Intelligent influenza early warning system based on community multi-modal data fusion

Abstract

The invention relates to the technical field of infectious disease monitoring and early warning, and discloses an intelligent influenza early warning system based on community multi-modal data fusion. The community-level multi-modal data fusion architecture is constructed, medical health data, environmental data, crowd activity data and network behavior data are integrated, spatial-temporal features are dynamically extracted and fused in combination with a deep learning model, and the problem of community monitoring blind areas caused by a single data source of an existing early warning system is solved; a long short-term memory network and convolutional neural network cascade architecture is utilized to capture a localized propagation rule, and the defect that a region-level prediction model cannot adapt to community heterogeneity is overcome; the risk score is generated in real time, the grading response instruction is triggered, a'monitoring-early warning-intervention 'closed loop is established, the early warning timeliness is remarkably improved, a basic-level response chain scission gap is filled, early prevention and control of flu outbreak are finally achieved, and public health resource consumption is reduced.

Description

Technical Field

[0001] This invention relates to the field of infectious disease monitoring and early warning technology, specifically to an intelligent influenza early warning system based on community multimodal data fusion. Background Technology

[0002] Influenza, an infectious respiratory disease, causes approximately 290,000 to 650,000 deaths globally each year, posing a particularly severe threat to young children, the elderly, and those with chronic illnesses, while also imposing a significant socioeconomic burden. Existing influenza early warning systems generally employ artificial intelligence technology to integrate multi-source data: international research improves prediction accuracy by combining influenza-like illness (ILI) data with social media behavior data, and some models integrate climate and population movement data to provide early warnings 7-10 weeks before outbreaks; domestic research integrates weather data, internet search indices, and other multi-source electronic data to construct adaptive models, effectively reducing prediction errors. These systems primarily rely on regional or national-level data and employ time-series analysis, neural networks, and other methods for epidemic prediction.

[0003] However, existing technologies have significant drawbacks:

[0004] Limited forecasting scope: It focuses on macro-level forecasts at the regional or national level, lacks the ability to monitor at the community level, and cannot capture localized transmission characteristics.

[0005] Insufficient data integration: Reliance on a single data source (such as only medical records or meteorological data) and failure to fully integrate multimodal information such as community medical and health data, environmental data, and population activity data leads to delayed early warning;

[0006] Lack of real-time model capability: Most systems use static parameter models, which make it difficult to dynamically optimize parameters to adapt to the rapid changes in the epidemic, affecting the timeliness of intervention.

[0007] These problems have resulted in blind spots in the existing system's "outpost perception" stage in the community, making it unable to meet the early warning needs at the grassroots level. To address this, an intelligent influenza early warning system based on community multimodal data fusion is proposed. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent influenza early warning system based on community multimodal data fusion to solve the problems mentioned in the background.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent influenza early warning system based on community multimodal data fusion, comprising:

[0010] Data acquisition module: used to acquire community-level medical and health data, environmental data, population activity data, and online behavior data in real time;

[0011] Data processing module: used to perform spatio-temporal synchronization, heterogeneous field standardization, and feature fusion on the medical and health data, environmental data, population activity data, and network behavior data to generate a structured input data set;

[0012] AI prediction module: processes the structured data set using a deep learning model and outputs influenza risk scores and incidence predictions;

[0013] Early warning response module: generates hierarchical early warning signals based on the risk scores and pushes intervention instructions to the target terminals.

[0014] Preferably, the data processing module performs the following operations:

[0015] Perform time series indexing on the number of influenza-like illness (ILI) cases and vaccination rates in the medical and health data;

[0016] ** Perform one-hot encoding on the aggregation events in the population activity data;

[0017] Perform standardization processing on the search index in the network behavior data;

[0018] Generate composite features of meteorological data and social activity data, and process the keywords in the network behavior data through word embedding technology; **

[0019] Extract the impact features of historical data on future influenza transmission through time lag effects. **

[0020] Preferably, the AI prediction module includes a cascaded deep learning architecture:

[0021] LSTM sub-module: used to capture the temporal evolution law of influenza cases;

[0022] CNN sub-module: used to extract spatial transmission features between communities;

[0023] Attention mechanism sub-module: used to enhance the mutation detection sensitivity of LSTM and CNN;

[0024] Among them, mean square error (MSE) and mean absolute error (MAE) are used as loss functions for model training, and the Dropout mechanism and early stopping strategy are applied for optimization.

[0025] Preferably, the hierarchical early warning response unit triggers a response according to the following rules:

[0026] [[ID=*45]]When the risk score X ≤ 20, output a blue early warning signal and perform daily monitoring;

[0027] When the risk score 20 < X ≤ 40, output a green early warning signal and trigger vaccination reminders and enhanced prevention and control;

[0028] When the risk score is 40 < X ≤ 60, a yellow warning signal is output and public protection tips are triggered;

[0029] When the risk score is 60 < X ≤ 80, an orange warning signal is output and the flow restriction of gathering places is triggered;

[0030] When the risk score X > 80, a red warning signal is output and the suspension order of large-scale activities is triggered.

[0031] Preferably, the medical and health data includes the number of ILI cases in community clinics, vaccination rates, and respiratory visit records; the environmental data includes temperature, humidity, temperature difference, and wind speed; the population activity data includes holidays, school opening days, and the passenger flow density in business districts; the network behavior data includes the search index of influenza-related keywords.

[0032] Preferably, the CNN sub-module is used to extract the following features, including:

[0033] The spatial transmission relationship between communities;

[0034] The local pattern of meteorological data; [[ID=1​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​This invention constructs a community-level multimodal data fusion architecture, integrating medical and health data, environmental data, population activity data, and network behavior data. It combines this with a deep learning model to dynamically extract and fuse spatiotemporal features, solving the problem of blind spots in community monitoring caused by the single data source in existing early warning systems. By utilizing a cascaded architecture of long short-term memory networks and convolutional neural networks to capture localized transmission patterns, it overcomes the limitation of regional prediction models in adapting to community heterogeneity. Through real-time generation of risk scores and triggering tiered response instructions, a closed loop of "monitoring-early warning-intervention" is established, significantly improving the timeliness of early warnings, filling gaps in grassroots response, and ultimately achieving early prevention and control of influenza outbreaks while reducing the consumption of public health resources.

