Public health early warning system based on composite multi-modal large model

Abstract

The application relates to the technical field of artificial intelligence, and particularly provides a public health early warning system based on a composite multimodal large model, which comprises the following steps: acquiring meteorological, environmental and health multi-source heterogeneous data through a data acquisition agent; performing cross-modal attention fusion through a composite multimodal large model by a feature fusion agent to generate a joint feature vector; constructing a meteorological-pollution-health three-layer causal correlation model by a causal modeling agent to calculate an average causal effect; identifying a pollution source and a diffusion path by a pollution tracing agent; and quantifying a comprehensive health risk index and grading early warning by a risk early warning agent. The collaborative agent adopts a responsive flow and a back pressure mechanism to realize data pipeline flow control, dynamically reconstructs a directed acyclic graph topology based on a confidence score, and combines multi-agent reinforcement learning to hierarchically schedule resources and trigger resampling and retraining. The application realizes fine and adaptive early warning driven by causal inference, and significantly reduces the false positive rate and delay.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a public health early warning system based on a composite multimodal large model. Background Technology

[0002] Existing public health early warning systems typically employ threshold-triggered mechanisms, automatically issuing warnings when meteorological data (such as temperature and wind speed) or environmental pollutant concentrations (such as PM2.5 and PM10) exceed preset thresholds. This type of warning is crude and broad, often simply stating "poor air quality, reduce outdoor activities," lacking detailed assessments of specific health risks (such as respiratory or eye diseases) and failing to differentiate the vulnerabilities of different population groups.

[0003] In recent years, large-scale modeling and intelligent agent technologies have begun to be applied in the industrial field. However, existing industrial intelligent agent systems still follow the hard-coded rule approach for scenario adaptation, where developers pre-write a large number of "IF-THEN" conditional statements to match specific scenarios and trigger corresponding intelligent agents. When facing complex scenarios in the public health field (such as sandstorms simultaneously causing a high incidence of respiratory and ophthalmological diseases), this approach faces problems such as rule combination explosion and inability to generalize to unknown scenarios, leading to delayed warnings or false alarms.

[0004] Therefore, there is an urgent need for an intelligent early warning method that can adapt to complex public health scenarios and dynamically adjust the execution process. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a public health early warning system based on a composite multimodal large model to solve the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a public health early warning system based on a composite multimodal large model, comprising: Multiple specialized intelligent agents, including at least: The data acquisition agent is used to collect multi-source heterogeneous data, including meteorological time-series data, environmental spatial data, and health-related data. The feature fusion agent is used to perform cross-modal attention fusion on the multi-source heterogeneous data using a pre-trained composite multimodal large model, extract deep features and generate a joint feature vector; A causal modeling agent is used to construct a three-layer causal association model based on the joint feature vector. The three layers include at least a meteorological-pollution layer, a pollution-health layer, and a composite causal layer, and the average causal effect is calculated. A pollution source tracing intelligent agent is used to identify pollution sources and diffusion paths based on the joint feature vector and average causal effect. A risk warning intelligent agent is used to calculate a comprehensive health risk index and generate warning information based on the joint feature vector and average causal effect. And a collaborative intelligent agent, configured to: respond to scene-triggered events, dynamically orchestrate the execution flow of the multiple dedicated intelligent agents using a directed acyclic graph, and dynamically allocate tasks and schedule resources for the multiple dedicated intelligent agents according to real-time data status, computing load or task changes.

[0007] The beneficial effects of this invention lie in the fact that the public health early warning system based on a composite multimodal large model provided by this invention achieves refined assessment and personalized early warning of public health risks by integrating meteorological, environmental, and health data through a multimodal large model. Addressing the limitations of hard-coded rules in industrial intelligent agents, which cannot adapt to complex scenarios, this invention introduces a multi-agent reinforcement learning and directed acyclic graph dynamic orchestration mechanism. This enables the system to adaptively adjust the agent execution process and resource allocation based on real-time data status, scenario semantics, and computing load, handling complex scenarios such as sandstorms concurrent with eye diseases without exhaustively enumerating preset rules. Furthermore, by combining causal modeling intelligent agents, the system effectively eliminates confounding factors between meteorology, pollution, and health, significantly reducing false alarm rates and early warning delays. This represents an upgrade from data-driven to causal inference and from fixed processes to dynamic adaptation, improving the accuracy and intelligence level of public health early warnings. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic architecture diagram of a system according to an embodiment of the present invention.

[0010] Figure 2 This is an illustrative principle of the intelligent agent orchestration process of a system according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0013] like Figure 1 As shown, the public health early warning system based on a composite multimodal large model includes: Multiple specialized intelligent agents, including at least: The data acquisition agent is used to collect multi-source heterogeneous data, including meteorological time-series data, environmental spatial data, and health-related data. The feature fusion agent is used to perform cross-modal attention fusion on the multi-source heterogeneous data using a pre-trained composite multimodal large model, extract deep features and generate a joint feature vector; A causal modeling agent is used to construct a three-layer causal association model based on the joint feature vector. The three layers include at least a meteorological-pollution layer, a pollution-health layer, and a composite causal layer, and the average causal effect is calculated. A pollution source tracing intelligent agent is used to identify pollution sources and diffusion paths based on the joint feature vector and average causal effect. A risk warning intelligent agent is used to calculate a comprehensive health risk index and generate warning information based on the joint feature vector and average causal effect. And a collaborative intelligent agent, configured to: respond to scene-triggered events, dynamically orchestrate the execution flow of the multiple dedicated intelligent agents using a directed acyclic graph, and dynamically allocate tasks and schedule resources for the multiple dedicated intelligent agents according to real-time data status, computing load or task changes.

[0014] In addition, dedicated intelligent agents also include data-enhanced intelligent agents, and corresponding intelligent agents can be added if other needs arise.

[0015] The following is an introduction to each specialized intelligent agent: In this embodiment, the data acquisition agent serves as the system's data entry point, employing a streaming computing architecture to achieve real-time acquisition, cleaning, and standardization of multi-source heterogeneous data. The implementation of this agent is described in detail below with specific steps.

[0016] 1. Data source access configuration The data acquisition agent is pre-configured with access parameters for multiple data sources, including: Meteorological data source: Access to minute-level observation data from automatic weather stations released by meteorological departments, including time-series variables such as temperature, humidity, wind speed, wind direction, and air pressure, in standard CSV or NetCDF format.

[0017] Environmental data sources: Hourly pollutant concentration data from national air quality monitoring stations, including indicators such as PM2.5, PM10, SO2, NO2, CO, and O3, as well as aerosol optical thickness (AOD) grid data retrieved from satellites.

[0018] Health data source: Access the regional health information platform or hospital information system to obtain anonymized electronic medical record data, including structured fields such as consultation time, diagnostic ICD code, patient age, gender, and underlying medical history, as well as unstructured data such as chief complaint text.

[0019] The data acquisition agent uses Apache Kafka as the message middleware, creating an independent topic for each data source and setting up corresponding consumer groups. For meteorological data with high real-time requirements, a second-level polling retrieval mode is used; for batch-updated health data, an incremental synchronization mode is used.

[0020] 2. Real-time data stream processing The data acquisition agent has a built-in Apache Flink-based streaming engine that performs the following processing steps on each incoming data record: (1) For numerical data (such as temperature and pollutant concentration), the data acquisition agent uses a combination of statistical methods and physical rules to identify outliers: Physical boundary rules: Preset reasonable value ranges for each variable, such as temperature between -50℃ and 60℃, and PM2.5 concentration between 0 and 1000 μg / m³. Data outside these ranges are directly marked as abnormal and entered into the processing queue.

