A public health level estimation method based on prior knowledge and artificial intelligence

Abstract

The application discloses a public health level estimation method based on prior knowledge and artificial intelligence, relates to the fields of public health, general artificial intelligence and space-time big data, and comprises the following steps: preprocessing multi-source heterogeneous data, generating a multi-channel space-time tensor, and extracting corresponding geographical numerical feature vectors; converting the geographical numerical feature vectors into natural language thinking chains and encoding the natural language thinking chains into prior knowledge vectors; inputting the prior knowledge vectors and the geographical numerical feature vectors into a public health evaluation intelligent agent to generate expert feature vectors; constructing a public health level estimation model; adopting a space-time block verification strategy to perform basic pre-training and knowledge injection fine-tuning on the public health level estimation model; inputting the multi-channel space-time tensor of a region to be estimated into the trained public health level estimation model to obtain an optimized estimation result of the public health level. The application provides decision support and data support for the related fields of public health and social governance.

Description

Technical Field

[0001] This invention relates to the fields of public health, general artificial intelligence, and spatiotemporal big data. Specifically, it relates to a method for estimating public health levels based on prior knowledge and artificial intelligence. Background Technology

[0002] In the field of public health, accurate and high-resolution health surveillance data is the cornerstone of policy-making and resource allocation. Estimating health data (such as cardiovascular disease mortality) at a 100m x 100m grid has extremely high scientific significance and application value.

[0003] Traditional health data is often aggregated at the administrative level (such as streets or districts), which contains an ecological fallacy. An area that appears healthy overall may conceal extreme health deficiencies (such as urban villages next to chemical plants).

[0004] High-precision grid data can reveal these hidden health inequalities. Governments can then use this data for "surgical" interventions, such as building more community hospitals in specific high-risk grid areas, improving green spaces, or implementing stricter environmental regulations, thereby significantly improving the efficiency of public health spending.

[0005] Studying the health effects of environmental factors (air pollution, urban design) often requires expensive cohort studies and is difficult to isolate confounding factors.

[0006] Generative estimation methods based on big data and AI are equivalent to building a health model of a digital twin city. Researchers can use this model to perform low-cost counterfactual reasoning, such as simulating "how much would cardiovascular mortality decrease if the green space ratio in this area increased by 10%", providing a scientific basis for urban planning.

[0007] In emergency situations such as extreme heat or infectious disease outbreaks, it can quickly identify potential areas with high prevalence of underlying diseases and areas with high concentrations of vulnerable populations (such as grids with severe aging populations and high building density), thereby guiding the priority deployment of medical rescue forces.

[0008] Traditionally, estimating the level of public health in a region has relied primarily on Pathway 1 and Pathway 2.

[0009] The first approach is based on statistical interpolation from monitoring stations. It utilizes limited data from disease surveillance sites (such as sentinel hospitals) and employs geostatistical methods such as Kriging interpolation or inverse distance weighting (IDW) to extrapolate values ​​for unknown areas. Its advantages include simple calculation and ease of understanding.

[0010] The drawbacks of Path One are twofold. First, it suffers from a severe smoothing effect, assuming that health conditions change smoothly and gradually in space, failing to capture abrupt changes caused by factors such as road barriers and factory emissions. Second, it ignores environmental heterogeneity, utilizing only spatial distance information while neglecting the influence of key pathogenic factors such as population, environment, and socioeconomic factors.

[0011] Pathway Two is the classic Land Use Regression (LUR) model. Pathway Two establishes a multiple linear regression model: Health = β0 + β1 (population) + β2 (road network density) + β3 (green space ratio) + ϵ. The advantage of Pathway Two is that it incorporates environmental covariates, making it more scientific than simple interpolation and offering a degree of interpretability.

[0012] Pathway Two suffers from the limitation of linear assumptions, which presuppose a simple linear relationship between environmental factors and health. However, reality is often non-linear (for example, moderate green spaces are beneficial, but excessively dense forests may breed mosquitoes). Pathway Two also struggles with high-dimensional heterogeneous data. Traditional regression models are ineffective for complex unstructured features in remote sensing imagery, such as textures and semantic combinations of points of interest (POIs), often requiring manually designed features, which is inefficient and prone to missing crucial information. Furthermore, Pathway Two lacks spatiotemporal dynamism, typically being a static model that fails to capture the complex dynamic processes of health risk evolution over time and spatial interactions (neighborhood effects).

[0013] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0014] In response to the problems in related technologies, this invention proposes a public health level estimation method based on prior knowledge and artificial intelligence to overcome the aforementioned technical problems existing in existing related technologies.

[0015] Therefore, the specific technical solution adopted by the present invention is as follows: This invention provides a method for estimating public health levels based on prior knowledge and artificial intelligence, the method comprising: S1. Using a pre-defined spatial ETL pipeline, preprocess multi-source heterogeneous data to generate multi-channel spatiotemporal tensors of standard resolution grids, and extract the geographic numerical feature vectors corresponding to each standard resolution grid. S2. Based on the preset dynamic semantic mapping rules, the geographic numerical feature vector is transformed into a natural language thought chain; using a text embedding model, the natural language thought chain is encoded into a prior knowledge vector. S3. Input the prior knowledge vector and geographic numerical feature vector into the preset public health assessment agent, use the public health assessment agent to search the epidemiological knowledge base for macro-reasoning, and generate expert feature vectors. S4. Construct a public health level estimation model; input multi-channel spatiotemporal tensors into the public health level estimation model, generate spatiotemporal feature vectors through feature extraction and spatiotemporal modeling, and dynamically fuse the spatiotemporal feature vectors with expert feature vectors using a gating fusion mechanism to output the estimation result of the public health level; S5. A spatiotemporal block verification strategy is adopted to perform basic pre-training and knowledge injection fine-tuning on the public health level estimation model. Among them, the knowledge injection fine-tuning uses prior knowledge vectors and expert feature vectors as supervision signals, and optimizes the parameters of the public health level estimation model through loss function constraints to obtain the trained public health level estimation model. S6. Input the multi-channel spatiotemporal tensor of the region to be estimated into the trained public health level estimation model to obtain the optimized public health level estimation result.

[0016] The beneficial effects of this invention are as follows: 1. This invention improves prediction accuracy. Compared to traditional linear models, it can capture complex nonlinear environment-health relationships, thereby generating more accurate 100-meter-level health risk maps.

[0017] 2. This invention enhances interpretability. Traditional deep learning is a black box. This invention not only outputs predicted values, but the agent can also simultaneously output a chain of reasoning in natural language, greatly enhancing the decision-maker's trust.

[0018] 3. This invention enables the accumulation and reuse of knowledge. As this invention operates, the intelligent agent can deposit successful reasoning patterns into a knowledge base, continuously evolving and becoming smarter with each use. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart according to an embodiment of the present invention; Figure 2 This is one of the technical roadmaps according to embodiments of the present invention; Figure 3 This is the second technical roadmap according to an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0022] According to an embodiment of the present invention, a method for estimating public health levels based on prior knowledge and artificial intelligence is provided.

[0023] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-3 As shown, according to an embodiment of the present invention, a public health level estimation method based on prior knowledge and artificial intelligence includes: S1. Using a pre-defined spatial ETL pipeline, preprocess multi-source heterogeneous data to generate multi-channel spatiotemporal tensors of standard resolution grids, and extract the geographic numerical feature vectors corresponding to each standard resolution grid. In this optional embodiment, the multi-source heterogeneous data is preprocessed using a preset spatial ETL pipeline to generate a multi-channel spatiotemporal tensor of a standard resolution grid, and the geographic numerical feature vector corresponding to each standard resolution grid is extracted, including the following steps: S11. Read the original coordinate system of all multi-source heterogeneous data, select the appropriate projection coordinate system according to the study area, and reproject all data to the appropriate projection coordinate system. S12. Calculate the boundary range of the study area, divide the grid according to the preset grid resolution, and generate a grid index; S13. Perform rasterization on the vector data to obtain vector rasterized features; S14. Virtually slice the remote sensing image based on affine transformation, and extract the corresponding pixel region as visual feature according to the grid boundary. S15. Map the point data to the grid through spatial connections and perform feature aggregation to obtain the point data aggregated features; S16. Align the vector raster features, visual features, and point data aggregation features according to the grid index and time dimension, combine them into a multi-channel spatiotemporal tensor, and normalize the numerical features of each grid to generate the geographic numerical feature vector corresponding to each grid.

[0024] In this optional embodiment, the vector rasterization features include building volume features and road network density features; wherein, the building volume features are obtained by calculating and aggregating the building base area and the intersection area within the grid; the road network density features are obtained by calculating the effective length of roads within the grid using a line segment clipping algorithm.

[0025] It should be explained that the ETL (Extract, Transform, Load) process maps all data to a single spatial reference system and grid index. Inputting multi-source heterogeneous data in formats including remote sensing imagery, vector road networks, POIs, and population raster (CSV, Geo TIFF, and vector types), a unified spatial reference of 100m×100m grid in a projected coordinate system (such as UTM) is used. The final output is a standardized feature tensor. The specific process is as follows: I. Automatic unification and projection of spatial reference systems.

