preprocessing process includes: filling missing values ​​for variables with a missing rate of less than 20% using the MICE multiple imputation method; identifying outliers using the 3σ principle and replacing them with the median; standardizing numerical features using Z-scores; and aggregating all data by week as the time unit and district / county as the spatial unit to form a unified spatiotemporal data cube with dimensions [T×N×D], where T=260 weeks, N=120 districts / counties, and D=35 feature dimensions.

[0050] Step 2: Construction of the Spatiotemporal Graph Structure As shown in Figures 4-6, a spatial graph structure G = (V, E) is constructed with districts and counties as nodes, where |V| = 120: (1) Geographic adjacency matrix A geo Adjacent relationships are determined based on administrative boundaries. If two districts / counties share a boundary, then A... geo[i,j] = 1, otherwise 0. Considering potential inaccuracies in boundary conditions, districts and counties less than 10km apart are also considered adjacent. The geographic association topology of the nodes is as follows: Figure 4a As shown; the matrix representation of the corresponding geographic adjacency graph is as follows. Figure 4b As shown.

[0051] (2) Functional adjacency matrix A func The daily average population flow between districts and counties is calculated using mobile signaling data, and the normalized data is used as edge weights. This matrix reflects the actual connection strength between districts and counties. The functional relationship topology of the nodes is shown below. Figure 5a As shown; the matrix representation of the corresponding functional adjacency graph is as follows Figure 5b As shown.

[0052] (3) Similarity adjacency matrix A sim Calculate the cosine similarity of the feature vectors (population structure, economic development level, medical resources, etc.) of each district / county, setting a threshold θ=0.7. Connect district / county pairs with similarity greater than the threshold. The feature similarity topology graph of the nodes is as follows: Figure 6a As shown; the matrix representation of the corresponding feature similarity topology graph is as follows: Figure 6b As shown in the image, the colors (from blue to red) correspond to the adjacency strength of the merged node pairs. Combined with the values ​​of the color bars on the right (0.0~0.7), their specific meanings are as follows: Dark blue (bottom of the color bar, corresponding value ≈ 0.0): This indicates that the overall adjacency strength between the node pairs in the corresponding row and column is extremely low, and there is almost no fusion association based on geographical, functional, or feature similarity.

[0053] Light blue / cyan (corresponding to values ​​of 0.1~0.4): indicates that there is a weak to moderate degree of comprehensive association between node pairs, which is a moderate integration result after weighting geographical adjacency, functional interaction, and feature similarity.

[0054] Orange (corresponding to a value of 0.5~0.6): indicates a high overall adjacency strength between node pairs, indicating a strong degree of integration and correlation between the two in geographical, functional, and characteristic dimensions.

[0055] Red (top of the color bar, corresponding value ≈ 0.7): indicates the highest comprehensive adjacency strength between the corresponding node pairs. It is a strong integration result of multi-dimensional associations based on geographical, functional, and feature similarities, and the overall correlation between the two is the closest.

[0056] The combined adjacency matrix is ​​obtained by weighted summation: A = 0.4A geo + 0.4A func + 0.2·A sim The weights are determined by optimization on the validation set through grid search.

[0057] Step 3: Multimodal Feature Fusion like Figure 7 As shown, three parallel feature encoders are designed: (1) Health feature encoder: A 3-layer MLP (128-64-32) is used to encode the health indicators (COPD incidence, proportion of high-risk groups, average age, smoking rate, etc.) after regional aggregation, and output a 32-dimensional feature vector H. health ∈ R^{N×32}.

[0058] (2) Environmental-meteorological feature encoder: A temporal convolutional network (TCN) is used to process the environmental-meteorological time series of the past 4 weeks. The kernel size is 3 and the dilation factor is [1,2,4]. The output is a 32-dimensional feature vector H. env ∈ R^{N×32}.

[0059] (3) Spatial location encoder: A learnable location embedding layer is used to encode the geographical coordinates (longitude and latitude) of the districts and counties into a 32-dimensional feature vector H. spatial ∈ R^{N×32}.

[0060] A multi-head cross-attention mechanism is used for feature fusion, and the importance weights of each modality feature are calculated and then summed in a weighted manner. H = α1·H health + α2H env + α3·H spatial The attention weights [α1, α2, α3] are obtained by Softmax normalization, and the fused features H ∈ R^{N×32}.