[0021] Statistical outlier detection: For data points that are within the physical boundaries but significantly deviate from the historical distribution, a sliding window is used to calculate the mean and standard deviation of the current data. When a data point deviates from the mean by more than a preset multiple (such as 3 times the standard deviation), it is identified as an outlier.

[0022] Detected outliers are not deleted directly, but are marked with confidence scores and then submitted to subsequent processing decisions (such as whether interpolation or resampling is needed).

[0023] (2) For data points that are missing or marked as anomalous, the data acquisition agent selects a completion strategy based on the data type and missing pattern: Missing time-series data: For missing points in the time series of meteorological or pollutant concentration data, linear interpolation or prediction methods based on historical patterns are used to fill in the gaps. For example, if PM2.5 data for a certain hour is missing at a monitoring station, the weighted average of the data from the two hours before and after that station, or the median of the data at the same time over the past seven days, is used.

[0024] Spatial data gaps: For spatial gaps in gridded environmental data, inverse distance weighted interpolation or kriging interpolation is used to estimate the value of the missing grid using measurements from nearby monitoring stations.

[0025] Missing text data: For unstructured text in medical records, if the missing field does not affect the core diagnostic information, it is left blank; if it is a critical field, a default value (such as "unknown") is generated according to the context rules.

[0026] (3) The cleaned data needs to be converted into a unified spatiotemporal reference: Time alignment: The timestamps of all data are unified to UTC+8 time zone, and data with different sampling frequencies are aligned to the same time granularity (e.g., hourly) through resampling. Meteorological minute data are averaged hourly, and health data are aggregated according to the hour of medical visit.

[0027] Spatial coding: Geographic coordinate information is appended to each data point. For data from monitoring stations, the station's latitude and longitude are used; for grid data, the coordinates of the grid's center point are used; for individual health data, it is converted to latitude and longitude coordinates based on the patient's residential address using a geocoding service.

[0028] Data structuring: The processed data is encapsulated into data packets of a uniform format. Each data packet contains the following fields: {data_id, timestamp, lng, lat, data_type, value, confidence_score, source}. The confidence_score is the confidence score evaluated by the data acquisition agent based on the anomaly detection and completion process. Its value ranges from 0 to 1, with an initial value of 1, which is decreased by the corresponding amount with each correction.

[0029] 3. Data Output and Persistence The standardized data packets are pushed to the downstream feature fusion agent through a message queue, and are also written to a time-series database (such as InfluxDB) or a data lake for persistent storage, for subsequent offline analysis and model training.

[0030] The data acquisition agent also provides a data quality monitoring interface, which outputs indicators such as flow rate, missing rate, and anomaly rate of each data source in real time, for the collaborative agent to make global scheduling decisions.

[0031] The feature fusion agent is responsible for performing cross-modal deep fusion of multi-source heterogeneous data (meteorological time series, environmental space, health text / numerical data) acquired by the data acquisition agent, and outputting a unified joint feature vector. The specific implementation includes the following steps.

[0032] Step 1: Heterogeneous Data Vectorization and Embedding For the original data of different modalities, the following adaptation processing is performed: Meteorological time series data: Let the sequence of a certain meteorological variable (such as temperature, humidity, wind speed) within a time window be... It can be mapped to sequence fragment embeddings through linear projection or convolutional layers. At the same time, learnable positional encoding is added. To preserve the chronological order.

[0033] Environmental spatial data: This includes pollutant concentration values ​​and their geographic coordinates at monitoring stations. Joint encoding. Coordinates are converted into high-dimensional embeddings using sine / cosine functions and concatenated with concentration values ​​to form spatial embeddings. .

[0034] Health-related data: Unstructured text (such as chief complaint) is segmented using a tokenizer to obtain a token sequence. Each token is converted into a vector using a pre-trained embedding matrix. Structured numerical values ​​(such as age) are mapped to the same dimension using a linear projection layer. Finally, all health token embeddings are... .

[0035] The multimodal input sequence is obtained by concatenating all the embedded sequences. ,in This represents the total number of tokens.

[0036] Step 2: Cross-modal attention fusion The feature fusion agent loads a complex multimodal large model, which is based on the Transformer architecture and its core is a multi-head cross-modal attention mechanism. For the input sequence... Calculate the query, key, and value matrix:

[0037] in The scaling dot product attention is defined as:

[0038] Multi-head attention combines the outputs of multiple independent attention heads and then performs a linear transformation:

[0039] This mechanism forces the model to learn deep co-occurrence patterns among meteorological, environmental, and health modalities, such as the association between high temperatures and increased ozone concentration, and between ozone concentration and increased asthma consultation rates.

[0040] Step 3: Extraction of spatiotemporal and evolutionary features The model uses positional encoding internally. For spatial coordinate embedding, use timestamp embedding. Encoding time points. The model automatically extracts them using attention weights: Spatiotemporal evolution characteristics: the migration rate of pollutant concentration over time and spatial gradient .

[0041] Chemical reaction chain characteristics: For example, the transformation from primary pollutant NOx to secondary pollutant O3 is captured by the model through time-dependent attention.

[0042] Population health vulnerability characteristics: By combining meteorological and environmental characteristics with population attributes (age, underlying diseases) in health data, sensitivity representations under specific exposure combinations are extracted through cross-attention.

[0043] Step 4: Domain Fine-tuning and Physical Constraints To prevent general-purpose large models from exhibiting illusions in vertical domains, LoRA lightweight fine-tuning is employed. A physics constraint term is added to the training loss.

[0044] in The mean squared error between the predicted value and the true label. The residuals of the atmospheric diffusion equation are:

[0045] here For pollutant concentration, For wind field, The diffusion coefficient is... This is the source term. This constraint ensures that the model output conforms to the laws of physics.

[0046] Step 5: Output joint feature vector Take the hidden state of the last layer of the large model A fixed-dimensional joint feature vector is obtained through average pooling:

[0047] This vector simultaneously contains cross-modal semantic associations and causal clues. The feature fusion agent will... The spatiotemporal labels and confidence scores are stored in the feature database for use by downstream agents.

[0048] The causal modeling agent is responsible for constructing three-layer causal models: weather-pollution, pollution-health, and complex causal models, enabling causal inference from environmental exposure to health outcomes. The specific implementation includes the following three parts.

[0049] 1. Meteorological-Pollution Layer: Construction and Solution of Physical Information Neural Differential Equations 1.1 Spatiotemporal grid discretization Divide the target geographic region into Each of the two-dimensional spatial grid cells is a unit. In time The status is:

[0050] in This refers to the concentration of pollutants (such as PM2.5 and O3). For wind speed components, For temperature, Humidity. Time step. Set according to the data sampling frequency (e.g., minute level).

[0051] 1.2 Construction of Neural Differential Equations The rate of change of pollutant concentration over time is composed of the sum of physical constraint terms and data-driven residual terms:

[0052] The physical constraint terms are based on the atmospheric advection-diffusion equation:

[0053] in For the advection term, For turbulent diffusion term ( (where is the diffusion coefficient) Sources and sinks (such as emissions and sedimentation).

[0054] The data-driven residual term is generated by a fully connected neural network. Parameterization, with the current concentration as the input. and meteorological elements The output is the compensation value for unmodeled physical processes (such as nonlinear chemical reactions and sudden emission sources).