[0026] Read the coordinate reference system (CRS) of all input metadata. Since latitude and longitude (WGS84) cannot be directly measured at the meter level, determine the longitude of the region center, select a suitable projection coordinate system (such as UTM projection), and reproject all data.

[0027] Mathematical principle (coordinate transformation): Let the original coordinates be... (Latitude, Longitude), Target Projected Coordinates are (meters). Establish the transformation function. : The Gauss-Krüger or UTM projection formula (approximate expansion) is usually used: ; ; In the formula, Input the geodetic latitude and longitude; The longitude of the central meridian (automatically calculated by the agent based on the longitude range of the input data): ; The scaling factor (UTM is 0.9996). The coordinates are the converted Cartesian coordinates (unit: meters).

[0028] II. Standard Grid Indexing: Calculate the bounding box covering the study area and generate a grid ID matrix with a step size of 100m.

[0029] 1. Boundary calculation formula: ; 2. Grid row and column calculation: For any spatial point Its grid index for: .

[0030] In the formula, The grid resolution is set to 100 (meters) here. These are the row and column numbers of the matrix. Let x be the x-value of all Cartesian coordinates in the plane. This array contains the y-values ​​of all Cartesian coordinates. `min()` returns the minimum value of the array. `max()` returns the maximum value of the array.

[0031] III. Vector-to-Grid Conversion.

[0032] A. Building Data Processing (Volume Calculation).

[0033] The input is the sequence of projected plane corner points for each building. and height .

[0034] 1. Polygon construction involves closing a sequence of corner points to form a polygon.

[0035] 2. The area is calculated using the shoelace formula to determine the base area. : .

[0036] In the formula, This refers to the calculated building base area. This represents the total number of vertices (i.e., the number of corner points) of the polygon (building base). The vertices are indexed from 1 to... Traverse sequentially. ) is the first The planar projected coordinates of each vertex (e.g., longitude / latitude or X / Y values ​​in a projected coordinate system). Let the coordinates of the next vertex be arranged in order, when (i.e., the last point) It refers to the first point. This means that the polygon must be closed. To use the summation sign, the cross product results of all adjacent vertex pairs are accumulated. The absolute value sign ensures that the final calculated area is positive (because the internal summation result may be negative depending on whether the vertices are arranged clockwise or counterclockwise).

[0037] 3. Grid Cutting: If a building spans multiple grids, calculate its value in each grid. Intersecting area within: Area (Poly) ).

[0038] 4. Feature Aggregation: Calculate the building volume features within the grid. : .

[0039] In the formula, For grid The final aggregated building volume feature value reflects the building density and spatial congestion within the grid. The row and column indices of the target grid in the entire study area matrix (i.e., the unique coordinate identifier of the grid). To represent the summation symbol, all buildings within the grid are summed. To be compatible with grid The collection of all buildings whose spatial extents overlap. For set Each individual building in the complex. For buildings Base polygons and mesh The geometric intersection area of ​​the boundary is calculated only for the portion of the building that falls within the grid boundary, rather than the total base area of ​​the building, to ensure the accuracy of spatial statistics. For buildings The height (usually in meters). Therefore, multiplication is used to obtain the local volume of the building within the grid by multiplying the intersection area by the height.

[0040] B. Road network data processing (Line Clipping).

[0041] Input: Road centerline segment And the intermediate turning point.

[0042] Using the Cohen-Sutherland clipping algorithm, the grid for each road is calculated. The cutoff length within. Let the mesh boundary be... For line segments Solve The parameters are obtained from the intersection points with the four boundaries. and .

[0043] 1. Effective length : .

[0044] In the formula, (Also written in the text) ( ) represents a specific road segment L that, after being cut, falls onto the grid. The effective cutoff length inside is the portion of the road that is actually contributed to the grid. Parametric equations of road segments , where t is a scalar parameter between 0 and 1; These are the parameter values ​​when the road enters the grid boundary; These are the parameter values ​​when the road leaves the grid boundary; The coordinates of the entry point obtained after substituting the parameters. That is, the first intersection point between the road and the grid boundary; The coordinates of the departure point obtained after substituting the parameters. That is, the second intersection of the road and the grid boundary; for Norm (Euclidean distance), which is used to calculate the entry point and departure point The straight-line distance between them.

[0045] 2. Final characteristics: .

[0046] In the formula, (Rho road) represents the final road network density feature of this grid. In physics, density usually refers to the length of a road per unit area. or ). To sum up, we add up the effective lengths of all different road segments that pass through the grid to get the total length of the roads within the grid. The square of the side length represents the base area of ​​the grid.

[0047] IV. Image Tiling.

[0048] Input: Geo TIFF imagery, including affine transformation matrices .

[0049] Instead of physical cutting (to avoid generating a large number of small files), a virtual view is created, and pixel coordinates are calculated from the grid coordinates.

[0050] Mathematical principle (inverse affine transformation): from projected coordinates Mapped to image pixel rows and columns : ; ; In the formula, The spatial resolution (pixel size) of the image. The coordinates are the top left corner of the image.

[0051] For each grid According to its geographical boundaries Calculate the corresponding pixel window Read the pixel matrix as a tensor .

[0052] V. Spatial Join.

[0053] Inputs: Health data, population data, POIs .

[0054] Label all point data directly using the grid index formula. Aggregation logic: .

[0055] In the formula, Data i,j This is the final calculated grid feature value. It's the result of aggregation for a specific grid cell (coordinate index: row i, column j). For example, it could be the total population within the grid, the average health index, or the number of POIs. {...}: The concept of a set. Here, it refers to the set of all data points that meet specific conditions. d is a single original data point. It represents a record in the original dataset, such as a specific POI (like a restaurant), a census point, or a health record. | represents the mathematical set notation **Such that**. Its left side is the element d of the set, and its right side is the filtering condition that these elements must meet. GridID(dx, dy) is the grid index mapping function. This function maps the spatial coordinates of data point d... Convert to the corresponding grid number. Dx and dy are the spatial coordinates of data point d, i.e., the longitude and latitude (or the X and Y values ​​in the projected coordinate system) of that point. These are the index coordinates of the target grid. i represents the row number, and j represents the column number.

[0056] Equations in the formula The meaning is to filter out only those that fall within this specific grid. Data points within the range. For population / health data, AggFunc = Sum (total) or Mean (mean). For POIs, AggFunc = Count (count by type) or KDE kernel density calculation.

[0057] The formula for calculating KDE kernel density is as follows: For n two-dimensional data points, the set is: , where each point It is a two-dimensional column vector.

[0058] any point on a two-dimensional plane Kernel density estimate at [location] It can be represented as: .

[0059] In the formula, These are the positions of n grid points; n is the sample size and the number of grid points. The kernel function defines the distribution at each grid point. The shape and weighting of the surrounding data distribution; Bandwidth is a scalar value that controls the spread or width of the kernel function; T is the transpose symbol in matrix calculations.

[0060] To implement an automated pipeline architecture, a DAG (Directed Acyclic Graph) task queue needs to be maintained.

[0061] Task 1: Data Self-Check. Scan all input files. Check if the projected coordinate systems of all data are consistent. If not, perform reprojection and coordinate unification. Check for NaN values ​​or coordinate anomalies (e.g., Lat > 90).

[0062] Task 2: Mesh Generator. Calculate the global spatial extent and generate a standard mesh spatial dataset.

[0063] Task 3: Parallel Processing. Worker A: Processes data such as buildings and roads. Worker B: Processes rasterized remote sensing imagery. Worker C: Processes CSV table data.

[0064] Task 4: Tensor Assembly. Merge the results from Workers A / B / C using (i,j,Year) as the primary key. Fill in missing values ​​(e.g., set a grid with no buildings to 0). Save as a .npy file (NumPy format) or HDF5 file for use in training the public health level estimation model.

[0065] S2. Based on the preset dynamic semantic mapping rules, the geographic numerical feature vector is transformed into a natural language thought chain; using a text embedding model, the natural language thought chain is encoded into a prior knowledge vector. In this optional embodiment, based on preset dynamic semantic mapping rules, geographic numerical feature vectors are transformed into natural language thought chains; using a text embedding model, the natural language thought chains are encoded into prior knowledge vectors, including the following steps: S21. According to the preset dynamic semantic mapping rules, map each numerical feature in the geographic numerical feature vector to the corresponding natural language description. S22. Based on the analytical logic of environmental epidemiology and the type of area to which the grid belongs, the natural language description is constructed into a natural language thought chain; S23. Input the natural language thought chain into the pre-trained text embedding model for encoding, and output it as a prior knowledge vector.

[0066] It needs to be explained that the process involves translating dry geographical data (such as coordinates and density) into semantic scenarios that LLM can understand, guiding LLM to perform epidemiological reasoning, and ultimately outputting a prior knowledge vector that can be absorbed by the neural network. ).

[0067] The core design principle of this step lies in constructing a Structured CoT (CoT). To maximize the reinforcing effect of the agent (i.e., the public health assessment agent) on the predictive model (i.e., the public health level estimation model), the instructions cannot be open-ended chatter; they must follow the analytical logic of environmental epidemiology: Exposure identification - Pathway analysis - Receptor sensitivity assessment - Hypothesis. The instruction design is divided into three parts: Context (context injection), Instruction (inference instructions), and Output Protocol. The construction of the prior knowledge vector includes the following steps: I. Context Semantic Mapping.

