Infection risk dynamic assessment method based on deep learning

Abstract

The invention discloses an infection risk dynamic assessment method based on deep learning, and relates to the technical field of intelligent monitoring. Multi-source heterogeneous data are integrated by calling a pre-trained deep learning model, a basic regeneration number (R0) value set is calculated based on an SEIR infection model, the basic regeneration number (R0) value set serves as a node initial feature of a space-time diagram network, and a double-layer dynamic network is constructed to generate a regional vulnerability scoring matrix. And calculating a global infection risk index through the linear combination vulnerability score and the R0 numerical set, and generating a prediction infection density matrix and a hierarchical prevention and control strategy when the index exceeds a preset threshold. And further comparing a prediction result with actually measured infection data, dynamically adjusting a symptom keyword weight and a medical resource dependency parameter, and triggering online incremental learning of the model to optimize evaluation precision. Dynamic adaptation of multi-source data fusion modeling and prevention and control strategies is achieved, and the timeliness of infection risk assessment and the scientificity of prevention and control decisions are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to a method for dynamic infection risk assessment based on deep learning. Background Art

[0002] In the monitoring and prevention of public health events, existing infection risk assessment technologies have significant shortcomings. Traditional methods rely heavily on static infection models and isolated data analysis, making it difficult to cope with the dynamic spread of infection risks and the real-time demands of complex multi-source data. The SEIR infection model based on fixed parameters cannot effectively capture the instantaneous impact of factors such as population mobility and virus mutations, resulting in prediction results lagging behind the actual development of the epidemic. In addition, symptom monitoring data and the distribution of susceptible populations are often separated, the semantic features of symptom keywords are not fully explored, and there is a lack of fine-grained modeling of the spatial dynamic changes of susceptible populations, causing risk assessment results to deviate from actual transmission trends.

[0003] Existing technologies lack the ability to integrate heterogeneous data from multiple sources, further exacerbating assessment bias. Differences in protocols and formats make cross-modal alignment of real-time mobile data such as mobile phone signaling and traffic checkpoints difficult with medical records and inventory data. The rigidity of resource scheduling mechanisms is also a prominent issue. Hospital bed and protective material allocation strategies are often based on historical experience or static thresholds and cannot be dynamically adjusted based on real-time infection risks.

[0004] To address the above issues, this field urgently needs a dynamic infection risk assessment method that can deeply integrate multi-source dynamic data, realize collaborative modeling of infection transmission mechanisms and real-time risk situations, and support closed-loop feedback optimization, so as to break through the limitations of static models and isolated data analysis, and improve the timeliness of infection risk prediction and the dynamic adaptability of prevention and control strategies. Summary of the Invention

[0005] In response to the above-mentioned shortcomings of the prior art, the present invention provides a dynamic infection risk assessment method based on deep learning.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0007] The deep learning-based dynamic infection risk assessment method includes the following steps:

[0008] S1. Call the pre-trained deep learning model to obtain multi-source heterogeneous data in real time, perform cross-modal alignment processing, and output the real-time basic reproduction number (R0) R0 value set based on the SEIR infection model;

[0009] S2. Using the R0 value set as the node initial feature vector of the space-time graph network, constructing a two-layer space-time graph network based on the dynamic flow risk matrix, and generating a regional vulnerability score matrix in combination with the multi-source heterogeneous data;

[0010] S3. Calculate the global infection risk index based on a linear combination of the regional vulnerability score matrix and the R0 value set;

[0011] S4. When it is detected that the global infection risk index exceeds a preset infection threshold, a predicted infection density matrix and a hierarchical prevention and control strategy including prevention and control level parameters are generated;

[0012] S5. Compare the predicted infection density matrix with the measured infection density matrix, dynamically update the weight parameter according to the comparison result, and output it to the execution terminal, wherein:

[0013] Calculating a spatial overlap index between the predicted infection density matrix and the measured infection density matrix;

[0014] When the spatial overlap is lower than a preset threshold, the weight of the symptom keyword is increased proportionally;

[0015] When the consumption rate of medical resources continues to exceed the consumption threshold, the dependency parameter of the associated area is reduced;

[0016] When the correction direction is consistent, online incremental learning of the model is performed to trigger retraining of the deep learning model.