[0055] 1.3 Solving and Training Use ordinary differential equation solvers (such as the fourth-order Runge-Kutta method or Dormand method) Prince method) from initial concentration Integrating yields the predicted trajectory .

[0056] Define a loss function that fits the observed data to the residuals of the physical equations:

[0057] in For physical constraint weights (e.g., take...) After training, it outputs a continuous, physically consistent pollutant concentration field. .

[0058] 2. Pollution-Health Layer: Counterfactual Matching and Attribution Risk Score 2.1 Calculation of Exposure Dose Patient data is extracted from longitudinal electronic health records (EHRs), including at least the time of visit, diagnostic ICD code, age, gender, history of underlying medical conditions, and latitude and longitude of residence. The residential coordinates of each patient are then compared with the aforementioned concentration field. Perform spatiotemporal correlation to calculate the pre-onset period Cumulative exposure dose within 72 hours (e.g., 72 hours):

[0059] 2.2 Propensity Score Matching to Construct Counterfactual Comparisons Propensity score matching (PSM) or coarse-exact matching (CEM) was used to match one or more control individuals for each patient. Matching covariates included: age, sex, BMI, smoking history, season of visit, and socioeconomic status (SES). The required exposure dose for control individuals was determined. Significantly lower than the patient's exposure dose (e.g.) , It can be set to the 25th percentile of the background concentration.

[0060] 2.3 Individual treatment effect and attribution risk score Based on the matching results, the counterfactual incidence rate of patients was estimated using the actual incidence rate of individuals in the control group. Individual treatment effect (ITE) is defined as:

[0061] in As an indicator of disease incidence under actual high exposure, This serves as an indicator of disease incidence when hypothetical exposure is reduced to control levels. Actual disease incidence rate. This represents the patient's actual illness (taken as 1 or 0). The counterfactual probability is estimated from the incidence rate in the control group.

[0062] The attributable risk score is obtained by averaging the individual treatment effects of all patients:

[0063] This score quantifies the excess health risk resulting from pollutant exposure.

[0064] 3. Composite Causal Layer: Decoupling Representation Learning and Global Causal Graph 3.1 Multimodal Coding and Decoupled Representation Meteorological variables respectively Pollutant concentration field Attribution risk score and health outcome variables (e.g., whether the disease has occurred) are encoded to obtain their respective initial latent feature vectors. .

[0065] Decoupled representation learning is employed to decompose each latent feature into common causal factors. and unique noise factors ,satisfy:

[0066] The training objectives include two aspects: Maximize the mutual information of shared causal factors between weather and health:

[0067] This is achieved using InfoNCE loss, where positive sample pairs are from the same time period. and Negative samples are pairs with randomly shuffled times.

[0068] Minimize the mutual information between shared causal factors and idiomatic noise factors:

[0069] This is achieved through adversarial training or mutual information lower bound estimation.

[0070] 3.2 Construction of Global Directed Acyclic Graph Based on extracted common causal factors Causal discovery algorithms (such as NOTEARS, DAG) are employed. Construct a global directed acyclic graph (GNN or PC algorithm) The nodes in the graph include meteorological variables, pollutant variables, and health outcome variables, while the edges represent direct causal drivers. For example, the algorithm outputs:

[0071] It may also include indirect paths via intermediate variables.

[0072] 3.3 Calculation of Average Causal Effect exist Above, respectively using meteorological variables and pollutant variables As a treatment variable, health outcome As an outcome variable, the backdoor criterion is used to identify the set of confounding variables that need adjustment. (That is, all variables on the backdoor path from the processing variable to the outcome variable).

[0073] The average causal effect (ATE) is defined as the expected difference when the intervention variable is used:

[0074] In the formula This indicates an intervention operation. In actual calculations, this is achieved by adjusting the backdoor variable set. :

[0075] Similarly, calculate These two ATE values ​​quantify the pure causal driving force of meteorological conditions or pollutants on health outcomes after excluding confounding variables.

[0076] The final output of the causal modeling agent is a physically consistent pollutant concentration field. (For use by pollution source tracing intelligent agents); Attribution risk score (For risk warning agents to adjust thresholds); Global Directed Acyclic Graph and average causal effect (Provides the risk warning agent with interpretable warning basis).

[0077] The pollution source tracing agent is responsible for identifying the geographical location, emission intensity, and pollutant diffusion path of pollution sources based on the joint feature vector output by the feature fusion agent and the average causal effect provided by the causal modeling agent. The specific implementation includes the following steps.

[0078] 1. Source term inversion based on concentration field 1.1 Constructing a receptor model Let the set of monitoring stations be . Each site In time The measured pollutant concentration was Assuming there are potential sources of pollution in the area. Each source The emission rate The receptor model represents the observed concentration as a linear superposition of contributions from each source:

[0079] in To start from the source Arrive at the station The contribution coefficient (depending on meteorological conditions and source) (receptor geometry) To observe noise.

[0080] 1.2 Principal Component Analysis Based on Positive Deterministic Matrix Decomposition Positive definite matrix factorization (PMF) or nonnegative constrained least squares method is used to analyze the given monitoring data. and weather field Under the given conditions, solve for the emission rate vector. The source and source component spectra. The optimization objective is:

[0081] To improve the spatial resolution of the inversion, the physically consistent concentration field output by the causal modeling agent is utilized. Instead of discrete site data, the entire gridded concentration field is used as input to invert the possible source intensities within each grid.

[0082] 2. Inverse Particle Tracking Based on the Adjoint Method 2.1 Forward Diffusion Model Pollutant diffusion satisfies advection Diffusion equation:

[0083] in For source terms (emission distribution). Given a meteorological field. and diffusion coefficient It can positively simulate concentration distribution.

[0084] 2.2 Construction of the adjoint equation To reverse-locate the source position, an adjoint variable is introduced. Its adjoint equation is:

[0085] in The functional derivative of the observed term (usually taken as the indicator function at the monitoring point). In the reverse time direction (from... arrive Solving the adjoint equation yields... This is proportional to the sensitivity of the observed concentration at that location. Specifically, the source term... With observed concentration The sensitivity between them can be expressed as:

[0086] 2.3 Source Region Identification For each monitoring station or area of ​​interest, the sensitivity field is obtained by solving the adjoint equation. The spatial coordinates corresponding to the maximum sensitivity value are taken as the most likely source region. For observations at multiple time points, the sensitivity fields at each time step are accumulated:

[0087] The high-value areas indicate the spatial location of the pollution source.

[0088] 3. Ranking of source contributions for causal effects Average causal effect provided by the causal modeling agent This characterizes the driving force of different pollutants on health outcomes. The pollution source tracing agent uses this information to weight and rank multiple identified potential sources, prioritizing those that emit "high-causal-effect pollutants".

[0089] Source The types of pollutants emitted are: The corresponding average causal effect is .source The health-weighted contribution is defined as:

[0090] in For source pollutants The emission rates (from the inversion results in step 1). After sorting, the top contributors will be... Individual sources are the main sources of pollution.

[0091] 4. Diffusion path identification Based on the aforementioned forward or reverse simulations, the spatial trajectory of pollutants can be mapped. Specifically, a forward diffusion model is used to simulate tracer particles released from the identified main source, combined with wind field data. Calculate particle trajectories:

[0092] By performing Lagrange tracking on a large number of particles, the evolution of concentration distribution over time can be obtained, thereby depicting the diffusion direction, migration path, and deposition area of ​​pollutants.