[0068] LLM is insensitive to purely numerical values ​​such as NDVI=0.13 or building density=0.8. A semantic mapper needs to be built to discretize continuous numerical values ​​into natural language descriptions. A mapping function needs to be defined. (Language space). Building density mapping is as follows: ; Instead of inputting all POI coordinates, the POI coordinates are aggregated into the dominant function area. (For example, if a chemical plant has a high weight, it is marked as an industrial-dominant area even if there is only one).

[0069] II. Dynamic Template Construction.

[0070] The agent dynamically assembles instructions based on the data in each grid.

[0071] The instruction structure template is as follows: "# Role You are a space epidemiologist. Your goal is to provide a "risk prior vector" for a neural network that predicts cardiovascular disease mortality.

[0072] 1. Environmental Background (Automatically Generated) - Location: Within the grid of [{lat}, {lon}], Year: {year}.

[0073] - City morphology: {Desc_build} (volume density: {val_vol}).

[0074] - Green infrastructure: {Desc_ndvi} (NDVI vegetation index: {val_ndvi}).

[0075] - Key stressors: - Distance from main road: {dist_road} meters ({Desc_road_noise}).

[0076] - Main POI type: {list_pois} (potential contamination / stress sources).

[0077] - Demographics: Population density is {val_pop}. {Desc_aging} (inferred from historical data).

[0078] 2. Knowledge Retrieval (RAG) (The system inserts a relevant fragment from the knowledge base here, such as: "Studies show that traffic noise > 60dB is associated with a 5% increase in the incidence of hypertension.") 3. Reasoning Task (Thinking Chain) Analyze the interactions between the above factors.

[0079] - Step 1 (Air Quality): Do building density and pollution sources (POI / road) combine to cause pollutant retention? - Step 2 (Lifestyle): Does a lack of green space mean a low level of physical activity? - Step 3 (Noise): Is the noise exposure likely to be high? 4. Output Requirements Generate a JSON object that contains a text digest for embedding.

[0080] III. Output Protocol & Embedding.

[0081] This is the key interface for the agent. The output of LLM must include a field specifically for vectorization.

[0082] The output is defined as follows: { "Reasoning Chain": "Although the green cover is moderate, this grid is adjacent to the highway and located downwind of the chemical industrial park..." Risk Level: High Confidence level: 0.85 "Embedded text": "Due to the combined effects of traffic noise exposure and industrial pollution accumulation in poorly ventilated canyon areas, the cardiovascular risk is high." } The system extracts the embedding_text field and calls a text embedding model (such as OpenAI text-embedding-3-small or the local BERT): .this This refers to the input vector in step two. The embedding-text must be highly condensed and have causal logic so that the neural network's attention mechanism can capture keywords (such as industrial pollution, traffic noise).

[0083] IV. Prompt Optimization Strategy.

[0084] Design scenario-adaptive instructions. Before constructing the instructions, use simple rules to determine the grid type and select different instruction variants. Variant A (urban center type) emphasizes building density, heat island effect, noise, and living pressure; the instruction needs to be supplemented with "focusing on the impact of urban heat island effect on the aging population." Variant B (industrial edge type) emphasizes point source pollution (factories), wind direction, and protection distance; the instruction needs to be supplemented with "focusing on the diffusion mechanism of point source emissions." Variant C (rural natural type) emphasizes accessibility to medical resources (distance to the nearest hospital) and aging population; the instruction needs to be supplemented with "focusing on the accessibility of medical facilities."

[0085] The following is an example of a complete workflow generated by the command: Suppose the system is processing grid IDG_1024.

[0086] A. Data Reading: Road network: 50 meters from the highway.

[0087] Buildings: Average height 30 meters (high), density 0.7 (dense).

[0088] POI: No factories, but all office buildings.

[0089] B. Semantic conversion: Road distance <100m is converted to high traffic exposure level; density >0.5 is converted to low ventilation conditions.

[0090] C. Command Assembly (Variation A - Urban Center), Background: High traffic exposure levels, low ventilation, office area. Analyze cardiovascular disease risk.

[0091] D. Agent Reasoning (LLM): High traffic noise increases the risk of high blood pressure. Poor ventilation leads to the accumulation of vehicle exhaust fumes (nitrogen oxides / NOx). Office areas imply a sedentary lifestyle. The output text reads: "Risk factors: sedentary lifestyle, traffic noise, nitrogen oxide (NOx) accumulation. Synergistic effects may exist." E. Vector Injection: Convert "Risk Factors..." into vectors. .

[0092] F. Model Prediction: ST-Agent-GNN Receiver By combining image features, a high mortality rate was predicted.

[0093] V. Automated Prompt Engineering (APE): During the fine-tuning phase, instructions can be optimized in reverse.

[0094] Let P be the instruction template. The goal is to find the optimal template P. * This minimizes the prediction error: ; Leveraging the self-reflective capabilities of the LLM (Local Management Module). If a grid prediction has a significant error (e.g., the model predicts low risk, but the actual risk is high), provide feedback to the LLM: "Your previous analysis indicated 'low risk,' but the actual data is very high. Please analyze the reasons and modify your instruction concerns." The LLM will output new concerns: "I overlooked the fact that although the area has good greenery, it is located under a high-voltage power line corridor. Next time, checks for 'electromagnetic radiation' or 'infrastructure NIMBY effect' should be added." The system automatically updates the background section of the instruction template.

[0095] S3. Input the prior knowledge vector and geographic numerical feature vector into the preset public health assessment agent, use the public health assessment agent to search the epidemiological knowledge base for macro-reasoning, and generate expert feature vectors. In this optional embodiment, the prior knowledge vector and geographic numerical feature vector are input into a preset public health assessment agent. The public health assessment agent is then used to retrieve an epidemiological knowledge base for macro-level reasoning to generate expert feature vectors, including the following steps: S31. The public health assessment intelligent agent uses data exploration tools to obtain statistical metadata of the target area based on the input prior knowledge vector and geographic numerical feature vector. S32. Generate query vectors based on statistical metadata, retrieve knowledge fragments from the pre-built epidemiological knowledge base; S33. Combine statistical metadata with knowledge fragments to generate a risk reasoning chain according to a preset reasoning template; S34. Input the risk reasoning chain into the pre-trained text embedding model for encoding, and output the expert feature vector.

[0096] In this optional embodiment, the preset public health assessment agent includes: The planning layer is used to parse user tasks, retrieve epidemiological knowledge based on statistical metadata, and generate expert feature vectors. The execution layer is used to call data preprocessing scripts, model inference interfaces, and post-processing tools to complete the prediction task. The critique layer is used to perform statistical rationality verification on the estimation results output by the public health level estimation model. When the verification results do not meet the preset standards, the public health level estimation model is triggered to re-infer the process. The memory module is used to maintain short-term session context information and a long-term success case library, and reuses historically effective expert feature vector generation patterns through geographic metadata similarity retrieval. A state machine is used to automatically switch between the initialization, knowledge injection preparation, execution, critique, and reporting states according to the order of task progress based on public health level estimation.

[0097] It should be explained that the agent adopts a Re-Act (Reasoning + Acting) architecture and deeply integrates RAG (Retrieval Enhanced Generation) technology. The agent here undertakes two key responsibilities: macroscopic reasoning and closed-loop control. Macroscopic reasoning involves analyzing geographic metadata to generate structured prior knowledge vectors (…). The data is directly input into the neural network. Closed-loop control is the complete process of autonomously orchestrating data preprocessing, model inference, and result verification. Specifically, the agent is designed as a Planner-Executor-Critic ternary structure.

[0098] Planner (L1 - Cognitive Layer): Responsible for understanding the task, retrieving epidemiological knowledge, and generating prior hypotheses. Executor (L2 - Operation Layer): Responsible for calling the Python toolchain (preprocessing scripts, model inference interface). Critic (L3 - Monitoring Layer): Responsible for verifying the statistical reasonableness of the prediction results and triggering recalculation when necessary.

[0099] The construction of the public health assessment intelligent agent includes the following steps: I. Domain Knowledge Base Construction and Retrieval (RAG Module)

[0100] In order for intelligent agents to provide professional prior knowledge, an environmental epidemiology knowledge base must be established.

[0101] A. Knowledge Base Content: This includes a literature database and a rule database. The literature database includes abstracts of academic papers on the relationship between cardiovascular diseases and environmental factors (PM2.5, noise, green view rate). The rule database includes rules such as "chemical plants less than 500m from residential areas are considered high-risk" and "an aging population greater than 20% is considered a sensitive group."

[0102] B. Vectorized storage: Using an embedding model (such as text-embedding-3-large) to convert text into vectors V. d They are stored in vector databases (such as Milvus / FAISS).

[0103] C. Retrieval Mathematical Principle: When the agent analyzes a grid, it generates a query vector q. It then retrieves the knowledge fragment with the highest relevance. : .

[0104] II. Prior Knowledge Generation and Vector Injection.

[0105] This is the core path of the agent-enhanced prediction model. The agent does not directly output predicted values, but instead outputs an expert feature vector. The specific steps are as follows: A. Environment Overview Reading: The agent calls the Data_Profiler tool to obtain statistical metadata (not pixel-level, but statistical level) of the target region R.