[0017] Beneficial effects

[0018] Compared with the known public technology, the technical solution provided by the present invention has the following beneficial effects:

[0019] 1. Based on the coupled architecture of the SEIR infection model and a spatiotemporal graph network, this invention organically combines the virus transmission mechanism with real-time data features, significantly improving the spatiotemporal resolution of infection risk prediction. Dynamic parameter calibration and incremental model learning driven by a closed-loop feedback mechanism provide the system with continuous adaptive optimization capabilities, breaking through the rigid limitations of traditional static models.

[0020] 2. The present invention adopts cross-modal attention fusion technology to nonlinearly associate the semantic features of symptom keywords with the statistical features of the distribution of susceptible populations, dynamically capture the evolution trend of the epidemic through the time attenuation factor, improve the sensitivity of early clustered epidemic identification, and reduce the underreporting rate of symptom signals. The innovative hierarchical prevention and control strategy generation mechanism incorporates multi-dimensional indicators such as medical resource status and environmental risks into the dynamic decision-making system, realizing the full link from data perception to prevention and control execution. Through the nonlinear fusion of symptom semantic features and regional vulnerability scores, the potential correlation patterns of epidemic spread are effectively captured, providing a scientific basis for precise prevention and control.

[0021] 3. The present invention is applicable to precise prevention and control decision support under sudden epidemic situations. It realizes precise intervention through a layered dynamic prevention and control mechanism and a two-layer spatiotemporal graph network: the bottom graph tracks the exposure risk of high-risk places in real time, and the upper graph dynamically adjusts the transmission weight between regions in combination with the medical resource dependence. It cooperates with the hierarchical response strategy to realize flexible resource allocation, supports seamless connection with the smart city platform, and reduces deployment costs through edge computing nodes. Through the collaborative innovation of multi-source data drive, dynamic network modeling and closed-loop feedback optimization, an intelligent infection prevention and control technology system has been constructed. While improving the accuracy of risk assessment, it also enhances the flexibility and scalability of prevention and control strategies, providing strong technical support for responding to complex and changing public health events. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] To more clearly illustrate the technical solutions of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0023] Figure 1 Schematic diagram of the steps of the present invention;

[0024] Figure 2 This is a pre-screening flow chart of the present invention;

[0025] Figure 3 This is a flow chart of the cross-modal alignment process of the present invention;

[0026] Figure 4 This is a flow chart of the hierarchical prevention and control strategy of the present invention. DETAILED DESCRIPTION

[0027] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0028] Application Overview:

[0029] Existing technologies often rely on static transmission models or single-dimensional data analysis, making it difficult to achieve accurate predictions in dynamic scenarios. Traditional methods fail to effectively integrate multi-source data such as population mobility, environmental viral load, and the status of medical resources, resulting in models unable to respond to changes in the epidemic's spread in real time. Especially during public health emergencies, existing technologies lack the ability to model the spatiotemporal heterogeneity of virus transmission pathways and the lags in resource allocation. This results in inadequate adaptability of prevention and control strategies, which can easily lead to resource runs and control loopholes.

[0030] To address these issues, the inventors discovered a dynamic correlation between virus transmission intensity, population mobility patterns, and environmental virus concentrations. By constructing a coupled architecture between the SEIR infection model and a spatiotemporal graph network, they achieved collaborative modeling of mechanisms and data. During their research, they discovered that medical resource dependence has a nonlinear regulatory effect on interregional transmission risk, and that the spatiotemporal distribution characteristics of symptom keywords can effectively characterize early infection clustering signals. Consequently, they proposed a closed-loop assessment mechanism based on the dynamic fusion of multi-source data to establish a real-time mapping relationship between transmission risk and prevention and control strategies.