[0093] 5. Final output of the pollution source tracing intelligent agent: List of pollution sources: geographical coordinates, emission intensity, types of pollutants emitted, and health-weighted contribution value for each source; Diffusion path: The spatial trajectory of pollutants from the source area to the sensitive area (which can be represented as a series of latitude and longitude points or vector lines). Confidence assessment: Based on the variance of the inversion residuals and the associated sensitivity field, the uncertainty range of each output result is given.

[0094] The risk warning agent is responsible for calculating the comprehensive health risk index (HRI) based on the joint feature vector output by the feature fusion agent and the average causal effect provided by the causal modeling agent, and generating differentiated warning reports and handling suggestions according to the risk level. The specific implementation includes the following steps.

[0095] 1. Extract risk-related features The risk warning agent reads the following inputs from the feature database: Joint eigenvectors This includes meteorology environment Deep semantic features of health; Average causal effect This represents the increase in health outcome risk caused by an increase in the concentration of a single pollutant; Real-time pollutant concentration (From data acquisition agent or concentration field interpolation); Population exposure parameters: including the density of the exposed population. The proportion of vulnerable groups (such as the proportion of the elderly) Children's proportion Prevalence of underlying diseases wait).

[0096] 2. Calculate the comprehensive health risk index For a spatial unit (or a population group) within the region, the comprehensive health risk index is defined as the weighted cumulative risk of each pollutant, with the weights determined by both causal effects and population vulnerability.

[0097] Set up a joint monitoring area Key pollutants (such as PM2.5, O3, NO2, etc.), the first The real-time concentration of the pollutants is (Unit: μg / m³ or ppm), its average causal effect is (This represents the increase in the probability of a health outcome resulting from each unit increase in concentration). Population vulnerability weight. Calculated based on different age groups and underlying medical history:

[0098] in For the preset sensitivity coefficient (e.g.) (This can be adjusted based on local epidemiological data). Exposure population weight. Take the ratio of the population density of the area to the baseline density.

[0099] Comprehensive Health Risk Index The calculation formula is:

[0100] In some scenarios, synergistic effects between pollutants also need to be considered. Optionally, add interactive elements:

[0101] in The causal effects of joint exposure (which can be replaced by interaction effect terms provided by the causal modeling agent). These are the normalization coefficients. To simplify implementation, a linear weighted formula is generally used.

[0102] 3. Dynamic thresholds and risk level classification The risk warning agent maintains a set of dynamic thresholds, which can be adaptively adjusted based on the quantiles of historical risk indices or the upper limit of daily predictions. Common methods include: Four risk levels are defined: low (green), medium (blue), high (yellow), and very high (red). Calculate the past Risk index within 30 days (e.g., within 30 days) quantiles; Current risk index Comparison with quantiles: →Low risk; →Medium risk; →High risk; →Extremely high risk.

[0103] Or preset a fixed threshold according to health department standards (e.g.) ).

[0104] 4. Generate early warning information and handling suggestions The risk warning agent utilizes the generative capabilities of a complex multimodal large model (or template-based dynamic imputation) to generate a natural language warning report. The report content includes: Current risk level, main contributing pollutants and their concentrations; Warnings about vulnerable populations (e.g., "Older adults and children are at higher risk of asthma attacks"); Behavioral suggestions are generated separately for environmental protection departments (e.g., "suggest reducing road sweeping and suspending production at high-emission enterprises"), medical institutions (e.g., "open more respiratory clinics"), and the general public (e.g., "wear masks and reduce outdoor activities").

[0105] Example output format: Warning level: High (Yellow warning); Main risk source: PM 2.5 (concentration 85 μg / m 3 Causal effect 0.012), O3 (concentration 160 μg / m³) 3 (Causal effect 0.009). Affected populations: Attributable risk increased by 28% in people aged 65 and over, and by 15% in children. Recommended measures: Environmental protection departments: implement staggered production in industrial parks; Health Commission: issue health protection tips and open temporary oxygen inhalation points in communities; Public: avoid outdoor exercise in the morning and keep doors and windows closed.

[0106] 5. Results push and closed-loop feedback Early warning information is pushed to environmental protection, health, and other departments, as well as public platforms, via message middleware. Simultaneously, the risk warning intelligent agent records the trigger time, risk index, and actual health outcome (subsequent changes in outpatient volume reported by the hospital system) of this warning in the database. This data is then used by the collaborative intelligent agent for offline evaluation and reward function updates (e.g., if outpatient volume does not increase significantly after the warning, it indicates the warning is timely and effective, and a positive reward is given; conversely, if there is a missed or false alarm, a negative reward is given), enabling the system to self-evolve.

[0107] The data augmentation agent is responsible for receiving data packets with confidence scores below a preset threshold. It improves data quality and updates confidence scores through methods such as missing value completion, noise suppression, or data generation, and then re-injects the processed data packets into the data pipeline. The specific implementation includes the following steps.

[0108] 1. Input data packet parsing The data augmentation agent subscribes to the data pipeline and receives data packets forwarded by the upstream agent (usually a data acquisition agent or a feature fusion agent). Each data packet contains: Payload Actual data (such as meteorological observations, pollutant concentrations, and health text snippets); Metadata: Spatiotemporal tags (timestamps) Geographic coordinates ), data source identifier; Confidence score This characterizes the reliability of the data.

[0109] when (For example When the data packet is routed to the data augmentation agent, the collaborative agent routes the data packet to the data augmentation agent.

[0110] 2. Imputation (Missing Value Completion) If certain fields are missing in the data packet (such as the PM2.5 value of a certain monitoring station for a certain hour is missing), the data augmentation agent will fill them in using spatiotemporal interpolation or historical pattern prediction.

[0111] 2.1 Spatiotemporal interpolation (e.g., Kriging) Set missing position The value to be estimated is Utilize the neighborhood 1 effective monitoring point concentration value Construct a weighted average:

[0112] Weight This was obtained by solving the Kriging equations, which are based on the variogram. Describes spatial correlation. In simple cases, inverse distance weighting (IDW) can be used:

[0113] in .

[0114] 2.2 Historical Pattern Prediction For missing time series data, a simple cyclic prediction model (such as exponential smoothing or ARIMA) can be trained using historical data from the same location. Let the concentration at the same historical time be... Predict the current value:

[0115] in The smoothing parameter can be optimized to 0.3–0.7. For pollutants with obvious periodicity (such as diurnal variations in O3), a seasonal decomposition method can be used.

[0116] 3. Noise Suppression (Filtering) When there are obvious abnormal fluctuations in the observations in the data packet (such as instantaneous spikes or negative values), noise suppression is performed.

[0117] 3.1 Sliding Median Filtering For length of (As shown in 5) A sliding window replaces the value at that moment with the median of the sequence within the window:

[0118] This method is robust to impulse noise and preserves edges.

[0119] 3.2 Kalman Filter (Suitable for Time Series Smoothing) Constructing a state-space model:

[0120] in This represents the actual state (such as the actual concentration). The observed values ​​are used. Kalman filtering iteratively updates the predictions and corrections, outputting a smoothed state estimate. It can also provide the uncertainty of the estimate (covariance matrix).

[0121] 4. Data Generation When data is severely missing or cannot be recovered through interpolation / filtering (e.g., all monitoring stations in a certain area go offline at the same time), the data-augmented agent initiates a generative model to synthesize data based on the context of nearby time and region.