[0048] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the overall system architecture of the present invention;

[0050] Figure 2 This is a flowchart of the multimodal data processing of the present invention;

[0051] Figure 3 This is a diagram of the spatiotemporal dual-stream deep learning model of the present invention;

[0052] Figure 4 This is a flowchart of the early warning response closed loop of the present invention;

[0053] Figure 5 This is a diagram illustrating the system evolution mechanism of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1-5This invention discloses an intelligent influenza early warning system based on community multimodal data fusion. By constructing a comprehensive data collection network covering medical health, environmental monitoring, population activity, and online behavior, the system achieves dynamic perception of influenza transmission risk at the community level. Relying on a deep learning-based spatiotemporal feature fusion architecture, it generates real-time risk scores and triggers tiered response commands, forming a closed loop of "monitoring-early warning-intervention." This embodiment aims to verify the system's adaptability to the heterogeneous transmission patterns in communities, addressing core issues of traditional early warning models such as blind spots in grassroots monitoring, insufficient data fusion, and delayed response times, ultimately improving early prevention and control capabilities for influenza outbreaks.

[0056] 1. Multimodal data acquisition and dynamic processing mechanism

[0057] 1.1 Real-time Medical and Health Data Capture System

[0058] Data collection terminals are deployed in community health service centers to automate the acquisition of three types of core data through the medical information system interface:

[0059] 1) Dynamic surveillance system for influenza-like illnesses

[0060] Establish a standardized data pipeline to continuously capture case information that meets clinical diagnostic criteria. The specific implementation process includes a multi-level validation mechanism.

[0061] First, the system automatically verifies the spatiotemporal tag integrity of the case records to ensure that each case record contains a timestamp accurate to the minute and a location code based on the community geographic grid;

[0062] Secondly, community clinicians review the logical consistency between the symptom description text and the diagnostic code to rule out false reports caused by non-infectious febrile diseases.

[0063] Finally, disease control experts conduct epidemiological link analysis on clustered cases. When the system detects more than three times the baseline value of cases in the same school or residential community within 24 hours, it automatically triggers a contact tracing procedure and generates a preliminary transmission chain report. This data pipeline is optimized through stream processing technology, compressing data latency to a fraction of that of traditional systems, achieving near real-time monitoring.

[0064] 2) Scheme for constructing a panoramic view of vaccination

[0065] By connecting to the regional immunization program platform via a dedicated line, a two-tiered immunization analysis model covering individuals and communities can be established.

[0066] At the individual level, the system records metadata such as the precise vaccination time, vaccine strain subtype, age of the recipient, and basic health status for each dose of vaccine. Based on the immunological model, it calculates the protection decay curve and scientifically divides the post-vaccination period into four stages: the immune establishment period (protection ramp-up stage from 0 to 14 days after vaccination), the highly effective protection period (antibody peak stage from 15 to 28 days), the decay period (linear decline in protection from 29 to 90 days), and the basic protection period (residual protection stage after 90 days).

[0067] At the community level, the system integrates parameters such as vaccination coverage, high-risk population density, and vaccination time dispersion from each microgrid to generate a dynamically updated herd immunity heat map, visually displaying the distribution of immune barrier strength in different areas of the community. For example, in aging community grids, the system will automatically increase the immune weight coefficient for the elderly population, more accurately assessing the effectiveness of herd immunity.

[0068] 3) Respiratory Disease Visit Analysis Engine

[0069] Develop a machine learning-based syndrome monitoring algorithm to identify potential outbreaks by analyzing multi-dimensional features in electronic medical records from community clinics.

[0070] When the system detects that the rate of seeking medical attention for acute respiratory symptoms increases by more than two standard deviations from the historical baseline within 48 hours, it automatically initiates a three-level early warning and verification process.

[0071] Simultaneously, correlation analysis was conducted to identify abnormal fluctuation patterns in the prescription volume of antipyretic drugs, such as a sudden increase in the prescription frequency of drugs like ibuprofen and acetaminophen.

[0072] Further text mining techniques are used to perform cluster analysis on the symptom description fields. When keywords such as "persistent high fever" and "body aches and pains" show an abnormally high density, the system will generate a syndrome warning report and submit it to disease control experts for review. This engine can achieve high-frequency updates of risk assessment results during the winter flu season.

[0073] The standardized case dataset is directly input into the transmission dynamics analysis unit of the feature engineering module, and the spatiotemporal map of vaccination provides an immune background dimension for the risk prediction model.

[0074] The automated pipeline enables seamless flow of case data from collection to analysis, solving the problem of time lag in traditional manual reporting; the cluster case identification engine provides key signal input for early outbreak warning.

[0075] 1.2 Intelligent Sensing Network for Environmental Data

[0076] 1) Meteorological parameter fusion and analysis system

[0077] By accessing minute-level monitoring data streams from the meteorological department in real time through a secure data channel, a composite environmental stress assessment model with three dimensions is constructed:

[0078] In terms of temperature, the system continuously calculates the deviation between the ambient temperature and the optimal temperature range for influenza virus transmission (5-15℃). When the daily average temperature is below the critical value, a continuous low temperature accumulation counter is activated to record the duration of adverse weather conditions.

[0079] In terms of humidity, a non-linear relationship model between relative humidity and virus survival time is established. When humidity exceeds 75% for more than 48 consecutive hours, a high humidity environment risk alarm is automatically generated.

[0080] In terms of wind speed, a wind speed change monitoring algorithm was designed to assess the increased risk of airborne pathogen transmission by calculating the rate of change of wind speed per unit time.

[0081] When these three dimensions of indicators create a synergistic effect, the system will raise the environmental risk level to an orange alert. For example, during a cold wave, if the system detects a sudden drop in temperature of 10°C accompanied by a doubling of wind speed, it will immediately trigger an environmental risk escalation protocol.

[0082] 2) Dynamic mapping platform for microenvironment quality

[0083] Deploy IoT sensor networks in key community locations to build a sophisticated environmental monitoring system:

[0084] In educational institutions, multi-parameter environmental sensors are deployed in classrooms to collect real-time data on carbon dioxide concentration (reflecting ventilation efficiency), window opening frequency (monitoring natural ventilation behavior), and infrared sensor personnel density (calculating the number of people per unit area).

[0085] In commercial complex scenarios, the system accesses the central air conditioning control system to obtain fresh air exchange rate data, combines it with the video passenger flow analysis system to calculate peak-hour passenger flow density, and uses WiFi probe technology to analyze the length of time customers stay in enclosed areas.

[0086] In public transportation hubs, passenger compartment crowding monitoring devices and handrail contact frequency sensors are installed. All data is spatially matched through a community geographic grid coding system to generate a three-dimensional environmental risk heat map with a time dimension, intuitively presenting the real-time risk status of each micro-area.