[0106] Input: Work area boundary. Output: {"Average road network density":0.8,"Main POI":["Power plant","Hospital"],"Average building height":25m}.

[0107] B. Chain-of-Thought Generation: Combining the knowledge retrieved from RAG, a natural language description is generated. .

[0108] The template is "Given [facts], and considering [expert rules], I assume..."; the result is... "The area contains a thermal power plant. Epidemiological literature indicates that SO2 emissions are associated with higher cardiovascular disease mortality. High building density (average spacing of 25 meters) suggests poor pollutant dispersion conditions. Risk bias: High." C. Semantic Embedding: ... Convert to mathematical vectors This allows it to be read by the neural network. .

[0109] In the formula, It is a pre-trained text embedding model (Encoder). It needs to be fed into the Cross-Attention layer of ST-Agent-GNN.

[0110] III. Tool Orchestration.

[0111] The intelligent agent needs to manipulate actual Python code. A standard tool interface is defined. The tool definition (JSON Schema) is as follows: 1. Define the data preprocessing capability interface: Construct a utility function description named `run_preprocessing`. Define its functional semantics as an "execution space ETL pipeline," responsible for transforming multi-source heterogeneous raw data into standardized grid tensors. Simultaneously, set `region_code` and `year` as required input parameters to limit the processing scope.

[0112] 2. Define the core inference capability interface: Construct a utility function description named `run_inference`. Define its functional semantics as "calling a deep neural network for prediction", clearly defining it as the entry point of the computing engine. Crucially, the parameter definition explicitly introduces `agent_knowledge_vector` (agent knowledge vector) as an input parallel to `tensor_path` (data path), thus establishing the "knowledge injection" technical path at the interface level.

[0113] 3. Define the statistical validation capability interface: Construct a utility function description named `spatial_statistics`. Define its functional semantics as "calculating spatial autocorrelation indices (such as Moran's I)", giving the agent the ability to self-evaluate the results. Set `result_grid_path` as the input parameter to perform post-processing quality analysis on the generated raster data.

[0114] Mathematical principles (Moran's I - used in the Critic layer): The agent uses... The indicators determine whether the generated health data has reasonable spatial clustering characteristics (health data is usually spatially continuous and should not be random noise).

[0115] In the formula, This is the spatial weight matrix. If... (Random distribution) The agent should determine that the prediction has failed, reject the output, and retry.

[0116] IV. Agent State Machine. To ensure the orderly execution of tasks, the agent maintains an internal state machine: A. State: Initialization. In this state, the user inputs: "Generate health monitoring data for area A in 2024." The agent needs to parse the intent and check if the original data for area A exists.

[0117] B. State: Knowledge Injection Preparation. In this state, the agent needs to read the basic geographic statistics of Area A. Retrieve the knowledge base. Generate... And transformed into .

[0118] C. State: Execution. In this state, the agent's actions need to call `run_preprocessing` (if not cached); and simultaneously `un_inference`, passing in... parameter.

[0119] D. State: Criticism. In this state, the agent needs to read the prediction results and calculate Moran's I and Outlier Z-score.

[0120] The logical judgment is as follows: if Max(Mortality) > the theoretical upper limit, an anomaly is indicated, and a rollback correction is initiated. If Moran'sI < 0.2, the space is discontinuous, indicating that the model may collapse.

[0121] E. Status: Report. In this status, the agent needs to generate a visual heatmap and write an analysis report.

[0122] V. Memory Module Design. The agent needs to remember user preferences and historical corrections. Short-term memory remembers the context of the current session (e.g., changing the color of the previous image to red). Long-term memory stores historical successful instructions and related data. .

[0123] If the user reports that the previous prediction of the industrial zone was very accurate, the agent will use the next generated prediction... Store it in a high-value sample library, and the next time a similar region is encountered (through metadata similarity matching), directly fine-tune and use that vector.

[0124] The intelligent agent interaction process is as follows: 1. Task initiated.

[0125] User command: The user issues a command to Agent_Core (LLM master controller) to request the generation of health monitoring data for a specific area (area X).

[0126] 2. Phase One: Knowledge Injection – Perception and Reasoning. In this phase, the intelligent agent is not merely a simple transmitter of information, but actively gathers information and forms expert opinions.

[0127] Obtaining metadata: Agent_Core first calls Data_Profiler (statistical tool) to request basic geographic metadata of region X (such as POI distribution, road network density, etc.).

[0128] Environmental profiling: Data_Profiler provides statistical results, such as: the area has a high density of factories and a scarcity of green space.

[0129] Knowledge retrieval: Agent_Core initiates targeted searches to RAG_DB (knowledge base) based on the environmental profile, such as querying "the relationship between dense chemical plants and cardiovascular disease risk".

[0130] Theoretical basis: RAG_DB returns relevant literature conclusions or epidemiological evidence.

[0131] Generate a priori vector: Agent_Core combines environmental profiling and theoretical basis to perform internal reasoning and generate a mathematical prior knowledge vector. ).

[0132] 3. Phase Two: Model Execution and Hybrid Computation. In this phase, the agent transmits expert intuition to the deep learning model.

[0133] Invoking Prediction: Agent_Core sends a prediction request to Model_Engine (ST-GNN model). Crucially, the input includes not only the basic data (Data) but also the data generated in the previous step. .

[0134] The calculation returns: Model_Engine combines the data and prior vectors to perform inference operations and returns the preliminary prediction results (Result_Grid) to Agent_Core.

[0135] 4. Stage Three: Self-Criticism (i.e., quality control). In this stage, the intelligent agent performs logical verification on the initial results generated by the AI ​​to prevent AI illusions.

[0136] Initiate quality control: Agent_Core submits the prediction results to Critic_Module, requesting a check on the reasonableness of the results.

[0137] By verifying that the Critic_Module calculates statistical indicators (such as Moran's I) to confirm that the data has reasonable autocorrelation in space and no outliers, it sends a pass / fail feedback to Agent_Core.

[0138] 5. Mission complete.

[0139] Result delivery: After confirming that the results are correct, Agent_Core will display the final health risk map and accompanying analysis report to the user.

[0140] S4. Construct a public health level estimation model; input multi-channel spatiotemporal tensors into the public health level estimation model, generate spatiotemporal feature vectors through feature extraction and spatiotemporal modeling, and dynamically fuse the spatiotemporal feature vectors with expert feature vectors using a gating fusion mechanism to output the estimation result of the public health level; In this optional embodiment, the public health level estimation model includes: A multimodal feature encoder is used to encode the vector rasterized features and visual features in the input multi-channel spatiotemporal tensor respectively, and then concatenate and fuse the encoding results to output multimodal fused features; The spatial interaction module is used to construct a spatial adjacency graph with standard resolution grid as nodes, and to aggregate neighborhood information of the multimodal fusion features through a graph attention network to output spatial features; The temporal evolution module is used to construct a sequence of spatial features of historical years in chronological order, and capture the temporal cumulative effect of health indicators through a long short-term memory network, outputting a spatiotemporal feature vector. The intelligent agent knowledge fusion layer is used to receive spatiotemporal feature vectors and expert feature vectors through a gating fusion mechanism, dynamically adjust the fusion ratio of the two using learnable gating weights, generate the final fused features, and map and output the public health level estimation results.

[0141] In this optional embodiment, a spatial adjacency graph is constructed using a standard resolution grid as nodes. A graph attention network is then used to aggregate neighborhood information of the multimodal fusion features, and the output spatial features include the following steps: S421. Treat each standard resolution grid as an independent graph node, set a geographic distance threshold, establish adjacency edges between grid nodes whose distance is less than the geographic distance threshold, and construct a spatial adjacency graph covering the study area. S422. For the central grid node and all neighboring grid nodes in the spatial adjacency graph, the multimodal fusion features are mapped to the attention computation space using a shared feature transformation matrix, and the attention coefficients of the neighboring nodes to the central node are calculated through the attention mechanism. S423. Normalize the attention coefficients to obtain the normalized attention weights for each neighboring node. S424. The multimodal fusion features of the neighborhood nodes are weighted and summed according to the normalized attention weights to obtain the spatial features.

[0142] It needs to be explained that a large deep learning model for spatiotemporal prediction is constructed. This model does not just passively receive data, but actively combines the prior knowledge (expert experience) provided by the agent through the attention mechanism to adjust the feature weights. The model (i.e., the public health level estimation model) is named ST-Agent-GNN (Spatio-Temporal Agent-Guided Graph Neural Network).

[0143] The model consists of four cascaded modules: A) A Multimodal Encoder, used to process imagery and structured data separately. B) A Spatial Interaction Module (Spatial GAT), used to capture geographic interactions between grids. C) A Temporal Evolution Module (Temporal LSTM), used to capture historical trends. D) An Agent-Knowledge Fusion Layer, which uses the agent's text vectors to weight and correct spatiotemporal features.

[0144] The model construction includes the following steps: I. Multimodal Feature Encoding.

[0145] The goal is to transform unstructured images and structured vector features Mapped to the same dimension of the latent space.

[0146] A. The Visual Channel uses a pre-trained ResNet-50 as the backbone network, removing fully connected layers and retaining only convolutional layers to extract texture and structural features. Mathematical formula: .