[0061] Step 4: Spatiotemporal Graph Neural Network Modeling like Figure 8 and Figures 9a-9c As shown, the spatiotemporal graph neural network model ST-GNN is constructed: Spatiotemporal graph neural network model architecture as follows Figure 8 As shown, the graph attention calculation node i and its neighboring nodes are as follows: Figure 9a As shown, bullish attention is as follows Figure 9b As shown, the attention weights of each neighboring node are adjusted as follows: Figure 9c As shown.

[0062] Figure 9cIn the diagram, red corresponds to attention head 1, and the red arrows represent the attentional connections between the central node and surrounding nodes under attention head 1. Cyan corresponds to attention head 2, and the cyan arrows represent the attentional connections corresponding to attention head 2. Yellow corresponds to attention head 3, and the yellow arrows represent the attentional connection paths under attention head 3. Light green corresponds to attention head 4, and the light green arrows represent the attentional connections corresponding to attention head 4.

[0063] (1) Spatial graph convolutional layer: A 2-layer graph attention network (GAT) is used, with 4 attention heads per layer and a hidden dimension of 64.

[0064] Multi-head attention outputs are aggregated through splicing or averaging. Each layer of GAT is followed by residual connections and layer normalization.

[0065] (2) Time series modeling layer: A 2-layer GRU is used with a hidden state dimension of 128. The input is the spatial graph convolution output sequence of the past 12 weeks {H^{(t-11)}, H^{(t-10)}, ..., H^{(t)}}, and the output is the feature representation H_st ∈ R^{N×128} that integrates spatiotemporal information.

[0066] The model training uses the cross-entropy loss function: Where c represents the risk level category (low / medium / high), and a spatiotemporal smoothing regularization term is also included: The Adam optimizer is used, with an initial learning rate of 0.001, cosine annealing scheduling, a batch size of 32, a maximum number of training epochs of 200, and an early stopping patience value of 20.

[0067] Step 5: Disease Risk Prediction and Dynamic Simulation like Figure 9d As shown, risk prediction and dynamic simulation are performed based on the output of the spatiotemporal graph neural network: (1) Risk level prediction: Input H_st into the fully connected layer (128-64-3) and the Softmax layer, and output the probability distribution of COPD risk level P ∈ R^{N×3} for each district and county, corresponding to three risk levels: low / medium / high.

[0068] (2) Multi-timescale prediction: By adjusting the output steps of GRU, risk prediction for the next 1 week, 2 weeks and 4 weeks can be achieved, supporting short-term early warning and medium-term trend analysis.

[0069] (3) Dynamic Scenario Simulation: A scenario simulation interface is designed to allow users to modify the environmental parameters input for future time periods (such as assuming a 20% increase in PM2.5 concentration or a 5°C decrease in temperature). The model then performs forward inference again and outputs the risk prediction results under that scenario. Supported scenario types include: air quality deterioration scenarios, extreme weather scenarios, and public health intervention scenarios.

[0070] Step Six: Interpretability Analysis and Visualization like Figure 10a As shown, the specific process for providing multi-dimensional interpretability analysis is as follows: (1) Attention weight visualization: extract the attention weights of the multimodal fusion layer and the graph attention layer, and generate a heatmap of feature importance ranking and inter-region influence relationship.

[0071] (2) Risk Spatiotemporal Distribution Map: Overlay the risk levels of each district and county on the GIS map to generate a spatiotemporal distribution heat map, which supports time series animation playback and intuitively displays the spatiotemporal evolution of risk.

[0072] (3) Risk transmission path analysis: Based on graph attention weight, identify the key paths and core node areas of risk transmission to provide a basis for precise prevention and control.

[0073] (4) Decision support report: Automatically generates analysis reports containing risk prediction results, key influencing factors, and recommended measures to assist public health decision-making.

[0074] Risk evolution in different scenarios, such as Figure 10b As shown. Light green background: represents the low-risk range, covering the range of 0.0 to 0.3 on the vertical axis, which is the lowest risk range.

[0075] Light yellow background: Represents the medium risk level range, covering the range of 0.3 to 0.6 on the vertical axis, which is a range of moderate risk.