[0087] For example, in the food court area of ​​a shopping mall during winter, the system continuously monitors CO2 concentrations exceeding 1500ppm and the average customer stay time exceeds 90 minutes, marking the area as a high-risk red zone.

[0088] The meteorological stress index is output to the environmental response engine of the feature engineering, and the microenvironmental heat map is overlaid with the community geographic grid system to provide site-level input for the spatial propagation model.

[0089] Through multi-dimensional environmental parameter fusion analysis, the complex influence mechanism of meteorological conditions on virus transmission is revealed; the micro-environment monitoring network enables precise location of high-risk transmission sites.

[0090] 1.3 In-depth Quantitative Model of Crowd Activity

[0091] Education Scene Aggregation Effect Evaluation System

[0092] 1) The dynamic risk assessment model designed for the back-to-school season includes three core modules:

[0093] The basic parameter acquisition module accurately records static parameters such as the school calendar, student enrollment data, and boarding student ratio of each educational institution;

[0094] The historical epidemic correlation module retrieves influenza incidence data from the same period in the past five years at the school to construct a baseline model for seasonal disease transmission.

[0095] The real-time dynamic monitoring module is connected to the school's electronic attendance system and obtains the change curve of the absenteeism rate due to illness every two hours.

[0096] The system uses a non-linear decay algorithm to calculate the daily risk intensity value after the start of the school year. This value reaches its peak on the 3rd to 5th day after the start of the school year, and then gradually decays according to an exponential curve. For example, for a middle school with 1,200 students, the system calculates the risk coefficient for the fifth day of the school year as 0.78 (high risk level) based on its 40% boarding student ratio and the average incidence rate of 12% in the second week of the school year.

[0097] 2) Business Activity Heat Analysis Framework

[0098] By integrating mobile operator signaling data, electronic payment location data, and video surveillance data, a risk assessment model for commercial venues is constructed.

[0099] The passenger flow density analysis module calculates the instantaneous number of people per unit area based on changes in mobile phone signal strength. For example, it triggers a primary alarm when it detects more than 3 people per square meter.

[0100] The dwell time characteristic analysis module identifies the proportion of people who stay for extended periods, for example, marking customers who stay for more than 120 minutes as high-risk individuals;

[0101] The movement pattern tracking module analyzes the community epidemic risk level of the customers' origin, for example, when the proportion of visitors from high-risk areas exceeds 30%, the venue risk rating is increased.

[0102] When these three indicators simultaneously reach the warning threshold, the system automatically raises the warning level to orange. For example, during the Spring Festival promotion period in a community shopping center, the system detected a peak customer density of 5 people per square meter, 40% of customers stayed for more than 2 hours, and 35% of customers came from nearby communities with influenza outbreaks, thus generating an orange cluster risk warning.

[0103] The school prevention and control strategy library is directly connected to the early warning and response module for the risk coefficient of educational gatherings; the graph neural network is used to input the spatial transmission channel of the heat map of passenger flow in commercial places.

[0104] Quantify the amplification effect of crowd gathering behavior to identify super-spreading scenarios; provide behavioral basis for differentiated prevention and control strategies.

[0105] 1.4 Network Behavior Semantic Parsing System

[0106] 1) Search Intent Deep Recognition Engine

[0107] Constructing a three-level semantic knowledge graph of influenza:

[0108] The symptom description layer includes core keywords such as fever, chills, and muscle aches;

[0109] The treatment needs layer integrates the names and administration methods of antiviral drugs such as oseltamivir and peramivir.

[0110] The preventative measures section includes information on vaccination sites, mask-wearing guidelines, and other protective measures.

[0111] The system uses a BERT-based deep learning model to perform triple analysis on the search query:

[0112] The urgency scoring module identifies high-risk word combinations such as "high fever of 41 degrees Celsius" and "difficulty breathing";

[0113] The spatial positioning module maps IP address resolution and user historical location data to the community grid;

[0114] The time pattern analysis module detects abnormal fluctuations in search volume.

[0115] For example, if the system simultaneously detects a 300% surge in searches for "flu specific drugs" and a 25% share of phrases related to "persistent high fever" within a certain community grid, it will immediately trigger a red alert for network behavior in that area.

[0116] 2) Public opinion dissemination dynamics tracking mechanism

[0117] Establish a full lifecycle monitoring system for social media topics:

[0118] The initial spread rate analysis module records the forwarding growth rate of a topic in the first hour after its emergence. For example, when the growth rate exceeds 500% / hour, it is marked as a rapidly spreading topic.

[0119] The In-Depth Assessment module calculates the participation rate of key opinion leaders in the medical field, and improves the topic's authority rating when the participation rate of certified physician accounts exceeds 15%.

[0120] The duration monitoring module tracks how long a topic remains on the trending topics list and establishes a risk correlation model by combining topic content with the characteristics of influenza transmission. For example, if the topic #ElementarySchoolSuspendedDuringFlu# reaches the top ten of the local trending topics list within 2 hours, and three chief physicians from the respiratory departments of top-tier hospitals participate in the discussion, the system will raise the risk rating of the topic to the highest level.

[0121] The network behavior feature vector is input into the LSTM network of the time series analysis channel; public opinion risk rating participates in dynamic weight adjustment decision-making.

[0122] Transforming unstructured network data into early warning signals for the epidemic, capturing the surge in medical demand and social psychological fluctuations.

[0123] 1.5 Multi-source data spatiotemporal alignment system

[0124] Developing an intelligent data coordinator to solve the challenge of cross-source data integration:

[0125] Time-dimension synchronization module: Creates a dynamic sliding time window mechanism to coordinate the time alignment between second-level meteorological data and hour-level medical data. For example, it resamples minute-level meteorological data into hourly averages to match the frequency of medical data.

[0126] The spatial dimension unification module employs the Geohash grid coding system to accurately map points of interest such as shops and schools to the community management grid, achieving spatial matching. The data anomaly handling module is equipped with an intelligent interpolation algorithm; for example, when medical data is delayed by more than 4 hours, it predicts missing values ​​based on historical data from the same period and trends in neighboring areas. The system sets execution times and performs global data consistency checks at intervals to ensure accurate fusion of multi-source data within a unified spatiotemporal coordinate system.

[0127] Composite feature generation engine: Nonlinearly combines meteorological data and social activity data. For example, when the temperature drops by more than 8°C and the passenger flow density in the business district exceeds 5 people / ㎡ during holidays, a "low temperature-gathering" composite risk index is generated. Keywords in network behavior data (such as "persistent high fever" and "oseltamivir") are transformed into 128-dimensional semantic vectors through Word2Vec word embedding technology and input into the sentiment analysis unit of the LSTM submodule.