[0147] In the formula, For grid Remote sensing image slices; These are the learnable parameters of the CNN. The extracted visual feature vector (e.g., 2048-dimensional).

[0148] B. The Semantic Channel processes numerical features such as road network density, POI density, and building volume. Multilayer Perceptron (MLP) is used for feature transformation. Mathematical formula: .

[0149] In the formula, This is the normalized numerical eigenvector. Let represent the weight matrix and bias. σ is the ReLU activation function. This is a semantic feature vector.

[0150] C. Channel splicing mathematical formula: .

[0151] In the formula, For grid The final multimodal fused feature vector. This is a complete mathematical description of the grid state after merging information from different sources, and it will serve as the input to subsequent graph neural networks (GNNs) or temporal models (LSTMs). `Concat()` is a vector concatenation operation, which joins two feature vectors end-to-end along the channel dimension. For example, if... The length is 64. The length is 32, and the length of the concatenated vector is 96. This represents a visual feature vector. It originates from the processing of remote sensing satellite images and is usually extracted by convolutional neural networks (CNNs, such as ResNet), capturing visual information such as surface texture and green cover. This represents semantic features. It originates from the processing of unstructured or structured geographic data (such as POI distribution, road network statistics, and population density), and is usually encoded by a multilayer perceptron (MLP) to capture the functional attributes of the city. It is the set of real numbers. This represents the total dimension of the model's hidden layers. The concatenated vector It is a dimension A real-valued vector. Typically equal Dimensions plus Dimensions.

[0152] II. Spatial Modeling Based on Graph Attention Networks.

[0153] Cardiovascular diseases exhibit spatial clustering (e.g., grids surrounding an industrial area may all be at high risk). This invention employs a Graph Attention Network (GAT) to dynamically learn the importance of neighbors based on node features (e.g., upwind neighbors have greater influence than downwind neighbors).

[0154] Spatial modeling based on graph attention involves the following logical steps: 1. Spatial graph construction: Define each 100-meter grid as a graph node, and connect adjacent grids based on a set geographical distance threshold (e.g., 1.5km) to construct a spatial adjacency graph.

[0155] 2. Attention coefficient calculation: For the central grid and its neighboring grids, the correlation between them is calculated using their multimodal feature vectors, and attention weights are generated by Softmax normalization to quantify the influence of different neighbors.

[0156] 3. Neighborhood Feature Aggregation: Based on attention weights, the features of all neighboring grids are weighted and summed to dynamically integrate surrounding environmental information.

[0157] 4. Node state update: Output an enhanced feature vector that incorporates spatial context information, which serves as the input for subsequent time series modules.

[0158] The implementation steps of a graph attention network are as follows: A. Graph construction: Nodes: Each 100m grid is a node.

[0159] Edges: Edges exist between grids that are less than a threshold δ (e.g., 1.5km apart).

[0160] B. Attention Coefficient Calculation: For grid u and its neighbor v, calculate the correlation score: .

[0161] In the formula, `v` is the raw attention score or correlation coefficient. It quantifies the importance of neighboring grids `v` to the center grid `u`. This value is unnormalized and can be any real number. `Leaky ReLU` is the LeakyRectified Linear Unit activation function. This is a non-linear function used to increase the model's expressive power. It allows negative inputs to have small non-zero outputs, preventing neuron death. `a` is the attention weight vector. This is a learnable parameter vector that the model adjusts during training to score the concatenated feature set. `T` (superscript) stands for transpose. It transforms the column vector `a` into a row vector for dot product with subsequent vectors. This is the feature transformation matrix. It is a learnable parameter matrix used to map the original input feature vector to a high-dimensional feature space suitable for attention computation. The feature vector of the central grid u (such as a combination of population, POI, and road network features). Let v be the feature vector of the neighboring grid. || represents the vector concatenation operation. This transforms the feature vector of the center node. Features of neighboring nodes Connect the two ends together to form a longer vector.

[0162] Normalization yields the final attention weights .

[0163] In the formula, This is the normalized attention coefficient. This is the key point explicitly marked in the image. Its value is between 0 and 1, and for a center node u, the sum of α of all its neighbors is 1. It determines how much information is absorbed from the neighbor v when aggregating features. exp(⋅) is the exponential function (e^(-α / v)). x Its function is to convert the raw score into a single value. Convert to positive numbers and amplify the differences in high scores. ∑ represents the summation. This is the denominator of the classic Softmax operation. Let k be the first-order neighborhood set of the central grid u. That is, the set of all grids k that are spatially adjacent to u (or within a threshold distance). k is the neighborhood set. This is any neighboring node in the set. This term is used to traverse all neighbors to calculate the sum, thereby achieving normalization.

[0164] C. Feature aggregation: .

[0165] In the formula, Grid features that incorporate information about the surrounding environment.

[0166] III. Temporal Modeling.

[0167] The health data is a time series (past 5 years). The cumulative exposure effect is captured using an LSTM (Long Short-Term Memory) model in time-series evolution modeling.

[0168] Temporal evolution modeling includes the following logical steps: 1. Feature sequence construction: Arrange the grid space feature vectors of each historical year (such as the past 5 years) output by the graph attention module in chronological order to construct a historical feature sequence.

[0169] 2. Recurrent network input: Input the feature sequence into a Long Short-Term Memory (LSTM) network and use its recurrent structure to process the time-dimensional dependencies.

[0170] 3. Memory state update: Through the gating mechanism (forget gate, input gate) inside the LSTM, forward propagation is carried out at time steps to dynamically update the cell state in order to capture the long-term evolution trend of health indicators.

[0171] 4. Spatiotemporal feature output: Extract the hidden state vector of the last time step as the final spatiotemporal feature representation that aggregates the historical evolution rules, and transmit it to the agent knowledge fusion layer.

[0172] For grid (I,j), input its spatial feature sequence {h} over the past T years. 1 ,h 2 ,…,h T}: The hidden state of the last time step is taken as the spatiotemporal feature representation: .

[0173] IV. Agent-Knowledge Reinforcement. This is the core reinforcement step of this solution, where the model heeds the agent's suggestions. Design logic: The agent generates a prior knowledge vector based on macroscopic data (…). This will be used as a query or context to filter spatiotemporal features through a cross-attention mechanism.

[0174] The knowledge reinforcement and fusion of intelligent agents involves the following logical steps: 1. Dual-source vector reception: Simultaneously receive the spatiotemporal feature vector (representing data patterns) output by the temporal evolution module, and the prior knowledge vector (representing expert experience) generated by the agent through knowledge retrieval and reasoning.

[0175] 2. Gating coefficient calculation: Construct a learnable gating network, concatenate two sets of vectors and map them through the Sigmoid function to adaptively calculate the gating coefficients used to balance data and knowledge weights.

[0176] 3. Feature Dynamic Modulation: The spatiotemporal feature vector and the prior knowledge vector are weighted element by element using gating coefficients to dynamically adjust the retention ratio of the two.

[0177] 4. Enhance feature output: Fuse the modulated two-way features (e.g., by weighted summation) to generate a semantically aligned and knowledge-enhanced final feature representation, which is then input into the prediction and regression layer.

[0178] Scenario example: Data features (HST) indicate that a certain grid has a high density of buildings and a low amount of green space. (Agent knowledge) Information included: This area is located downwind of the chemical plant, and dense building density leads to pollutant retention. The cross-attention mechanism amplifies the weights of features related to airflow in the HST.

[0179] Mathematical formula (Cross-Attention): 1. Define Query, Key, Value: (Features from the data); (Knowledge from the intelligent agent); (Knowledge from the intelligent agent).

[0180] A gating mechanism is employed to inject knowledge, enabling the model to balance data and expert opinions. Specifically, the gating mechanism learns adaptive weights z∈[0,1], dynamically adjusting the information flow like a soft valve. It automatically calculates the mixing ratio of data features and the agent's prior knowledge: z controls data retention, and (1-z) controls knowledge injection, thus achieving a dynamic balance between data-driven approaches and expert experience. The mathematical formula for the adaptive weight z is: .

[0181] In the formula, z represents the adaptive gating weight / coefficient. This is a vector with values ​​between [0,1]. It acts as a soft valve or control knob, determining whether the system relies more on data or expert knowledge. The closer the value is to 1, the more it relies on data; the closer the value is to 0, the more it relies on knowledge. σ (Sigma) is the sigmoid activation function. Its function is to compress the calculation results into the interval between 0 and 1, making them a valid weight ratio. W z This is the learnable weight matrix for the gating network. These are parameters automatically learned by the model during training, used to determine how many weights should be assigned in the current situation. H ST These are spatio-temporal features. They represent data. These are features extracted by the model from historical data, road networks, POIs, and other objective information (usually from LSTM or GNN modules). This is the agent's prior knowledge vector. It represents knowledge. This is an expert experience vector generated by the LLM agent after retrieving documents and reasoning. || or […] represents the vector concatenation operation. It combines data features and knowledge features, allowing the gating network to see both aspects of information simultaneously when making decisions.

[0182] The final feature fusion formula is: .