[0122] 4.1 Synthesis Based on Generative Adversarial Networks (GANs) A conditional generative adversarial network is pre-trained, taking spatiotemporal context features (such as the concentration of neighboring grid cells and weather conditions) as input and generating the target concentration value as output. Generator With discriminator Game Theory Training:

[0123] in These are conditional vectors (such as time encoding or spatial coordinates). During inference, the contextual features of the missing locations are input into the trained generator to obtain synthetic data.

[0124] 4.2 Simple Generation Based on Time Series Prediction For stationary processes, exponential smoothing or Prophet models can be used, extrapolating a window based on observations from previous time steps. For example, a cubic exponential Holt model can be used. The Winters model predicts values ​​for the next few hours, which are then used to fill in missing paragraphs.

[0125] 5. Confidence score update After processing, the data augmentation agent evaluates the new confidence score based on the reliability of the processing method and the residuals. Update rules:

[0126] in depending on: Interpolation / filtering method: If the variance of the interpolation residual is less than the threshold, If it is greater than the threshold, (To prevent the situation from worsening).

[0127] Generative methods: If the generated samples have high discriminator scores (i.e., high realism). ;otherwise (No improvement).

[0128] Alternatively, a probability expression can be used:

[0129] Where quality is a comprehensive score of the interpolation residuals or the confidence of the generative model.

[0130] 6. Output data packets The data-augmented agent will process the payload. Original metadata and updated confidence scores The data is repackaged into a data packet and published through the data pipeline to the intended downstream agent (i.e., the target to which it was originally sent when the compensation path was triggered). Simultaneously, the "enhancement history" (operation type, timestamp) is recorded in the data packet's metadata for subsequent auditing or model training.

[0131] In one embodiment of the present invention, please refer to Figure 2 The collaborative agent acts as the "commander" of the entire system, responsible for building data pipelines between agents, reconstructing dynamic directed acyclic graph (DAG) topology, scheduling resources based on multi-agent reinforcement learning, and monitoring and self-healing the system. The specific implementation includes the following four core modules.

[0132] 1. Responsive streaming data pipeline and backpressure mechanism 1.1 Data Pipeline Construction Collaborative agents construct data channels between agents based on reactive streaming protocols (such as the ReactiveStreams specification). Each specialized agent (data acquisition, feature fusion, causal modeling, etc.) simultaneously implements both Publisher and Subscriber interfaces, forming an asynchronous, non-blocking data stream network. The data packets transmitted within the channel are confidence-based packets. AwareDataPacket (CADP), structured as follows:

[0133] The payload consists of actual data (such as pollutant concentration values ​​and feature vectors), while the metadata includes spatiotemporal labels and source information. This is the confidence score.

[0134] 1.2 Back pressure mechanism Each agent monitors its own processing capacity (e.g., the number of data packets it can process per unit time) and the upstream transmission rate in real time. Let the upstream transmission rate be... Downstream processing capacity is When the processing capacity of the downstream agent is lower than the upstream transmission rate ( When this occurs, the downstream sends a back pressure signal to the upstream via a feedback channel, dynamically adjusting the upstream transmission rate. The adjustment rules are as follows:

[0135] in It uses a sigmoid-type smoothing function to ensure a smooth rate decrease. This mechanism prevents data backlog from causing system crashes and achieves traffic shaping.

[0136] 2. Confidence-driven dynamic topology reconstruction 2.1 Low-confidence trigger compensation path The collaborative agent continuously monitors the confidence score of data packets flowing through each node. When a data packet Below the preset threshold (For example And if the system load does not exceed the threshold, the collaborative agent performs topology reconfiguration: Identify the current node: The agent that detects the low-confidence data packet (such as a feature fusion agent) is the current node.

[0137] Establish a compensation path: Dynamically establish a directed edge (compensation path) between the current node and the preset data augmentation agent. This operation is achieved by modifying the DAG structure in memory without requiring a system restart.

[0138] Routing data: forward low-confidence data packets to the data-enhanced agent via a compensation path.

[0139] The data augmentation agent performs missing value completion, noise suppression, or data generation processing (see above for specific implementation details) and generates updated confidence scores for data packets. After processing, the data packet is re-injected into the data pipeline and continues to flow along the original downstream path.

[0140] 2.2 Probability of dropping under high load When the overall system load (such as CPU / GPU utilization) exceeds a preset threshold (For example When a data packet is received, the cooperative agent initiates a probabilistic discarding mechanism. For each incoming data packet, the probability of it being retained is determined. Positively correlated with confidence score:

[0141] in For the discard threshold (e.g.) The lower the confidence level, the higher the probability of discarding (i.e., ...). This strategy ensures that valuable computing resources are prioritized for high-value data.

[0142] 3. Multi-agent reinforcement learning scheduling 3.1 State Space and Input The collaborative agent employs a multi-agent reinforcement learning framework (such as MAPPO). Input state vector. Include: Resource utilization rate: ; Data flow rate: Current system throughput ; Task queue length: Number of tasks to be processed by each agent ; Model accuracy metrics: the prediction error (e.g., MAE) of each model over a recent period. Scene triggering events: natural language commands input by the user or scene types automatically detected by the system (encoded as one-hot vectors or semantic embeddings).

[0143] 3.2 Layered Action Space The Multi-Agent Reinforcement Learning (MARL) framework employs a centralized training, decentralized execution (CTDE) architecture. The policy network (Actor) takes the system state as input and outputs hierarchical actions. ,in For macroscopic topological actions, This refers to micro-level resource actions. The following details the definition of the action space, action output methods, and the execution mechanism of actions in the system.

[0144] (1) Macroscopic topological actions Definition and output Macro-level topological actions modify the structure of the current computational directed agglomerate (DAG). This graph uses specialized agents as nodes and data dependencies or control flows as directed edges. The action space is a compositional space containing three sub-dimensions: Agent activation vector ,in The total number of specialized intelligent agents available in the system (e.g., data acquisition, feature fusion, physical diffusion model, chemical diffusion model, ophthalmic monitoring, respiratory risk assessment, causal modeling, pollution source tracing, risk warning, data augmentation, etc.). Indicates the first One agent is activated. This indicates that the message has been blocked.

[0145] Data flow matrix ,in Indicates from the intelligent agent To intelligent agents There are directed edges (data flow direction). The matrix must satisfy the acyclic constraint.

[0146] Execution order constraints: implicitly expressed through topological sorting, or explicitly output as a priority vector. The smaller the value, the higher the execution priority.

[0147] To reduce the dimensionality of the action space, a parameterized topology generator is used in the actual implementation: the policy network outputs a latent vector. Then, it is transformed into a specific graph by a differentiable topological decoder (such as a graph generator based on Gumbel-Softmax). Alternatively, the action space can be predefined as several discrete modes (Mode 1: chemical analysis mode; Mode 2: physics + ophthalmology mode; Mode 3: pure data monitoring mode, etc.), and the macroscopic topological action is then derived from... Choose one from several modes. This implementation uses discrete mode selection to simplify engineering implementation.

[0148] Suppose the pattern library contains A predefined computation execution graph template (e.g.) The policy network outputs a probability distribution. :

[0149] Macro topological actions Pattern index is obtained by sampling according to this distribution. Each pattern corresponds to a fixed set of activated agents, data flow, and execution order (pre-designed by domain experts). For example: Mode 1 (Common Pollution): Activate the agent {data acquisition, feature fusion, chemical diffusion model, respiratory risk, risk warning}, with the flow being data acquisition → feature fusion → chemical diffusion → respiratory → warning.