[0076] Pink background: Represents the high-risk range, covering the range of 0.6 to 0.9 on the vertical axis, which is the highest risk range.

[0077] 2. Curve colors: Corresponding to different scenarios The blue curve represents the risk evolution trend under the baseline scenario, reflecting the risk changes in the absence of intervention / normal state.

[0078] Green Curve: Risk evolution trend under Scenario 1 (Air Quality Improvement 1), reflecting the risk changes after air quality improvement.

[0079] The red curve represents the risk evolution trend under scenario 2 (extreme weather), reflecting the risk changes under extreme weather scenarios.

[0080] In experimental validation, the dataset was divided chronologically into a training set (first 80%), a validation set (10%), and a test set (last 10%). Our method was compared with several baseline methods, including logistic regression (LR), random forest (RF), XGBoost, LSTM, and standard GCN. Experimental results show that our method achieves state-of-the-art performance on the risk prediction task, with an accuracy of 87.3% and an AUC of 0.912, representing improvements of 5.2% and 4.8% respectively compared to the best baseline methods. Ablation experiments demonstrate the effectiveness of key modules such as multi-level spatial graph structure, multimodal feature fusion, and spatiotemporal joint modeling.

[0081] A heatmap showing the spatiotemporal distribution of COPD risk in spring is shown below. Figure 11a As shown; a heatmap visualizing the spatiotemporal distribution of COPD risk in summer is shown below. Figure 11b As shown; a visual heatmap of the spatiotemporal distribution of COPD risk in autumn is shown below. Figure 11c As shown; a visual heatmap of the spatiotemporal distribution of COPD risk in winter is shown below. Figure 11d As shown; In the diagram: Green areas (0.0~0.2 range): correspond to areas with lower risk values. Light green dots / areas in the diagram represent areas with weaker risk levels.

[0082] Yellow / orange (0.4~0.8 range): corresponds to medium risk area, overall scene matching. The risk value of the yellow dotted line and orange area in the figure is at the middle level (close to the average risk value of 0.6).

[0083] Red areas (0.8~1.0 range): These correspond to areas with higher risk values. The red areas and red pentagrams in the diagram indicate that the spatial location has a higher degree of risk.

[0084] Overall, the color gradient from green to red corresponds to an increase in risk value from low to high, intuitively presenting the distribution of risk intensity in different spatial locations under this spring scenario.

[0085] In summary, the following improvements have been made to this solution: Through multi-source data fusion and spatiotemporal joint modeling, this system achieved significantly better performance than traditional methods in the COPD risk prediction task: accuracy reached 87.3%, an improvement of 5.2% over the best baseline method (XGBoost); AUC-ROC reached 0.912, an improvement of 4.8% over the best baseline method; the recall rate for identifying high-risk areas reached 91.5%, and the false positive rate was less than 8%. The multi-level spatial graph structure and graph attention network effectively captured the spatial dependencies between regions: accurately identified the transboundary transmission effect of environmental pollution (average impact distance 30-50km); discovered disease risk transmission pathways driven by population movement (inter-regional correlation coefficient r=0.68 for commuter flow >10,000 people / day); captured the seasonal variation pattern of COPD risk (42% higher risk in winter than in summer); and the dynamic simulation function supported the quantitative assessment of policy effects: under the simulated scenario of a 30% reduction in PM2.5 concentration, the number of high-risk areas decreased by 23%, and the number of medium-risk areas decreased by 18%; the intervention measure of increasing the smoking cessation rate by 15% was evaluated, which could reduce the incidence of COPD by 12% within 5 years; and the top 3 intervention measures that contributed the most to risk reduction (air quality improvement, smoking control, and screening of high-risk groups) were identified.