[0183] In the formula, This is the final fused feature vector. This is the final result after weighted mixing, representing an optimal trade-off between data facts and expert opinions, and will be directly used for the final prediction (output Y). ⊙ represents element-wise multiplication (Hadamard Product). This indicates that each element in the vector is weighted separately. tanh is the hyperbolic tangent activation function. It is used to perform non-linear transformations and standardization on the features, typically mapping the values ​​to the range [-1, 1], making the feature distribution more stable. For use in transforming spatiotemporal features The weight matrix. For use in transforming knowledge vectors The weight matrices are used to map the original features to the same semantic space for additive fusion. (1-z) represents the complementary weights for knowledge injection. The mathematical complementarity is such that if z controls how many data features are retained, then (1-z) controls how much expert knowledge is injected. The sum of the two is 1, ensuring a dynamic balance of information.

[0184] V. Prediction Head.

[0185] Finally, it is mapped to the [0,1] interval through a fully connected layer.

[0186] Mathematical formula: .

[0187] The data flow diagram code is as follows: “graph TD subgraph Input [Input Layer] RS [Remote Sensing Image I] --> CNN [ResNet-50] Vec[vector feature v] --> MLP1[MLP Encoder] end subgraph Feature [Feature Fusion Layer] CNN --> Concat MLP1 --> Concat Concat --> |x| GAT [Graph Attention Network GAT] end subgraph Temporal [Time-series layer] GAT --> |Sequence h_1...h_t| LSTM [Time-Sequence LSTM] end Subgraph Agent [Agent Reinforcement Layer] Prompt [Environment Description] --> LLM [Agent] LLM --> | Generate Text | Embedding API Emb --> | Knowledge Vector K_agent | Gate [Gating Fusion Mechanism] LSTM --> |Spatiotemporal Features H_st| Gate end subgraph Output [output layer] Gate -->|H_final| MLP2 [Prediction Header] MLP2 --> Result [Health Monitoring Data Y] end".

[0188] S5. A spatiotemporal block verification strategy is adopted to perform basic pre-training and knowledge injection fine-tuning on the public health level estimation model. Among them, the knowledge injection fine-tuning uses prior knowledge vectors and expert feature vectors as supervision signals, and optimizes the parameters of the public health level estimation model through loss function constraints to obtain the trained public health level estimation model. In this optional embodiment, a spatiotemporal block verification strategy is adopted to perform basic pre-training and knowledge injection fine-tuning on the public health level estimation model, including the following steps: S51. Divide the study area into several non-overlapping spatial blocks, and divide the data into training set and validation set according to the time dimension; S52. Input the multi-channel spatiotemporal tensor of the training set into the public health level estimation model, use real health data as the supervision signal, and perform basic pre-training of the public health level estimation model using the spatiotemporal hybrid loss function. S53. Load the parameters of the public health level estimation model after basic pre-training, freeze the underlying parameters of the multimodal feature encoder, use prior knowledge vector and expert feature vector as supervision signals, and use the training set and expert feature vector to fine-tune the public health level estimation model by knowledge injection. S54. Use the validation set to evaluate the performance of the public health level estimation model after knowledge injection fine-tuning, and adjust the hyperparameters of the public health level estimation model based on the evaluation results.

[0189] In this optional embodiment, loading the parameters of the pre-trained public health level estimation model, freezing the underlying parameters of the multimodal feature encoder, and using prior knowledge vectors and expert feature vectors as supervision signals to fine-tune the public health level estimation model through knowledge injection using the training set and expert feature vectors includes the following steps: S531. Load the parameters of the pre-trained public health level estimation model and freeze the underlying parameters of the multimodal feature encoder. S532. Input the multi-channel spatiotemporal tensor of the training set into the public health level estimation model. Through the forward propagation of the public health level estimation model, feature extraction and spatiotemporal modeling are completed, and the corresponding spatiotemporal feature vector is generated. S533. Input the spatiotemporal feature vector and the corresponding expert feature vector into the gated fusion layer of the public health level estimation model. Use the learnable gate weights of the public health level estimation model to dynamically fuse the spatiotemporal feature vector and the expert feature vector, generate the final fused feature, and output the public health level estimation result. S534. Using prior knowledge vectors and expert feature vectors as supervision signals, a total loss function is constructed by combining regression loss, spatial smoothing loss and semantic alignment loss. The parameters of the gated fusion layer and prediction output layer of the public health level estimation model are optimized and updated through backpropagation of loss, so as to realize the knowledge injection fine-tuning of the public health level estimation model.

[0190] It should be noted that this process is divided into three progressive stages: Base Pre-training, Instruction Tuning, and Reinforcement Feedback Optimization.

[0191] I. Training Data Partition Strategy (Spatio-Temporal Splitting). Block-Time Validation is employed.

[0192] A. Spatial division: Divide the study area into K non-overlapping blocks (e.g., large blocks of 5km×5km).

[0193] B. Time Division: Training Set ( This contains data from the first T-1 years, covering 80% of the spatial blocks. Validation set ( The test set contains data from the first T-1 years, covering the remaining 20% ​​of the spatial blocks (used to test spatial generalization ability). It contains all the data for year T (the last year) (used to test time prediction capabilities).

[0194] Phase 1: Base Pre-training.

[0195] Objective: To train the numerical intuition of ST-Agent-GNN so that it can map from X (images / population / POIs) to Y (health indicators) without connecting agent vectors.

[0196] The loss function is designed to use a spatio-temporal hybrid loss: ; A, This represents the regression accuracy loss. For extremely skewed healthy data (most regions have low risk, while a few regions have extremely high risk), Huber Loss is used instead of MSE to reduce the excessive interference of outliers on the gradient.

[0197] ; ; In the formula, This represents the true mortality rate. These are the model's predicted values. The Huber threshold controls sensitivity to outliers.

[0198] B This represents spatial Laplacian regularization, which enforces that geographically adjacent grids must have similar outputs if their input features are similar (smoothness constraint). It utilizes the Laplacian matrix from graph theory. .

[0199] In the formula, This is the prediction vector for the entire image. Let be the graph Laplace matrix. The adjacency weights for grids u and v are given (the weights should be small if the feature differences are large, and large if they are small). This is dynamically calculated by GAT.

[0200] Phase 3, Stage 2: Knowledge-Enhanced Instruction Tuning. This phase activates the gating fusion mechanism and cross-attention, enabling the model to learn to utilize prior vectors generated by the agent. .

[0201] Training methods: A. Freeze parameters: Freeze the underlying parameters of ResNet (visual) and MLP (semantic).

[0202] B. Activation parameters: Only train the Agent-Fusion Layer, Attention Weights, and Prediction Head.

[0203] C. Constructing instruction samples: For high-risk areas in the training set, have the agent generate a description: Due to the high density of chemical plants, the risk is extremely high. For low-risk areas, the generated description is: high green coverage and excellent environment. .

[0204] Adversarial Sampling: Intentionally providing false information (in a favorable environment) to high-risk areas forces the model to learn to distinguish them during training.

[0205] New term added to the loss function: semantic consistency loss ( ). Ensure that when When indicating high risk, the distribution of prediction results The average value should be high.

[0206] Introduce an auxiliary small mapping network ,Will Mapped to expected mortality range .

[0207] This forces the statistical distribution (mean, variance) of the prediction results to be consistent with the agent's macroscopic judgment.

[0208] Total fine-tuning loss: .

[0209] Phase 3: Agent Reinforcement Closed Loop (Agent RLHF).

[0210] Optimize the agent's instruction design and inference logic. If the model makes a wrong prediction, the agent needs to reflect on whether "the prior knowledge I provided was inaccurate or whether I missed a key Point of Interest (POI)?"

[0211] In terms of specific technology, this is considered a reinforcement learning problem.

[0212] Agent (Policy π): Takes geographic metadata as input and outputs text description. .

[0213] Environment (Prediction Model): Receiver (Right now ), outputting the prediction error.

[0214] Reward function (R): The higher the prediction accuracy, the higher the reward.

[0215] Formula definition: .

[0216] If incorporating the agent's knowledge reduces the MSE (smaller error), the reward is positive. If incorporating the agent's knowledge increases the MSE (misleads the model), the reward is negative.

[0217] Update strategy: Since fine-tuning LLM itself is costly, an optimization strategy using in-context learning can be used instead of directly updating LLM parameters.

[0218] Context-based learning optimization strategies aim to improve agent performance by maintaining a dynamic library of good examples without updating the LLM model parameters. The specific implementation steps are as follows: 1. Initialization of the best example library: Build a vectorized database specifically for storing environmental metadata-inference thought chain pairs that have performed well in historical tasks.

[0219] 2. Reasoning performance evaluation: After a single prediction task is completed, the decrease in prediction error before and after introducing the agent's prior knowledge vector is calculated to quantify the prediction gain reward brought by the reasoning logic.

[0220] 3. High-value sample accumulation: If the gain reward of a certain reasoning exceeds the set threshold, the system determines that the reasoning logic is valid and automatically stores it as a high-value sample in the excellent sample library.

[0221] 4. Contextual Similarity Retrieval: When the agent processes a new grid prediction task, it first calculates the similarity between the current environment metadata and the historical records in the best example library, and retrieves the most matching historical best examples.

[0222] 5. Few-shot context injection: The retrieved historical best examples are used as few-shot demonstrations and directly embedded into the prompt context window of the current task, guiding the agent to imitate historical successes to generate high-quality prior knowledge vectors.