[0150] Mode 2 (Dust Storm / Ophthalmology): Activate the agent {data acquisition, physical diffusion model, ophthalmology anomaly detection, feature fusion, comprehensive early warning}, with the flow being: data acquisition → physical diffusion; data acquisition → ophthalmology monitoring; output of physical diffusion and ophthalmology monitoring → feature fusion → comprehensive early warning (parallel branch).

[0151] (2) Micro-resource actions Definition and output Micro-level resource actions are responsible for allocating computing resources to each activated agent or triggering operational instructions. Resource allocation uses continuous variables, while operational instructions are discrete variables. Specifically, these include: Compute resource allocation vector ,in This indicates that the assignment is made to the intelligent agent. Resource allocation (such as GPU memory usage, CPU core count, thread pool size). Typically, constraints are placed on... (Total resources normalization).

[0152] Operation and maintenance instructions ,in .

[0153] The policy network simultaneously outputs micro-resource actions. To handle the mixed action space, the following approach can be used: For continuous resource allocation, the policy network outputs a vector. Resource share is obtained after Softmax normalization:

[0154] For discrete operation and maintenance commands, the policy network outputs logits. The probability distribution is obtained through softmax:

[0155] The actual instructions executed Obtained by sampling or taking the maximum value.

[0156] Ultimate micro-resource actions .

[0157] (3) Action execution mechanism When the policy network selects a mode Then, the collaborative agents perform the following steps to reconstruct the graph: Cancel the existing subscription relationship: Traverse all edges in the current computation execution graph and call the unsubscribe() method of each agent to disconnect the data pipeline.

[0158] Load target mode configuration: Read the mode from the configuration server or local cache. The corresponding list of activated agents, the list of data flow edges, and the parallel / serial flag.

[0159] Establish a new subscription relationship: Call the `subscribe(downstream)` method of the upstream agent sequentially according to the edge list. For parallel branches, use `subscribe` and set parameters such as `bufferSize`.

[0160] Start an idle agent: For a newly activated agent that was previously dormant, call its start() method to allocate initial resources.

[0161] Deactivate redundant agents: For agents that are no longer needed in the pattern, call the stop() method to release resources.

[0162] The entire refactoring process requires a latency of less than 100ms and is executed in an asynchronous, non-blocking manner to avoid blocking the data flow.

[0163] Resource quota adjustment: Collaborative agents set resource limits for each agent process or thread through containerized management interfaces (such as cgroups, Docker API, or KubernetesHPA). For example, for a feature fusion agent, the GPU memory limit is set to [value missing]. If resource adjustments require restarting the agent, a graceful shutdown and restart strategy will be adopted.

[0164] Operation and maintenance command execution: like If the data acquisition agent is not found, a RESAMPLE control message is sent to it, requesting it to increase the sampling frequency or switch to a backup data source.

[0165] like If so, a RETRAIN message is sent to the causal modeling agent or risk warning agent to trigger asynchronous model retraining (using the most recent incremental data).

[0166] like This requires specifying a hot switch of the agent to a backup model version.

[0167] All resource adjustments and command executions are sent via a message bus (such as RedisPub / Sub or gRPC streams) and the execution time and results are recorded for subsequent reward calculations.

[0168] (4) Policy learning combined with reward function The training objective of the policy network is to maximize the cumulative discount reward. Reward function. As defined previously:

[0169] in This represents the total time (in seconds) from when the data arrives to when the warning is issued. The cost of computing resources used by all agents in the current step (e.g., GPU hours × unit cost). The recent early warning accuracy rate (calculated based on post-outpatient volume feedback).

[0170] The Actor and Critic networks are updated using Proximal Policy Optimization (PPO). For hierarchical actions, hierarchical policy gradients can be used: the gradients of macro-actions and micro-actions are calculated separately, or the underlying feature extractor is shared. A simplified implementation can combine these into a single joint action vector for unified processing.

[0171] In one example, switching from normal pollution mode to sandstorm mode: Assuming the initial state is normal pollution mode (mode 1), the system detects normal resource utilization (CPU 30%), but the scene trigger event changes to "sandstorm emergency response". Policy network input state. This includes: meteorological characteristics (surge in PM10, low visibility), health data (increased ophthalmology appointment rates), and resource indicators. Network output probability distribution. Among them, mode 2 (dust storm / ophthalmology) had the highest probability. Sampling yielded... Meanwhile, for micro-level action output: allocate 0.25 GPU share to the physical diffusion model, 0.15 to ophthalmic monitoring, reduce the chemical diffusion model to 0.05, and keep the remaining allocation unchanged.

[0172] The collaborative agent performs topology reconstruction: disconnecting the chemical diffusion model, activating the physical diffusion model and ophthalmic monitoring, and establishing parallel data flows. Simultaneously, resource quotas are adjusted. The entire process is completed within 50ms, and the system seamlessly switches to the new computational execution graph.

[0173] 4. Monitoring and Closed-Loop Feedback 4.1 Data missing triggers re-sampling Collaborative intelligent agents monitor and collect data, and the data missing rate of the intelligent agent output is as follows: .like (For example And data confidence level (For example If this occurs, a data re-acquisition command is triggered. This command is issued via the message bus, and upon receiving it, the data acquisition agent switches to a high-frequency acquisition mode or activates a backup data source (such as satellite remote sensing data).

[0174] 4.2 Model Drift Detection and Retraining The collaborative intelligent agent performs streaming monitoring of the output residuals of the causal modeling intelligent agent and the risk warning intelligent agent. Page is used. The Hinkley test (PHT) detects conceptual drift. Let the residual sequence... Cumulative statistics:

[0175] in For drift threshold (e.g.) ).like If a concept drift occurs, asynchronous model retraining is triggered. The retraining process retrieves incremental data from the feature library over the past 7 days, initiates the LoRA fine-tuning process, and then hot-updates the model (seamlessly switching versions) upon completion.

[0176] 4.3 Fault Tolerance and Self-Healing The collaborative agent maintains the heartbeat status of each agent. If an agent loses three consecutive heartbeats, it is automatically removed from the execution graph, and an attempt is made to restart or switch to a backup agent. Simultaneously, abnormal events are logged for offline analysis.

[0177] The collaborative agent does not directly output early warnings or tracing results, but instead outputs control commands (re-sampling commands, model retraining commands, and topology modification commands) and resource allocation strategies. Through the above implementation method, the collaborative agent achieves closed-loop management across the entire chain, from data flow control, dynamic topology orchestration, reinforcement learning scheduling to monitoring and self-healing.

[0178] In one application embodiment, a method for issuing public health warnings for complex extreme weather events includes: Step S101: Input of multi-source heterogeneous data and semantic parsing of the scene.

[0179] Three types of real-time data streams are acquired in parallel by a data acquisition agent: Meteorological data stream: includes time-series data on wind speed, wind direction, temperature, humidity, and visibility; Environmental monitoring data stream: includes particulate matter (PM2.5, PM10), gaseous pollutants (SO2, NO2, O3), and pollen / allergen concentration data; Public health data streams include hospital ophthalmology and respiratory outpatient volumes, electronic medical record texts, and diagnostic ICD codes.