[0086] Attention mechanism and feature importance analysis provided model transparency: the top 5 factors influencing COPD risk were identified: PM2.5 concentration (contribution 28%), age structure (23%), smoking rate (19%), temperature (15%), and humidity (10%). Optimized model architecture and training strategies ensure practicality: single prediction time < 2 seconds (120 regions); model training time < 4 hours (GPU: NVIDIA RTX 3090); supports online incremental learning, with model update time of < 10 minutes for adding 1 week of data; The model demonstrated good generalization ability in cross-regional and cross-time tests: the accuracy remained above 82% on the test set of adjacent provinces (decreased by only 5.3%); the prediction accuracy remained above 80% on data outside the training period (the next 6 months); and the prediction response to extreme weather events (such as cold waves and smog) was timely, issuing warnings 3-7 days in advance. This system can be directly applied to multiple public health scenarios: Disease surveillance and early warning: Establish an early warning system for COPD to predict high incidence risks 1-4 weeks in advance; Medical resource allocation: guiding the dynamic allocation of emergency beds, respiratory therapy equipment, and medical staff; Environmental policy assessment: Quantitatively assess the health benefits of air pollution control policies to provide a scientific basis for government decision-making; High-risk population screening: accurately identify areas and populations that require key screening and intervention to improve screening efficiency; Health education and promotion: Based on risk prediction results, conduct targeted health education in high-risk periods and areas; Compared to traditional large-scale population screening, this solution significantly reduces costs: screening costs are reduced by approximately 40% by accurately identifying high-risk areas and time periods; early warning and timely intervention can reduce hospitalization costs caused by acute exacerbations of COPD; and the allocation of medical resources is optimized, improving resource utilization efficiency by more than 20%.

[0087] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for regional respiratory health risk situational awareness, characterized in that, The method steps include: S1. Collect multi-source heterogeneous sensing data of the target area, and perform spatiotemporal alignment and gridding mapping on the multi-source heterogeneous sensing data to construct a three-dimensional spatiotemporal data cube; S2. Based on multi-source heterogeneous sensing data, construct physical adjacency topology, interactive association topology, and homogeneous association topology, and integrate the above three topologies to generate a comprehensive spatial situation matrix. S3. Based on multi-source heterogeneous sensing data, a multimodal feature encoder is used to extract multimodal features and fuse them to obtain the fused node feature representation; S4. Construct a spatiotemporal deep learning model that includes a spatial attention aggregation mechanism and a temporal gating loop mechanism. In the model, the spatial attention coefficient matrix is ​​dynamically calculated based on the comprehensive spatial situation matrix using the spatial attention aggregation mechanism. The temporal gating loop mechanism is used to perform temporal extrapolation on the fused node feature representation. A future respiratory health risk situation map is generated based on the spatial attention coefficient matrix and the temporal extrapolation results. S5. In response to externally inputted public health strategies or environmental governance parameters, directly correct the corresponding feature layer values ​​in the spatiotemporal data cube; and trigger the spatiotemporal deep learning model to perform forward inference, outputting the future respiratory health risk evolution trajectory under this input.

2. The method for regional respiratory health risk situational awareness according to claim 1, characterized in that, The multi-source heterogeneous sensing data in S1 includes desensitized clinical statistical characteristics, environmental meteorological monitoring data, regional population attribute parameters, and geospatial data. The desensitized clinical statistical features include: average lung function indicators within the grid area, frequency of respiratory symptom search terms, and density of historical confirmed cases; The environmental meteorological monitoring data includes: time series of fine particulate matter, inhalable particulate matter, nitrogen dioxide concentration, and relative humidity. The regional population attribute parameters include: population aging coefficient, residential density index, and regional functional zoning type code; The three-dimensional spatiotemporal data cube is orthogonally composed of a spatial domain, a temporal domain, and a feature domain, and is used to fully digitize the regional health environment status.

3. The method for regional respiratory health risk situation perception according to claim 2, characterized in that, In S2, when constructing the physical adjacency topology, interactive association topology, and homogeneous association topology, a data-driven approach is adopted. The physical adjacency topology is constructed based on geographical boundary connections; the interactive association topology is constructed based on population flow intensity derived from mobile signaling data; and the homogeneous association topology is constructed based on the vector similarity of regional socioeconomic indicators. The specific process of S2 includes: Initialize scalar weights for the three matrices: physical adjacency, interaction association, and homogeneous association. During end-to-end training of the model, the scalar weights of the three matrices are updated through backpropagation of the loss function. The three matrices are weighted and superimposed using the updated scalar weights.