[0223] The present invention mainly consists of the following three steps.

[0224] 1. The agent maintains a high-scoring instruction library. 2. After each prediction, calculate R. 3. If R > Threshold, store the current analysis logic in the library as a future small-sample case.

[0225] In this invention, the training hyperparameters and configurations are shown in Table 1.

[0226] Table 1: Training Hyperparameters and Configuration Parameters Setting suggestions Principle Explanation Optimizer Adam W Standard choice for handling sparse gradients: Weight Decay prevents overfitting. Learning Rate 1e−4(Base),1e−5(Fine-tune) A large learning rate is used during pre-training, and a small learning rate is used during fine-tuning to avoid damaging the feature extractor. Batch Size 32 Regions (each Region contains N grids) Graph-based training is typically conducted on a subgraph-by-subgraph basis. Epochs 100 (Pre-train) + 50 (Fine-tune) Ensure full convergence Loss Weights λ1=0.1(Spatial),λ2=0.05(Align) Empirical data, ensuring regression accuracy remains the primary objective. Augmentation Random rotation, horizontal flip (images only) Increase the robustness of visual models The gradient flow diagram is as follows: Phase 1: Forward propagation (data flow, generating predictions).

[0227] The source of the gradient is the prediction error, so the prediction value needs to be generated through forward propagation first.

[0228] 1. Input is processed in parallel with dual input channels: Data path: The raw input data X (images, population, POIs, etc.) enters the GNN Backbone and is then processed to output feature vectors.

[0229] Knowledge Path: The agent's Prompt is encoded to generate Embedding K (expert knowledge vector), which is also transformed into feature vectors.

[0230] 2. Feature fusion: Features from both data and knowledge sources are combined in the Fusion Layer (e.g., through gating mechanisms or splicing).

[0231] 3. Generate predictions: The fused features are then calculated to output the final predicted value. (Pred).

[0232] The second stage: loss calculation (the origin of the gradient).

[0233] Three different errors are calculated based on the prediction results, and these three errors converge into the total gradient signal.

[0234] 1. Regression loss ( ): Calculate the predicted value The difference between the actual label Y (GT) and the actual label (such as Huber Loss) is used to measure "accuracy".

[0235] 2. Spatial smoothing loss ( ): Based on predicted values ​​only It performs its own calculations and checks whether the predicted values ​​of adjacent grids abruptly change, which is used to measure "whether it is smooth" or not.

[0236] 3. Semantic alignment loss ( ): Calculate the predicted value Consistency with the Embedding K distribution is used to measure "whether the agent's advice was heeded".

[0237] 4. Total Loss Aggregation: The weighted sum of the above three factors yields the total loss ( This is the starting point for backpropagation.

[0238] The third stage: backpropagation (the specific process of gradient flow).

[0239] The gradient (i.e., the error signal) starts from the total loss and flows backward along the network to update the model's weight parameters.

[0240] Path A: Main Task Update; Flow direction: →⋯→Backbone; Logically, the gradient generated by the total loss flows back to the GNN backbone. This forces the backbone network to adjust its parameters so that it can extract more accurate and spatially consistent image and geographic features next time.

[0241] Path B: Update the Fusion Strategy layer; Flow direction: →⋯→Fusion_Layer; Logically, the gradient generated by the total loss flows back to the Fusion Layer. This forces the model to learn when to trust the data and when to trust the agent, adjusting the weight parameters for feature fusion between the two.

[0242] Path C: Update Alignment Header (Alignment Specific Update); Flow direction: →Projection_Head; Logical: Semantic alignment loss ( The specific gradients generated are fed back to the projection head (usually located at the backbone end or the embedding interface). This forces the feature space learned by the model to be consistent with the agent's knowledge vector space.

[0243] S6. Input the multi-channel spatiotemporal tensor of the region to be estimated into the trained public health level estimation model to obtain the optimized public health level estimation result.

[0244] It needs to be explained that, for example Figure 1-3 As shown, this method includes the following steps: Step 1, Autonomous Data Preprocessing: Construct an automated spatial ETL (Extract-Transform-Load) pipeline, and use geometric calculations and affine transformations to uniformly clean and map multi-source heterogeneous data (remote sensing images, vector road networks, POIs, population raster) into standard 100-meter grid multi-channel spatiotemporal tensors.

[0245] Step 2, Core Model (ST-Agent-GNN) Construction: A deep network is built, incorporating CNN visual extraction, GAT spatial interaction, and LSTM temporal evolution. Its core innovation lies in the gating fusion mechanism, which uses cross-attention to receive prior knowledge vectors from the agent's input to weight features. Specifically, the gating mechanism learns adaptive weights z∈[0,1], dynamically adjusting the information flow like a soft valve. It automatically calculates the mixing ratio of data features and the agent's prior knowledge; z controls data retention, and (1−z) controls knowledge injection, thus achieving a dynamic balance between data-driven approaches and expert experience.

[0246] Step 3, Agent Architecture Design: A planning-execution-critique closed-loop structure is adopted. The agent first plans the execution process, then executes specific commands, and finally evaluates and corrects the execution results. The agent utilizes RAG to retrieve epidemiological knowledge, performs macro-level reasoning about the geographical environment, identifies factors that may have potential health impacts and their mechanisms of action, and generates expert feature vectors to enhance and guide the performance of the ST-Agent-GNN model. And monitor the reasonableness of the results.

[0247] Step 4, Fine-tuning and Training: Spatiotemporal block verification is used to prevent data leakage. The strategy is divided into two stages: basic pre-training and knowledge injection fine-tuning. Through loss function constraints, the ST-Agent-GNN model is forced to learn and heed expert advice from the agent.

[0248] Step 5, Prompt Design: Construct a dynamic semantic mapping system to transform dry geographic data into structured natural language thought chains (CoT), guiding the agent to generate high-quality, vectorizable prior causal logic.

[0249] The dual-brain collaborative architecture of AI Agent and AI Model in this invention breaks the traditional binary opposition of either a purely data-driven black box (deep learning) or a purely rule-driven white box (expert system). This method uses the LLM agent (i.e., the public health assessment agent) as the domain expert brain (responsible for qualitative reasoning, knowledge retrieval, and logical verification), and the graph neural network as the computational brain (responsible for quantitative fitting and spatiotemporal feature extraction), achieving a deep integration of qualitative analysis and quantitative computation.

[0250] This invention is based on a knowledge-guided mechanism. Traditional AI models passively learn from data. The agent in this method can proactively generate vectors containing prior epidemiological knowledge by analyzing environmental metadata. Through an innovative gated cross-attention mechanism, this knowledge vector can dynamically guide the neural network on which features to focus on (e.g., pollution sources in industrial areas and green spaces in residential areas), thereby significantly improving the model's environmental adaptability.

[0251] This invention presents an autonomous, closed-loop Re-Act (Reasoning and Acting) workflow. The agent not only participates in prediction but also manages the entire process. It possesses self-critical capabilities (Critic module), enabling it to verify the reasonableness of prediction results using statistical indicators (such as Moran's I) and autonomously trigger re-reasoning. This achieves an automated closed loop from data processing to result verification.

[0252] This invention possesses the capability to process multi-source heterogeneous big data. Through a multimodal encoder, it can simultaneously process unstructured remote sensing imagery and structured vector / statistical data, extracting information richness far exceeding that of traditional LUR models. This invention also exhibits powerful spatiotemporal modeling capabilities. Utilizing graph neural networks (GNNs) and LSTMs, it explicitly models the diffusion effects of health risks in geospatial space and their cumulative effects over time, overcoming the static limitations of traditional methods. Furthermore, this invention demonstrates higher robustness and generalization ability. By introducing prior knowledge from domain experts as constraints, the model is less prone to severe overfitting or producing predictions that defy common sense (AI illusions) in regions with sparse or noisy data.

[0253] According to one embodiment of the present invention, a public health level estimation system based on prior knowledge and artificial intelligence is also provided, the system comprising: The data preprocessing module is used to preprocess multi-source heterogeneous data using a preset spatial ETL pipeline, generate multi-channel spatiotemporal tensors of standard resolution grids, and extract the geographic numerical feature vectors corresponding to each standard resolution grid. The prior knowledge vector generation module is used to transform geographic numerical feature vectors into natural language thought chains based on preset dynamic semantic mapping rules; and to encode the natural language thought chains into prior knowledge vectors using a text embedding model. The expert feature vector generation module is used to input prior knowledge vectors and geographic numerical feature vectors into a preset public health assessment agent, and use the public health assessment agent to search the epidemiological knowledge base for macro-reasoning to generate expert feature vectors. The public health level estimation model construction module is used to construct a public health level estimation model. It inputs a multi-channel spatiotemporal tensor into the public health level estimation model, generates spatiotemporal feature vectors through feature extraction and spatiotemporal modeling, and dynamically fuses the spatiotemporal feature vectors with expert feature vectors using a gating fusion mechanism to output the estimation result of the public health level. The public health level estimation model optimization module is used to perform basic pre-training and knowledge injection fine-tuning of the public health level estimation model using a spatiotemporal block verification strategy. Among them, the knowledge injection fine-tuning uses prior knowledge vectors and expert feature vectors as supervision signals, and optimizes the parameters of the public health level estimation model through loss function constraints to obtain the trained public health level estimation model. The evaluation results module is used to input the multi-channel spatiotemporal tensor of the region to be estimated into the trained public health level estimation model to obtain the optimized public health level estimation results.