[0180] Simultaneously, it receives natural language commands input by the user, such as "Activate the sandstorm emergency response, focusing on the risk of respiratory diseases and allergic conjunctivitis." Using a multimodal large model (MLLM), it performs semantic embedding and entity recognition on the input commands, parsing out the core scene labels as "sandstorm" and "allergy," thereby determining that the current situation is a complex physical-biological exposure scenario.

[0181] Step S102: Dynamic topology reconstruction of DAG based on reinforcement learning.

[0182] The Coordinator Agent generates a dynamic computation graph topology based on the scene labels parsed in step S101 using a Multi-Agent Reinforcement Learning (MARL) policy network. Unlike the default chemical pollution analysis pipeline, this embodiment automatically activates dedicated operators for physical processes, constructing the following execution pipeline: The first parallel path (meteorological-physical branch): The meteorological data stream is input into the physical diffusion model, which solves the transport and sedimentation trajectories of dust particles at different particle sizes based on the Neural ODE.

[0183] The second parallel path (environment-biology branch): inputs the environmental monitoring data stream to the allergen concentration analysis module to quantify the exposure level of biological allergens such as pollen.

[0184] Step S103: Cross-modal feature fusion.

[0185] The spatiotemporal distribution feature vector of dust output from the first parallel path and the allergen concentration feature vector output from the second parallel path are aligned through cross-modal attention in the feature fusion agent to generate a unified joint feature representation. .

[0186] Step S104: Three-layer causal relationship modeling and health risk quantification.

[0187] The causal modeling agent performs hierarchical causal inference based on the joint feature representation generated in step S103: Meteorological-Pollution Layer: Using a Neural ODE constrained by Physical Information Neural Networks (PINNs), the driving effect of dust storm meteorological conditions on particulate matter concentration was simulated; Pollution-Health Layer: Using the hospital's longitudinal electronic health records (EHR) to construct individual-level counterfactual twins, calculate individualized treatment effects (ITE), and quantify the attribution risk (AR) of dust and allergens for conjunctivitis and respiratory diseases. Composite causal layer: By decoupling representation learning, the interference of confounding variables such as sudden temperature drops is removed, and a global causal graph is constructed.

[0188] Step S105: Differentiated comprehensive early warning output.

[0189] Based on the attribution risks quantified in step S104, the risk warning intelligent agent generates tiered warning information by combining current resource load and data confidence. For example, it provides control recommendations to environmental protection departments such as "suggest stopping outdoor construction and focusing on watering to suppress dust"; and provides health tips to the public such as "Orange warning for sandstorm: the risk of seeking medical attention for conjunctivitis is extremely high in the next 6 hours; it is recommended to wear goggles and reduce outdoor activities."

[0190] Preferably, in step S102, the DAG dynamic topology reconstruction further includes an anomaly handling mechanism: when the data acquisition agent detects that the data missing rate of the monitoring station exceeds the threshold, the collaborative agent dynamically cuts off the path to the pollution source tracing agent by issuing control commands, and instead activates the generative data augmentation agent to perform virtual data supplementation, and switches back to the original link after the data confidence is restored.

[0191] Preferably, in step S104, the counterfactual reasoning engine forces a strong causal relationship between the "high-pressure ridge" feature in the meteorological mode and the "hospitalization rate" feature in the health mode in the low-dimensional latent space by maximizing mutual information (MI), while suppressing the interference of confounding factors such as "day of the week".

[0192] This example demonstrates that this solution does not use a rigid "application scenario definition module," but rather internalizes scenario adaptation into an end-to-end dynamic capability encompassing input understanding, process orchestration, and model inference. This semantic and data-driven adaptation is better suited to handle complex and ever-changing real-world business environments than rule-table-based matching.

[0193] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A public health early warning system based on a composite multimodal large model, characterized in that, include: Multiple specialized intelligent agents, including at least: The data acquisition agent is used to collect multi-source heterogeneous data, including meteorological time-series data, environmental spatial data, and health-related data. The feature fusion agent is used to perform cross-modal attention fusion on the multi-source heterogeneous data using a pre-trained composite multimodal large model, extract deep features and generate a joint feature vector; A causal modeling agent is used to construct a three-layer causal association model based on the joint feature vector. The three layers include at least a meteorological-pollution layer, a pollution-health layer, and a composite causal layer, and the average causal effect is calculated. A pollution source tracing intelligent agent is used to identify pollution sources and diffusion paths based on the joint feature vector and average causal effect. A risk warning intelligent agent is used to calculate a comprehensive health risk index and generate warning information based on the joint feature vector and average causal effect. And a collaborative intelligent agent, configured to: respond to scene-triggered events, dynamically orchestrate the execution flow of the multiple dedicated intelligent agents using a directed acyclic graph, and dynamically allocate tasks and schedule resources for the multiple dedicated intelligent agents according to real-time data status, computing load or task changes.

2. The system according to claim 1, characterized in that, The feature fusion agent is configured as follows: The meteorological time-series data in the multi-source heterogeneous data is converted into sequence fragments through sliding window slicing; the environmental spatial data is converted into geographic coordinate embedding; and the unstructured text in the health-related data is processed by a word segmenter, while the structured numerical data is mapped into high-dimensional vectors through a linear projection layer. The sequence fragments, environmental data embedded with geographic coordinates, and the high-dimensional vector are used as inputs to the composite multimodal large model. The cross-modal attention mechanism of the composite multimodal large model is used to embed and map the inputs of different modalities into a unified semantic space. Spatiotemporal evolution features, pollutant chemical reaction chain features, and population health vulnerability features under specific meteorological and environmental combinations are extracted through location encoding and timestamp embedding. A lightweight fine-tuning technique is adopted and a physical constraint loss function is introduced. Historical labeled data is used to adapt the composite multimodal large model to the vertical domain so that the features extracted by the model conform to the physical laws of atmospheric chemistry and epidemiology. The hidden layer states or attention maps of the composite multimodal large model are output, and after pooling operations, a joint feature vector of fixed dimensions is generated and stored in the feature database.

3. The system according to claim 2, characterized in that, By embedding location coding and timestamps, spatiotemporal evolution features, pollutant chemical reaction chain features, and population health vulnerability features under specific meteorological and environmental combinations are extracted, including: The spatial coordinates are embedded using the location encoding built into the composite multimodal large model, and the temporal information is embedded using timestamps to extract the evolution law of pollutant concentration with time and space; the evolution law includes the diffusion direction, migration rate and accumulation area of ​​pollutants. Using the cross-modal attention mechanism of the aforementioned complex multimodal large model, the temporal dependence between different pollutant concentration sequences is analyzed, and the chemical reaction chain characteristics between pollutants are extracted; the chemical reaction chain characteristics include the quantitative characterization of the pathway and reaction rate from primary pollutant to secondary pollutant. The meteorological data and environmental data in the multi-source heterogeneous data are jointly encoded with the population attribute information extracted from health-related data. The composite multimodal large model is used to extract population health vulnerability characteristics under specific meteorological conditions and environmental pollutant combinations. The health vulnerability characteristics include the sensitivity of different age groups or people with underlying diseases to composite exposure.