[0128] Output spatiotemporally aligned multimodal datasets for unified processing by the feature engineering module; provide standardized spatiotemporal input matrices for deep learning models.

[0129] Establish a unified spatiotemporal reference system for cross-source data to solve the core technical bottleneck of multimodal fusion.

[0130] The multimodal data acquisition and dynamic processing mechanism, through the synergistic effect of a four-dimensional data acquisition system encompassing medical, environmental, behavioral, and network dimensions, constructs a real-time monitoring network covering the entire community. The spatiotemporal alignment engine transforms heterogeneous data into standardized input, laying the data foundation for subsequent feature extraction. This module breaks through the limitations of traditional single data sources, achieving three-dimensional capture of early warning signals of the epidemic.

[0131] 2. Multimodal Feature Engineering System

[0132] 2.1 Epidemiological characteristics

[0133] 1) Disease transmission dynamics feature extraction module

[0134] This module contains three core feature generators:

[0135] The short-term transmission trend analyzer uses a second-order time series differential algorithm to calculate the acceleration change value of the case growth rate. For example, when the acceleration exceeds three times the standard deviation of the historical mean for two consecutive days, it is marked as an abnormal transmission signal.

[0136] The periodic fluctuation feature extractor uses wavelet transform technology to identify propagation cycle patterns such as 7 days, 14 days, and 28 days, for example, it discovers the periodic decline in the consultation rate in the weekend effect analysis;

[0137] The spatial association pattern modeler constructs a propagation intensity matrix between communities and quantifies the mutual influence coefficients of geographically adjacent areas through covariance analysis.

[0138] For example, the system calculates the transmission correlation coefficient between two adjacent community grids, indicating that the spread of the epidemic has a significant spatial dependence.

[0139] 2) Vaccine protective effect modeling system

[0140] Establish a phased protection force attenuation model:

[0141] The immune establishment period feature extractor analyzes the slope of the protective efficacy ramp-up curve after vaccination, such as from 0 to 14 days, to identify individuals with a slow immune response.

[0142] The high-efficiency protection period feature monitor tracks the pattern of protective efficacy maintenance during the peak phase of the antibody and issues an abnormal alarm when the herd protection fluctuates by more than, for example, 10% during the peak phase.

[0143] The decay period characteristic analyzer calculates the rate at which protective efficacy decreases linearly over time and adjusts the decay coefficient based on the age of the recipient. For example, for the elderly population aged 65 and above, the system adjusts their decay coefficient to 1.3 times that of the general population, more accurately reflecting the effects of immunosenescence.

[0144] The output propagation dynamics feature vector is input into the LSTM time-series channel; the vaccine protective efficacy decay curve is used in the risk score calculation.

[0145] Deconstructing the spatiotemporal evolution of disease transmission provides key epidemiological parameters for predictive models.

[0146] 2.2 Environmental Response Feature Engine

[0147] 1) Meteorological Stress Cumulative Effect Analysis System

[0148] Design an evaluation model for the coupling effect of temperature and humidity:

[0149] The low temperature and high humidity synergistic effect module identifies dangerous combinations such as temperatures below 10°C and relative humidity exceeding 80%, and generates a synergistic risk coefficient.

[0150] The temperature change stress analyzer calculates, for example, drastic fluctuations in temperature exceeding 8°C within 24 hours and marks them as meteorological stress events.

[0151] The algorithm for accumulating adverse conditions performs risk accumulation calculations for areas experiencing unfavorable weather conditions for more than 5 consecutive days. For example, if a community experiences 6 consecutive days of low temperature and high humidity, the system will raise its environmental risk accumulation value to the highest level.

[0152] 3) Framework for extracting microenvironmental risk features

[0153] Different evaluation systems are designed for different types of venues:

[0154] In educational institutions, a balance coefficient between classroom personnel density and ventilation efficiency is used. A high-risk signal is generated when, for example, more than 2 people per square meter and CO2 concentration > 1000 ppm are detected.

[0155] In commercial settings, the matching degree between peak customer flow and spatial carrying capacity is calculated, and the carrying capacity threshold is dynamically adjusted based on the building volume ratio and ventilation system performance.

[0156] In transportation infrastructure scenarios, a risk model for exposure duration in enclosed spaces is established, marking passengers who stand continuously in subway cars for more than 30 minutes as high-risk individuals. For example, on subway lines during the morning rush hour, if the system identifies a car where passengers have an average exposure time of 45 minutes, it automatically raises the risk rating of that car to orange.

[0157] The environmental stress index is input into the dynamic risk assessment engine; the site risk rating is directly connected to the early warning and response strategy library.

[0158] To reveal the regulatory mechanism of environmental factors on transmission efficiency and to achieve site-specific risk assessment.

[0159] 2.3 Quantification System of Social Behavioral Characteristics

[0160] 1) Three-dimensional evaluation model of clustered events

[0161] Quantifying the risks of large-scale events from three dimensions:

[0162] The spatial dimension analyzer calculates the fit between the available area of ​​the site and the expected population density. When the expected density per square meter exceeds the safety threshold, a spatial risk value is generated.

[0163] The time-dimensional evaluator analyzes the non-linear relationship between activity duration and risk accumulation effect, and performs risk multiplication calculations for activities, for example, exceeding 4 hours.

[0164] The population structure analysis module assesses the age distribution and immune status of participants, for example, raising the risk level when the proportion of elderly people and children exceeds 40%.

[0165] For example, for community Spring Festival temple fair activities, the system calculated a comprehensive risk value of 0.86 (red risk level) based on data that a 5,000-square-meter venue is expected to accommodate 20,000 people for 8 hours, and that 45% of the participants are elderly or children.

[0166] 2) Population Flow Feature Extraction System

[0167] Constructing a cross-community mobility network analysis model:

[0168] The flow tracking module identifies the transmission path from high-risk areas to low-risk areas and draws a daily commuting heat map using mobile phone signaling data.

[0169] The flow intensity quantifier analyzes the dose-effect relationship between commuting scale and transmission risk, and establishes a risk coefficient table for different flow intensity ranges;

[0170] The population characteristic analyzer estimates the age structure of the mobile population through base station data and calculates the imported risk value by combining it with the epidemic data of the place of origin.

[0171] For example, during the morning rush hour, the system monitored 1,500 people commuting from communities where the outbreak occurred to this community every day, of which 35% were school-age children. Based on this, the imported risk coefficient was calculated to be 0.75.

[0172] Clustered risk values ​​are used to determine the early warning level; mobile network characteristics are input into the spatial propagation channel.