[0254] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A public health level estimation method based on prior knowledge and artificial intelligence, characterized in that, The method includes: S1. Using a pre-defined spatial ETL pipeline, preprocess multi-source heterogeneous data to generate multi-channel spatiotemporal tensors of standard resolution grids, and extract the geographic numerical feature vectors corresponding to each standard resolution grid. S2. Based on the preset dynamic semantic mapping rules, the geographic numerical feature vector is transformed into a natural language thought chain; using a text embedding model, the natural language thought chain is encoded into a prior knowledge vector. S3. Input the prior knowledge vector and geographic numerical feature vector into the preset public health assessment agent, use the public health assessment agent to search the epidemiological knowledge base for macro-reasoning, and generate expert feature vectors. S4. Construct a public health level estimation model; input multi-channel spatiotemporal tensors into the public health level estimation model, generate spatiotemporal feature vectors through feature extraction and spatiotemporal modeling, and dynamically fuse the spatiotemporal feature vectors with expert feature vectors using a gating fusion mechanism to output the estimation result of the public health level; S5. A spatiotemporal block verification strategy is adopted to perform basic pre-training and knowledge injection fine-tuning on the public health level estimation model. Among them, the knowledge injection fine-tuning uses prior knowledge vectors and expert feature vectors as supervision signals, and optimizes the parameters of the public health level estimation model through loss function constraints to obtain the trained public health level estimation model. S6. Input the multi-channel spatiotemporal tensor of the region to be estimated into the trained public health level estimation model to obtain the optimized public health level estimation result.

2. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 1, characterized in that, The process of preprocessing multi-source heterogeneous data using a pre-defined spatial ETL pipeline to generate multi-channel spatiotemporal tensors of standard resolution grids and extracting geographic numerical feature vectors corresponding to each standard resolution grid includes the following steps: S11. Read the original coordinate system of all multi-source heterogeneous data, select the appropriate projection coordinate system according to the study area, and reproject all data to the appropriate projection coordinate system. S12. Calculate the boundary range of the study area, divide the grid according to the preset grid resolution, and generate a grid index; S13. Perform rasterization on the vector data to obtain vector rasterized features; S14. Virtually slice the remote sensing image based on affine transformation, and extract the corresponding pixel region as visual feature according to the grid boundary. S15. Map the point data to the grid through spatial connections and perform feature aggregation to obtain the point data aggregated features; S16. Align the vector raster features, visual features, and point data aggregation features according to the grid index and time dimension, combine them into a multi-channel spatiotemporal tensor, and normalize the numerical features of each grid to generate the geographic numerical feature vector corresponding to each grid.

3. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 2, characterized in that, The vector rasterization features include building volume features and road network density features; wherein, the building volume features are obtained by calculating and aggregating the building base area and the intersection area within the grid; the road network density features are obtained by calculating the effective length of roads within the grid using a line segment clipping algorithm.

4. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 1, characterized in that, The method of transforming geographic numerical feature vectors into natural language thought chains based on preset dynamic semantic mapping rules, and encoding the natural language thought chains into prior knowledge vectors using a text embedding model, includes the following steps: S21. According to the preset dynamic semantic mapping rules, map each numerical feature in the geographic numerical feature vector to the corresponding natural language description. S22. Based on the analytical logic of environmental epidemiology and the type of area to which the grid belongs, the natural language description is constructed into a natural language thought chain; S23. Input the natural language thought chain into the pre-trained text embedding model for encoding, and output it as a prior knowledge vector.

5. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 1, characterized in that, The process of inputting prior knowledge vectors and geographic numerical feature vectors into a pre-defined public health assessment agent, and using the public health assessment agent to retrieve epidemiological knowledge bases for macro-level reasoning to generate expert feature vectors includes the following steps: S31. The public health assessment intelligent agent uses data exploration tools to obtain statistical metadata of the target area based on the input prior knowledge vector and geographic numerical feature vector. S32. Generate query vectors based on statistical metadata, retrieve knowledge fragments from the pre-built epidemiological knowledge base; S33. Combine statistical metadata with knowledge fragments to generate a risk reasoning chain according to a preset reasoning template; S34. Input the risk reasoning chain into the pre-trained text embedding model for encoding, and output the expert feature vector.

6. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 5, characterized in that, The preset public health assessment agent includes: The planning layer is used to parse user tasks, retrieve epidemiological knowledge based on statistical metadata, and generate expert feature vectors. The execution layer is used to call data preprocessing scripts, model inference interfaces, and post-processing tools to complete the prediction task. The critique layer is used to perform statistical rationality verification on the estimation results output by the public health level estimation model. When the verification results do not meet the preset standards, the public health level estimation model is triggered to re-infer the process. The memory module is used to maintain short-term session context information and a long-term success case library, and reuses historically effective expert feature vector generation patterns through geographic metadata similarity retrieval. A state machine is used to automatically switch between the initialization, knowledge injection preparation, execution, critique, and reporting states according to the order of task progress based on public health level estimation.

7. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 1, characterized in that, The public health level estimation model includes: A multimodal feature encoder is used to encode the vector rasterized features and visual features in the input multi-channel spatiotemporal tensor respectively, and then concatenate and fuse the encoding results to output multimodal fused features; The spatial interaction module is used to construct a spatial adjacency graph with standard resolution grid as nodes, and to aggregate neighborhood information of the multimodal fusion features through a graph attention network to output spatial features; The temporal evolution module is used to construct a sequence of spatial features of historical years in chronological order, and capture the temporal cumulative effect of health indicators through a long short-term memory network, outputting a spatiotemporal feature vector. The intelligent agent knowledge fusion layer is used to receive spatiotemporal feature vectors and expert feature vectors through a gating fusion mechanism, dynamically adjust the fusion ratio of the two using learnable gating weights, generate the final fused features, and map and output the public health level estimation results.

8. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 7, characterized in that, The process of constructing a spatial adjacency graph using a standard resolution grid as nodes, and then aggregating neighborhood information of the multimodal fusion features through a graph attention network to output spatial features includes the following steps: S421. Treat each standard resolution grid as an independent graph node, set a geographic distance threshold, establish adjacency edges between grid nodes whose distance is less than the geographic distance threshold, and construct a spatial adjacency graph covering the study area. S422. For the central grid node and all neighboring grid nodes in the spatial adjacency graph, the multimodal fusion features are mapped to the attention computation space using a shared feature transformation matrix, and the attention coefficients of the neighboring nodes to the central node are calculated through the attention mechanism. S423. Normalize the attention coefficients to obtain the normalized attention weights for each neighboring node. S424. The multimodal fusion features of the neighborhood nodes are weighted and summed according to the normalized attention weights to obtain the spatial features.

9. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 1, characterized in that, The method of employing a spatiotemporal block-based verification strategy to perform basic pre-training and knowledge-injection fine-tuning on the public health level estimation model includes the following steps: S51. Divide the study area into several non-overlapping spatial blocks, and divide the data into training set and validation set according to the time dimension; S52. Input the multi-channel spatiotemporal tensor of the training set into the public health level estimation model, use real health data as the supervision signal, and perform basic pre-training of the public health level estimation model using the spatiotemporal hybrid loss function. S53. Load the parameters of the public health level estimation model after basic pre-training, freeze the underlying parameters of the multimodal feature encoder, use prior knowledge vector and expert feature vector as supervision signals, and use the training set and expert feature vector to fine-tune the public health level estimation model by knowledge injection. S54. Use the validation set to evaluate the performance of the public health level estimation model after knowledge injection fine-tuning, and adjust the hyperparameters of the public health level estimation model based on the evaluation results.

10. The public health level estimation method based on prior knowledge and artificial intelligence according to claim 9, characterized in that, The steps of loading the pre-trained public health level estimation model parameters, freezing the underlying parameters of the multimodal feature encoder, using prior knowledge vectors and expert feature vectors as supervision signals, and fine-tuning the public health level estimation model using the training set and expert feature vectors through knowledge injection include the following: S531. Load the parameters of the pre-trained public health level estimation model and freeze the underlying parameters of the multimodal feature encoder. S532. Input the multi-channel spatiotemporal tensor of the training set into the public health level estimation model. Through the forward propagation of the public health level estimation model, feature extraction and spatiotemporal modeling are completed, and the corresponding spatiotemporal feature vector is generated. S533. Input the spatiotemporal feature vector and the corresponding expert feature vector into the gated fusion layer of the public health level estimation model. Use the learnable gate weights of the public health level estimation model to dynamically fuse the spatiotemporal feature vector and the expert feature vector, generate the final fused feature, and output the public health level estimation result. S534. Using prior knowledge vectors and expert feature vectors as supervision signals, a total loss function is constructed by combining regression loss, spatial smoothing loss and semantic alignment loss. The parameters of the gated fusion layer and prediction output layer of the public health level estimation model are optimized and updated through backpropagation of loss, so as to realize the knowledge injection fine-tuning of the public health level estimation model.

Images