4. The method for regional respiratory health risk situation perception according to claim 3, characterized in that, The specific process of S3 includes: Three parallel feature encoding branches are constructed, and the output dimension of each branch is uniformly mapped to the preset hidden layer dimension; In the first branch, the desensitized clinical statistical features are input into the multilayer perceptron encoder, which is configured as a three-layer fully connected structure with the number of neurons decreasing sequentially to obtain a health feature vector. In the second branch, environmental meteorological monitoring data containing 15 features and spanning 4 weeks is input into the temporal convolutional network encoder. After convolutional extraction of the time series, the environmental feature vector is output. In the third branch, the longitude and latitude coordinates of the grid center are input into the position encoder to obtain the spatial feature vector; The health feature vector, environmental feature vector, and spatial feature vector are simultaneously input into the cross-attention layer; The input feature vectors are mapped to query vectors, key vectors, and value vectors respectively; attention weight coefficients are calculated using the Softmax function, and the weight coefficients are used to perform a weighted summation of each feature vector to output the fused node feature representation.

5. The method for regional respiratory health risk situation perception according to claim 4, characterized in that, The temporal convolutional network encoder in the second branch adopts a hierarchical causal convolutional structure, and the dilation rate of the convolutional kernel is set to increase exponentially with the increase of the number of network layers.

6. The method for regional respiratory health risk situation perception according to claim 1, characterized in that, In S4, the specific execution logic of the spatial attention aggregation mechanism is as follows: Based on the comprehensive spatial situation matrix, the correlation scores between the central grid and its neighboring grids in the feature space are calculated. The relevance scores are mapped to normalized attention coefficients through non-linear activation; Based on the attention coefficients, the neighborhood features are weighted and summed to obtain the spatial attention coefficient matrix.

7. The method for regional respiratory health risk situation perception according to claim 1, characterized in that, In S4, the specific execution logic of the timing-gated loop mechanism is as follows: Receive the fused node features and the hidden state vector output from the previous time step; The feature vector at the current time step is concatenated with the hidden state vector output at the previous time step. Linear transformation and Sigmoid activation operations are then performed on the reset gate weight matrix and update gate weight matrix, respectively, to generate the reset gate value and update gate value. The reset gate value is multiplied element-wise with the hidden state vector output from the previous time step. The result of the multiplication is concatenated with the feature vector at the current time step. Then, the current candidate hidden state is generated by processing the candidate state weight matrix and the hyperbolic tangent activation function. Using the update gate value as a weighting coefficient, the hidden state vector output at the previous time step is weighted and summed with the current candidate hidden state to obtain the hidden state vector at the current time step. The generated hidden state vector at the current time step is used as the input for the next time step and iterated until the calculation of the preset time window is completed. The final hidden state vector is then input into the fully connected layer for dimension mapping to obtain the risk level probability value of each grid region, thus obtaining the time series simulation result.

8. The method for regional respiratory health risk situation perception according to claim 1, characterized in that, The specific process of S5 includes: Based on externally input public health strategies or environmental governance parameters, construct feature correction vectors; The feature correction vector is superimposed onto the corresponding time slice of the spatiotemporal data cube to construct a virtual situation space; By comparing the output results of the model on the original data cube and the virtual situation space, the magnitude of risk reduction and the time delay in the effectiveness of intervention measures are quantified.

9. The method for regional respiratory health risk situation perception according to claim 1, characterized in that, The method further includes: Extract the spatial attention coefficient matrix generated during model inference; Based on the spatial attention coefficient matrix, high-weight grid areas are connected and mapped onto a geographic map to obtain a risk propagation flow diagram.

10. A regional respiratory health risk situational awareness system, characterized in that, The system operates as a regional respiratory health risk situational awareness system as described in any one of claims 1-9, the system comprising: The multi-dimensional sensing module is used to collect multi-source heterogeneous sensing data and construct a three-dimensional spatiotemporal data cube; The situation topology construction module is used to calculate and fuse physical, interactive, and homogeneous multi-perspective adjacency relationships; The deep cognitive reasoning module has a built-in spatiotemporal deep learning model that is trained and configured to capture environmental lag effects using dilated convolution, aggregate spatial transmission risks using graph attention mechanisms, and infer temporal trends using gated recurrent units. The simulation and inference module is used to dynamically adjust the data cube based on the intervention parameters input from the outside, and to infer and generate the future trajectory of respiratory health risk under that input.

Images