[0173] Quantify the amplification effect of social activities on the transmission chain and identify key nodes in super-spreading scenarios.

[0174] 2.4 Cross-modal feature fusion system

[0175] A three-tiered integration strategy is adopted to achieve deep integration:

[0176] The primary feature fusion layer overlays meteorological data onto the geographic information system to generate a spatial distribution map with environmental pressure attributes. For example, temperature and humidity layers are overlaid with community grids to generate an environmental risk heat map.

[0177] The intermediate model fusion layer integrates population flow networks and epidemic distribution data through graph neural networks to construct a graph of inter-community transmission relationships;

[0178] The advanced decision fusion layer dynamically adjusts the contribution weights of each modality based on the attention mechanism. For example, it automatically increases the decision weight of meteorological features during cold waves and enhances the weight coefficient of crowd gathering features during the school opening season.

[0179] The system generates a feature contribution analysis report every 24 hours to continuously optimize the fusion strategy.

[0180] The output fused feature matrix is ​​used for training the deep learning model; the feature contribution analysis report is fed back to the data acquisition module.

[0181] Achieve deep complementarity and synergistic enhancement of multi-source information to improve feature discrimination capabilities.

[0182] The multimodal feature engineering system constructs a comprehensive feature profile of the influenza transmission chain through the synergy of three-dimensional feature extraction systems: propagation dynamics, environmental response, and social behavior. Cross-modal fusion technology overcomes the limitations of single data dimensions, providing highly discriminative input features for deep learning models. This module transforms raw data into decision parameters with epidemiological significance.

[0183] 3. Deep Learning Model Architecture and Training Mechanism

[0184] 3.1 Spatiotemporal Dual-Stream Processing Architecture

[0185] 1) Timing Analysis Channel (Enhanced LSTM)

[0186] This channel is designed with a gating mechanism:

[0187] Seasonal memory unit: Enhances the model's ability to identify seasonal epidemic characteristics through periodic weight adjustment, and automatically increases the memory strength of historical data during peak flu months;

[0188] Mutation response unit: Designed with a dynamic memory retention mechanism, when the case growth rate exceeds three times the standard deviation of the historical mean for two consecutive days, the memory weight of recent data is automatically enhanced;

[0189] Data input pipeline integrates three types of time series:

[0190] The medical data stream includes the daily number of ILI cases and the curve of changes in the consultation rate;

[0191] The trend of declining protective efficacy in the population as shown by the input of immune data streams;

[0192] Network behavior flow provides characteristics of search volume fluctuations.

[0193] The system accurately identifies patterns similar to historical epidemic curves through seasonal memory units and issues early warning signals in advance.

[0194] 2) Spatial propagation channel (graph convolutional network)

[0195] Constructing a community communication relationship graph:

[0196] Node definition system: Each community grid is transformed into an independent node, and the node feature vector contains six-dimensional features such as case density, environmental stress value, and population aggregation index;

[0197] Edge weight calculation module: Taking into account both population flow intensity and geographical proximity, it quantifies the connection strength between communities through commuting data;

[0198] The multi-layer graph convolution operation design adopts a three-order neighborhood propagation mechanism. The first layer extracts the influence of directly adjacent communities, the second layer analyzes indirect propagation paths, and the third layer identifies cross-regional propagation hotspots.

[0199] For example, in urban epidemic analysis, the system uses three-layer graph convolution to identify commercial centers as key hubs for cross-community transmission.

[0200] The time-series channel output connects to a dynamic risk assessment engine: spatial channel hotspot identification directly connects to regional prevention and control strategies.

[0201] At the same time, we can capture the temporal dynamics and spatial spread patterns of the epidemic's evolution and improve our ability to model complex transmission patterns.

[0202] 3.2 Adaptive Training Optimization System

[0203] 1) Incremental learning framework

[0204] Model training specifications: Backpropagation is performed using a dual loss function of mean squared error (MSE) and mean absolute error (MAE), with MSE accounting for 70% of the weight and MAE accounting for 30%; Dropout mechanism (dropout rate of 0.3) is applied to the fully connected layers of the network to prevent overfitting; an early stopping strategy is set—training is automatically terminated when the validation set loss decreases by less than 1% for 5 consecutive epochs.

[0205] Design a data flow-driven four-phase update mechanism:

[0206] The arrival of new data triggers the assessment of the degree of change in data distribution using a drift detection algorithm.

[0207] During the decision-making and diversion phase, the update strategy is selected based on the magnitude of the drift value—for slight drift, only the weights of the last two fully connected layers are fine-tuned, and for significant drift, global retraining is initiated.

[0208] The historical knowledge preservation stage employs flexible weighting solidification technology to impose constraints on important parameters and prevent catastrophic amnesia;

[0209] The validation and evaluation phase involves examining the performance of the updated model by retaining the dataset.

[0210] This framework enables automatic weekly model iterations, ensuring continuous improvement in prediction accuracy.

[0211] 2) Real-time performance monitoring system

[0212] Establish a three-dimensional quality tracking system: The short-term predictive quality monitor is set to compare the model output with the actual case development trend every 8 hours, and calculate the accuracy of the trend direction;

[0213] The mid-term early warning timeliness assessor tracks the effectiveness of the early warning time window and records the time interval from the issuance of the early warning to the outbreak of the epidemic;

[0214] The long-term model stability analysis module assesses the consistency of predictions across different epidemic seasons.

[0215] When the system detects that the prediction error exceeds the acceptable threshold for three consecutive days, it automatically triggers the emergency retraining process and notifies the technical team to intervene and diagnose the problem.

[0216] The incremental learning framework outputs a model update to the early warning decision module, and performance monitoring results trigger feature engineering optimization.

[0217] To achieve continuous self-optimization of model parameters and adapt to viral mutations and changes in population behavior.

[0218] 3.3 Attention Enhancement Submodule

[0219] 1) Cross-modal dynamic alignment mechanism

[0220] The system's ability to detect and respond to sudden changes in the epidemic in the community is improved by using a multi-level attention collaboration mechanism.

[0221] In the specific implementation, a cross-modal dynamic alignment mechanism is first constructed:

[0222] Use the final hidden state of the LSTM output from the time series analysis channel as the baseline query vector;

[0223] Simultaneously, the feature map generated by the spatial propagation channel is transformed into a key-value pair matrix;

[0224] An attention weight calculation model is used to achieve dynamic weighted fusion of spatiotemporal features.