4. The system according to claim 1, characterized in that, The causal modeling agent is configured as follows: The target region is discretized into a spatiotemporal grid, and a physical information neural differential equation is constructed. In the physical information neural differential equation, the pollutant concentration change rate is jointly determined by the physical constraint term based on the atmospheric diffusion equation and the neural network residual term. Solving the physical information neural differential equation yields a physically consistent pollutant concentration field. Based on the concentration field, the individual's pollutant exposure dose is calculated. Combined with longitudinal electronic medical records, a counterfactual control is constructed for each patient using a matching method. The individual treatment effect is calculated and weighted averaged to output the attribution risk score. Decoupled representation learning is employed to extract common causal factors from meteorological variables, pollutant concentration fields, and attribution risk scores. A global directed acyclic graph is constructed based on these common causal factors, and the average causal effect from meteorology to health outcomes and the average causal effect from pollutants to health outcomes are calculated on the global directed acyclic graph.

5. The system according to claim 4, characterized in that, The target region is discretized into a spatiotemporal grid, and physical information neural differential equations are constructed, including: The target area is divided into several two-dimensional spatial grid units, and the state of each grid unit at each time step includes pollutant concentration value and meteorological element value. A neural differential equation is constructed to describe the rate of change of pollutant concentration over time within each grid cell. The rate of change is composed of two superimposed parts: the first part is a physical constraint term determined by the advection-diffusion equation based on atmospheric physics, which characterizes the known dynamic processes of pollutant transport by wind and turbulent diffusion; the second part is a data-driven residual term fitted by a neural network, which is used to compensate for unmodeled physical processes, sudden emission sources or nonlinear chemical reactions present in actual observations. The physical constraint term is added to the data-driven residual term to form the global right-hand function of the neural differential equation.

6. The system according to claim 4, characterized in that, Based on the concentration field, the individual's pollutant exposure dose is calculated. Combined with longitudinal electronic medical records, a counterfactual control is constructed for each patient using a matching method. The individual treatment effect is calculated and weighted averaged, outputting an attribution risk score, including: Acquire longitudinal electronic medical record data, which includes at least the patient's consultation time, diagnosis result, age, gender, and history of underlying diseases; The coordinates of each patient’s residence are spatiotemporally correlated with the pollutant concentration field to calculate the cumulative pollutant exposure dose for each patient within a preset time window before the onset of illness. Propensity score matching or coarse-to-precise matching was used to match one or more control group individuals for each patient. The control group individuals were similar to the patients in observable covariates such as age, gender, season of visit, and history of underlying diseases, and the exposure dose of the control group individuals was lower than the preset threshold of the patient's exposure dose. Based on the matching results, the counterfactual incidence probability of each patient when their exposure dose is reduced to the level of the control group is estimated using the actual incidence of the individual in the control group. The individual treatment effect is calculated as the difference between the incidence probability at the actual exposure level and the counterfactual incidence probability. The individual treatment effects of all patients are averaged, and the attribution risk score is output.

7. The system according to claim 4, characterized in that, Decoupled representation learning is employed to extract common causal factors from meteorological variables, pollutant concentration fields, and attribution risk scores. A global directed acyclic graph (DAG) is constructed based on these common causal factors, and the average causal effect from meteorological conditions to health outcomes and from pollutants to health outcomes are calculated on this DAG, including: The meteorological variables, the pollutant concentration field, the attribution risk score, and the health outcome variables are encoded respectively to obtain their respective initial potential feature vectors; A decoupled representation learning method is adopted to split the initial latent feature vector into two mutually orthogonal parts: a common causal factor and a specific noise factor. Specifically, by maximizing the mutual information of the common causal factor between meteorological variables and health outcome variables, and simultaneously minimizing the mutual information between the common causal factor and the specific noise factor, the common causal factor carries shared causal information across modalities, while the specific noise factor carries interference information unique to each modality and unrelated to causality. Based on the extracted common causal factors, a global directed acyclic graph spanning meteorology, environment, and health is constructed using a causal discovery algorithm. The global directed acyclic graph contains direct causal edges from meteorological variables to pollutant variables and from pollutant variables to health outcome variables, as well as possible indirect causal paths transmitted through intermediate variables. On the global directed acyclic graph, meteorological and pollutant variables are used as treatment variables, and health outcome variables are used as outcome variables. The backdoor criterion is used to identify and block all backdoor paths. The average causal effect of the treatment variables on the outcome variables after eliminating the interference of confounding variables is calculated, including the average causal effect from meteorology to health outcome and the average causal effect from pollutants to health outcome.

8. The system according to claim 1, characterized in that, The collaborative intelligent agent is configured as follows: A data pipeline is constructed between agents based on a responsive streaming protocol. Each agent acts as both a publisher and a subscriber, and data packets with confidence scores are transmitted within the pipeline. When the processing capacity of a downstream agent is lower than the upstream transmission rate, a backpressure mechanism is triggered to reduce the upstream data transmission rate. When the load does not exceed the preset threshold, if the confidence score of a data packet is lower than the preset threshold, the directed acyclic graph topology is reconstructed, and the low-confidence data packets are routed to the preset data augmentation agent for processing. The processed data packets carry the updated confidence score and re-enter the data pipeline to continue flowing downstream. When the load exceeds the preset threshold, low-confidence data packets are discarded first. The probability of a data packet being discarded is negatively correlated with the confidence score; the lower the confidence score, the higher the probability of discarding. A multi-agent reinforcement learning framework is adopted, taking real-time monitored resource utilization, data flow rate, task queue length, model accuracy index and scene triggering events as input, and outputting hierarchical actions. The hierarchical actions include dynamically allocating tasks and scheduling resources for each dedicated agent. The monitoring data acquisition agent tracks the data missing rate and the prediction residuals output by each model. When the data missing rate exceeds a threshold, a data re-acquisition instruction is triggered. When the mean of the prediction residuals shifts statistically significantly, asynchronous model retraining and hot update are triggered.

9. The system according to claim 8, characterized in that, Reconstructing the topology of a directed acyclic graph includes: Obtain the confidence score carried by the data packet, and compare the confidence score with a preset confidence threshold; When the confidence score is lower than the confidence threshold, a topology reconstruction operation is triggered by the collaborative agent to establish a compensation path from the current node to the preset data augmentation agent in the original directed acyclic graph; the current node is the agent that detects low-confidence data packets with a confidence score lower than the confidence threshold. The low-confidence data packets are routed to the data augmentation agent via the compensation path to perform missing value completion, noise suppression, or data generation processing on the low-confidence data packets; After processing, an updated confidence score is generated for the low-confidence data packet, and the low-confidence data packet is re-injected into the data pipeline to continue flowing along the original downstream path.

10. The system according to claim 8, characterized in that, Employing a multi-agent reinforcement learning framework, this approach takes real-time monitored resource utilization, data flow rate, task queue length, model accuracy metrics, and scene-triggered events as input, and outputs hierarchical actions, including: A multi-agent reinforcement learning framework is adopted, with real-time resource status and scene semantic features as inputs; the real-time resource status includes resource utilization, data flow rate, task queue length and model accuracy indicators; the scene semantic features are derived from scene-triggered events input by the user. The multi-agent reinforcement learning framework outputs hierarchical actions, which include macro-topological actions and micro-resource actions: The macro-topology action is used to select or adjust the directed acyclic graph structure of the current computation execution graph, including determining the target dedicated agent to be activated, setting the data flow direction and execution order between the target dedicated agents; The micro-resource actions are used to allocate computing resource ratios to each dedicated intelligent agent, or to trigger data re-sampling and model retraining. The computation execution graph is reconstructed based on the macroscopic topological actions to switch from the default execution graph to a target computation execution graph optimized for the current scenario semantics and resource state; wherein, the nodes of the computation execution graph are dedicated intelligent agents, and the edges are data dependencies or control flows.

Images