[0225] When a specific community simultaneously experiences both a temporal signal of an abnormal surge in cases and a spatial signal of dense population activity, the system automatically increases the feature weight of that area to several times the baseline value, forming a precise location of the risk focus, and outputs it to the risk focus location system of the hierarchical early warning module.

[0226] 2) Abrupt signal amplification design

[0227] In the mutation detection enhancement mechanism, a gradient monitoring unit is designed to track the acceleration changes in the case growth rate in real time. When a significant abnormal fluctuation is detected, an attention gating system is triggered.

[0228] On the one hand, it increases the proportion of recent data in the query vector in terms of memory strength; on the other hand, it enhances the signal performance of high-risk areas in the spatial feature map and dynamically adjusts the trigger threshold based on the historical propagation characteristics of the community to adapt to the heterogeneous characteristics of different areas.

[0229] Finally, the attention-optimized fusion features are directly input into the hierarchical early warning module through the early warning decision direct connection channel, generating a visual risk heat map to identify high-risk spatiotemporal coordinates (such as the start of school week or a promotional period in a business district), and simultaneously triggering the response strategy generation engine to match scenario-based intervention measures, forming a closed-loop decision flow of "risk identification - focus reinforcement - precise intervention".

[0230] This design breaks through the limitations of traditional static fusion models, enabling the system to adaptively adjust the priority of the areas of interest, improve the detection sensitivity of sudden gathering events such as school outbreaks, and achieve continuous evolution and optimization of the model by feeding back response effect data to the weight allocation algorithm.

[0231] The optimized fusion feature is directly connected to the early warning decision module, generating a risk heat map to identify high-risk spatiotemporal coordinates (such as the start of school week / promotion period), and simultaneously activating the response strategy engine to match scenario-based measures.

[0232] Breaking through the limitations of static fusion models, it achieves adaptive positioning of risk focus; by feeding data back to the weight allocation algorithm, it constructs a closed-loop learning mechanism of detection-response-optimization.

[0233] The deep learning model architecture and training mechanism integrate the advantages of LSTM periodic perception and GCN spatial propagation through a spatiotemporal dual-stream architecture. The attention enhancement mechanism significantly improves the sensitivity to community mutation detection. The adaptive training system ensures the model's continuous evolution capability in dynamic environments through drift detection and knowledge protection techniques. This module forms a technical closed loop of "data feature extraction → spatiotemporal modeling → focus enhancement," providing a core driving force for accurate early warning.

[0234] 4. Hierarchical early warning and closed-loop control system

[0235] 4.1 Dynamic Risk Assessment Engine

[0236] Risk score generation uses a multi-source fusion algorithm:

[0237] The basic scoring model integrates three parts: time-series risk value (from LSTM channel), spatial propagation risk (from GCN channel), and environmental stress index (from feature factory).

[0238] The dynamic weight adjuster adjusts weight allocation based on real-time scenarios. For example, when medical data shows abnormal fluctuations, the time-series risk weight is increased to 0.7; when an outbreak occurs in a neighboring area, the spatial transmission risk weight is increased to 0.6; and when the meteorological department issues a red alert, the environmental pressure coefficient is increased to 0.5. The community vulnerability correction module adaptively adjusts the baseline risk value by considering factors such as the proportion of the elderly population, the prevalence of chronic diseases, and the accessibility of medical resources. For example, in aging communities, the system automatically lowers the baseline risk threshold by 20% to match the vulnerability characteristics of the population.

[0239] Receive the prediction results output by the deep learning model; trigger the generation mechanism of graded early warning signals.

[0240] This enables accurate and dynamic assessment of risk values, avoiding the misjudgment problems associated with traditional static thresholds.

[0241] 4.2 Intelligent Response Strategy Generation System

[0242] 1) Optimize the engine with measures

[0243] Constructing a multi-objective decision-making mathematical model:

[0244] The effectiveness dimension evaluator calculates the suppression efficiency of different intervention measures on the basic reproduction number (R0) and establishes a quantitative scoring table for the effectiveness of the measures;

[0245] The cost dimension analyzer calculates the human and material input requirements, including vaccine consumption, protective equipment usage, and medical staff working hours;

[0246] The feasibility assessment module analyzes the community's execution capacity and residents' expected compliance to generate an implementation difficulty coefficient.

[0247] The system solves the Pareto optimal solution set through linear programming, providing decision-makers with 3-5 optimization options.

[0248] For example, in an orange alert scenario in a certain community, the system outputs three solutions: Solution A focuses on school prevention and control (effectiveness score 85 / cost score 70), Solution B focuses on the management and control of commercial venues (effectiveness score 80 / cost score 65), and Solution C adopts a tiered approach across the entire area (effectiveness score 90 / cost score 85).

[0249] 2) Scenario-based response strategy library

[0250] Establish a hierarchical and categorized response system:

[0251] For example, configuring a three-level response process in an educational institution scenario:

[0252] During a yellow alert, enhanced morning health checks and upgraded classroom ventilation are implemented, requiring windows to be opened for ventilation no less than 6 times a day.

[0253] The orange alert has triggered staggered school arrival and departure times, spreading out arrival times within 40-minute intervals, while all large gatherings have been suspended.

[0254] When a red alert is issued, the school will switch to online teaching and implement comprehensive and in-depth disinfection of the campus environment, using chlorine-containing disinfectants to treat high-frequency contact surfaces daily.

[0255] Commercial venue management solutions include customer flow density control strategies:

[0256] For example, a yellow alert would involve setting up temperature screening points at the entrance;

[0257] For example, when an orange alert is issued, a 50% flow restriction measure is implemented, and protective reminders are broadcast every 30 minutes via the public address system;

[0258] For example, a red alert would close off non-essential areas and mandate that all personnel wear N95 protective equipment.

[0259] Nursing homes implement a gradual closed management system for their protective measures.

[0260] For example, a visitor health declaration system is implemented during a yellow alert;

[0261] For example, an orange alert would suspend group dining and recreational activities;

[0262] For example, a red alert triggers a complete lockdown, with a professional medical team stationed on-site to provide health monitoring.

[0263] Input risk assessment results and resource constraints; output prevention and control plans to community implementation terminals.

[0264] To balance the effectiveness of prevention and control with the cost of implementation, we can generate precise strategies that are highly operable.

[0265] 4.3 Closed-loop feedback verification mechanism

[0266] Example: A primary school's back-to-school season implementation case demonstrates the complete closed-loop process:

[0267] 1) Risk Identification Phase

[0268] The system detected the start date of school 7 days in advance. Retrieving the school's historical epidemic data, it was found that the average incidence rate in the second week of school over the past three years reached 15%. Environmental monitoring showed that the average CO2 concentration in the school's classrooms reached 1200ppm, and the ventilation efficiency rating was C. Combined with the weather forecast, it was predicted that the start of school would be affected by a cold wave and temperature drop.

[0269] Multi-source data input into the dynamic risk assessment engine triggers an orange alert signal.

[0270] 2) Early warning triggering stage

[0271] The comprehensive assessment generated a risk score of 68 (orange alert), and a response plan including 5 core measures was sent to the school administration:

[0272] ① Implement staggered school start and end times by grade level (30-minute interval);

[0273] ② Ultraviolet disinfection is added to the classrooms every day at noon;

[0274] ③Suspend gatherings such as opening ceremonies;

[0275] ④ Personalized protection guidelines are pushed to parents;

[0276] ⑤ Stockpile sufficient rapid testing reagents.

[0277] The solution is pushed to the school management terminal and parents' mobile devices through a standardized interface.

[0278] 3) Intervention Implementation Phase

[0279] The school strictly implements the staggered arrival time plan, spreading the arrival time of 2,000 students between 7:00 and 8:30; the logistics department conducts ultraviolet disinfection of classrooms every day from 12:00 to 14:00; the school clinic randomly selects 10% of students' throat swab samples every day; and parents report any abnormal body temperature during daily morning checks.

[0280] The execution data is transmitted back to the effect monitoring module in real time.

[0281] 4) Results Feedback Phase

[0282] Real-time monitoring shows that the student sick leave rate remained stable at 3.5% during the first week of school (compared to an average of 8.2% for the same period in history); the CO2 concentration in classrooms dropped below 800 ppm; and no school-related clusters of cases were found in the surrounding communities.

[0283] The cleaned validation data is then input into the model optimization pipeline.

[0284] 5) Model optimization stage

[0285] Successful experiences were fed back to the system: the feature engineering module increased the weight of school cluster features by 35%; a new "enhanced school prevention and control during cold waves" scheme was added to the strategy library; the evaluation weight of ventilation efficiency was added to the environmental monitoring indicators; and optimized parameters were injected into the decision flow for the next early warning cycle.

[0286] The early warning coverage quantification module: Real-time statistics on the reach rate of early warning information pushed to community residents' mobile devices, and monitoring social media platform interaction data (including forwarding volume, likes, and comment sentiment analysis); for example, when the reach rate is lower than the threshold or the proportion of negative feedback is greater than 30%, the multi-channel re-sending protocol is automatically activated (such as community broadcasting, smart information screen scrolling broadcast).

[0287] A closed-loop intelligent decision-making flow of "multi-source perception → strategy matching → execution feedback → model evolution" is formed, and the system is driven to iteratively optimize through practical verification.

[0288] The tiered early warning and closed-loop control system uses a dynamic risk assessment engine to transform predicted signals into tiered early warnings; the intelligent strategy generation system outputs actionable solutions based on a scenario library; and the closed-loop feedback mechanism continuously optimizes model parameters and the knowledge base through multi-dimensional verification data. This module bridges the value transformation path from "technology prediction" to "community prevention and control," constructing an adaptive and evolving intelligent decision-making ecosystem.

[0289] 5. System Implementation Performance Verification System

[0290] 5.1 Predictive Performance Verification Method

[0291] 1) Time-based assessment scheme

[0292] A rolling time window verification mechanism is adopted:

[0293] The training set consists of complete data from three consecutive flu seasons (approximately 300 observation points);

[0294] The test set was used to verify the predictive ability over different time spans in stages—the accuracy of the trend direction for short-term predictions (1-7 days) was evaluated every 24 hours.

[0295] The medium-term forecast (8-14 days) focuses on examining the peak time error (number of days earlier / later);

[0296] Long-term prediction (15-28 days) assesses the epidemic intensity bias rate (the percentage difference between predicted and actual incidence rates).

[0297] The verification process continued throughout the entire season, with more than 120 rolling tests conducted in total.

[0298] 2) Spatial verification framework

[0299] Design of a synthetic control experiment:

[0300] Two communities, A and B, with similar population structures were selected. A complete early warning system was deployed in community A, while traditional monitoring methods were used in community B.

[0301] The comparative indicators include the timeliness of initial case identification (the time difference between the appearance of a case and its marking by the system), the ability to detect cluster events (the ratio of the number of cluster events identified by the system to the actual number of outbreaks), and the accuracy of epidemic boundary positioning (the spatial overlap between the predicted hotspot areas and the actual distribution of cases).

[0302] Multi-indicator validation protocol: Short-term forecasts (1-7 days) are assessed with a mean absolute percentage error (MAPE) ≤ 15%; medium-term early warnings (8-14 days) require a sensitivity ≥ 90%; long-term forecasts (15-28 days) are evaluated using weighted interval score (WIS), with an acceptance threshold of WIS < 0.25. In model stability testing, the forecast error fluctuation range across annual flu seasons is required to be < 5%.

[0303] The experiment lasted for two full flu seasons, collecting data from more than 5,000 cases.

[0304] The verification results are fed back to the model training module to guide the direction of feature engineering optimization; and objective and comprehensive performance evaluation standards are established to drive technology iteration.

[0305] 5.2 Mechanism for Verifying Prevention and Control Effectiveness

[0306] 1) Evaluation of resource optimization effects

[0307] Establish a resource end-to-end tracking system:

[0308] The vaccine usage monitoring module records data from the entire chain, from planned allocation, cold chain transportation, vaccination sites to actual vaccination, and calculates the loss rate at each stage;

[0309] The human resource dispatch and analysis system collects parameters such as emergency team response time, working hours, and workload distribution.

[0310] The material allocation and assessment module uses RFID technology to track the efficiency of protective equipment flow from warehouse to end user.

[0311] The system generates a resource utilization efficiency report every month to identify areas for optimization.

[0312] 2) Analysis of transmission blocking effectiveness

[0313] Design a propagation chain tracing model:

[0314] The second-generation incidence rate statistics module calculates the proportion of secondary infections caused by primary cases through family follow-up surveys;

[0315] The family clustering analysis system establishes a household relationship map and marks the scale of the transmission chain within the same residential unit;

[0316] The site transmission assessment module calculates the percentage of cases associated with a specific site through cross-analysis of case activity trajectories. For example, in a community outbreak, the system identified a supermarket as a key transmission site (accounting for 65% of associated cases), and adjusted the prevention and control strategies for commercial venues accordingly.

[0317] Optimize the knowledge base of effectiveness data input strategies to guide adjustments to resource allocation strategies; quantify the actual value of prevention and control measures to optimize public health decision-making.

[0318] 5.3 Pathways to Realize Social Value

[0319] 1) Vulnerable Group Protection Mechanism

[0320] Develop a targeted protection strategy system:

[0321] The elderly protection module establishes a tiered closed management system for elderly care institutions, adjusting visitation policies according to risk levels.

[0322] The child protection network implements a home-school collaboration mechanism, with parents providing daily feedback on students' health status through parent groups.

[0323] The chronic disease patient channel has set up a community medical priority response mechanism, and family doctor teams conduct daily remote monitoring of high-risk patients.

[0324] This mechanism provides targeted protection for vulnerable populations during the winter peak season.

[0325] 2) Community Resilience Building System

[0326] A multi-layered response capability is constructed: the information transmission layer integrates multiple early warning systems, including SMS, public accounts, and community broadcasts, ensuring information coverage of over 90%; the material support layer establishes distributed emergency reserve points and sets up smart cabinets for protective materials in community property management centers; the capacity building layer develops online training courses on protective knowledge, improving residents' self-rescue and mutual-rescue capabilities through video teaching. The system organizes a community emergency drill every quarter.

[0327] 3) Technology migration verification scheme

[0328] The design incorporates a modular adaptation framework: for respiratory infectious diseases, the core spatiotemporal transmission model is retained, with incubation period parameters adjusted for RSV; for enteric infectious diseases, environmental media monitoring is enhanced, and data interfaces for water and food samples are added; for vector-borne infectious diseases, vector biological monitoring data is integrated to establish a correlation model between mosquito density and disease transmission. Each disease adaptation has undergone at least six months of field validation.

[0329] Social value assessment data supports policy formulation, and migration verification results expand the system's application scenarios; it realizes the transformation of technological value into social benefits and enhances the resilience of the public health system.

[0330] The system implementation effectiveness verification system establishes a scientific verification chain for technological effectiveness through a three-dimensional verification framework encompassing predictive performance, prevention and control effectiveness, and social value. Verification results drive system optimization and technology transfer, forming a value loop between research and development and application. This module provides an empirical basis for system promotion.

[0331] Summary of Implementation Examples:

[0332] This system constructs a complete closed loop of "data perception → feature extraction → risk prediction → early warning issuance → intervention execution → effect feedback → model evolution": It establishes a panoramic view of the community epidemic through real-time perception of multi-source data; high-dimensional spatiotemporal features are extracted by a feature factory to drive a deep learning model to generate dynamic risk predictions; the intelligent early warning engine triggers tiered response strategies accordingly; community terminals execute scenario-based precise interventions and provide real-time feedback on the effects; the model optimization module dynamically adjusts parameter weights through an adaptive learning mechanism. All links collaborate seamlessly through standardized interfaces, overcoming the technical bottlenecks of traditional early warning systems in data fusion, community adaptation, and response timeliness, achieving full-chain value coverage from early risk identification to the implementation of prevention and control measures.

Claims

1. An intelligent influenza early warning system based on community multimodal data fusion, characterized in that, It includes: Data acquisition module: used to obtain medical and health data, environmental data, population activity data, and network behavior data at the community level in real time; Data processing module: used to perform spatio-temporal synchronization, heterogeneous field standardization, and feature fusion on the medical and health data, environmental data, population activity data, and network behavior data to generate a structured input data set; AI prediction module: uses a deep learning model to process the structured data set and outputs influenza risk scores and incidence predictions; Early warning response module: generates graded early warning signals according to the risk scores and pushes intervention instructions to the target terminals.

2. The influenza intelligent early warning system based on community multimodal data fusion according to claim 1, characterized in that, The data processing module performs the following operations: Performs time series indexing on the number of influenza-like illness (ILI) cases and vaccination rates in the medical and health data; Performs one-hot encoding on the aggregation events in the population activity data; Performs standardization processing on the search index in the network behavior data; Generates composite features of meteorological data and social activity data, and processes keywords in the network behavior data through word embedding technology; Extracts the influence features of historical data on future influenza transmission through time lag effects.

3. The influenza intelligent early warning system based on community multimodal data fusion according to claim 1, characterized in that, The AI prediction module includes a cascaded deep learning architecture: LSTM sub-module: used to capture the temporal evolution law of influenza cases; CNN sub-module: used to extract spatial transmission features between communities; Attention mechanism sub-module: used to enhance the mutation detection sensitivity of LSTM and CNN; Among them, mean square error (MSE) and mean absolute error (MAE) are used as loss functions for model training, and the Dropout mechanism and early stopping strategy are applied for optimization.

4. The influenza intelligent early warning system based on community multimodal data fusion according to claim 1, characterized in that, The graded early warning response unit triggers responses according to the following rules: When the risk score X ≤ 20, output a blue early warning signal and perform daily monitoring; When the risk score 20 < X ≤ 40, output a green early warning signal and trigger vaccination reminders and enhanced prevention and control; When the risk score 40 < X ≤ 60, output a yellow early warning signal and trigger public protection tips; When the risk score 60 < X ≤ 80, output an orange early warning signal and trigger restricted flow in gathering places; When the risk score X > 80, output a red early warning signal and trigger instructions to postpone large-scale events.

5. The influenza intelligent early warning system based on community multimodal data fusion according to claim 1, characterized in that, The medical and health data includes the number of ILI cases in community clinics, vaccination rates, and respiratory visit records; the environmental data includes temperature, humidity, temperature difference, and wind speed; the population activity data includes holidays, school opening days, and passenger flow density in business districts; the network behavior data includes influenza-related keyword search indexes.

6. The influenza intelligent early warning system based on community multimodal data fusion according to claim 3, characterized in that, The CNN sub-module is used to extract the following features, including: Spatial transmission relationships between communities; Local patterns of meteorological data; Spatio-temporal distribution characteristics of population aggregation events.

7. The influenza intelligent early warning system based on community multimodal data fusion according to claim 4, characterized in that, The target terminals for pushing intervention instructions by the early warning response module include: Community residents' mobile devices; Community health service center monitoring platforms; Intelligent information release terminals in public areas.

8. The influenza intelligent early warning system based on community multimodal data fusion according to claim 1, characterized in that, The release coverage rate of the early warning response module is quantified by the following indicators: Terminal arrival rate of early warning information in the target community; Interaction rate and feedback data on social media platforms.

9. The influenza intelligent early warning system based on community multimodal data fusion according to claim 2, characterized in that, The feature fusion includes: Generating composite features of meteorological data and social activity data; Keywords in online behavior data are processed using word embedding technology.

Images