[0031] By acquiring mobile phone signaling data in real time, a population flow matrix is generated, and environmental virus monitoring data and medical resource status information are collected simultaneously. The SEIR infection model is used to calculate the basic reproduction number (R0) as the node feature of the spatiotemporal graph network, and the regional vulnerability score matrix is constructed in combination with symptom keyword extraction. When the global infection risk index exceeds the preset threshold, the system automatically triggers a graded prevention and control strategy, dynamically allocates medical resources, and outputs graded control instructions. During the continuous monitoring process, by comparing the spatiotemporal overlap of the predicted infection density and the distribution of actual cases, the symptom keyword weight coefficient and resource dependency parameters are corrected in real time, and the incremental learning mechanism is used to optimize the model prediction accuracy. For low-risk areas, the standard monitoring mode is continuously used to update the risk assessment results.

[0032] Compared with traditional technologies, existing technologies lack deep collaboration and dynamic feedback mechanisms for multi-source heterogeneous data, resulting in risk assessment lagging behind the actual development of the epidemic. This application innovatively establishes a two-way parameter transfer architecture for the SEIR-space-time graph network to achieve fusion modeling of virus transmission mechanisms and real-time data features. Different from traditional single-dimensional analysis, this application captures early infection signals through symptom semantic analysis, combined with a dynamic adjustment mechanism for resource dependence, significantly improving the reliability of risk assessment in complex scenarios. The closed-loop feedback system breaks through the limitations of static models and enables prevention and control strategies to have the adaptive ability to continuously optimize.

[0033] Through the above technical solutions, this application effectively addresses the issues of isolated multi-source data and rigid prevention and control strategies, ensuring real-time response while improving risk assessment accuracy. The dynamic parameter correction mechanism balances model sensitivity and stability, and the ability to model spatiotemporal heterogeneity provides scientific decision-making support for emergency response to major public health events, making it particularly suitable for precise epidemic prevention and control scenarios in highly mobile cities.

[0034] After introducing the basic concept of the present invention, embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] Example

[0036] The dynamic infection risk assessment method based on deep learning in this embodiment is as follows: Figure 1 As shown, the following steps are included:

[0037] S1. Call the pre-trained deep learning model to obtain multi-source heterogeneous data in real time, perform cross-modal alignment processing, and output the real-time basic reproduction number (R0) R0 value set based on the SEIR infection model;

[0038] S2. Using the R0 value set as the node initial feature vector of the space-time graph network, a two-layer space-time graph network is constructed based on the dynamic flow risk matrix, and a regional vulnerability score matrix is generated by combining multi-source heterogeneous data;

[0039] S3. Calculate the global infection risk index based on the linear combination of the regional vulnerability score matrix and the R0 value set;

[0040] S4. When it is detected that the global infection risk index exceeds the preset infection threshold, a predicted infection density matrix and a hierarchical prevention and control strategy including prevention and control level parameters are generated;

[0041] S5. Compare the predicted infection density matrix with the measured infection density matrix, dynamically update the weight parameters based on the comparison results, and output them to the execution terminal, where:

[0042] Calculate the spatial overlap index between the predicted infection density matrix and the measured infection density matrix;

[0043] When the spatial overlap is lower than the preset threshold, the weight of the symptom keyword is increased proportionally;

[0044] When the consumption rate of medical resources continues to exceed the consumption threshold, the dependency parameter of the associated area is reduced;

[0045] When the correction direction is consistent, the model is incrementally learned online, triggering retraining of the deep learning model.

[0046] Specifically, in step S1, multi-source heterogeneous data refers to data sets from different channels with different formats and features, such as hospital case data, community health monitoring data, and traffic flow data. These data cover a rich range of information dimensions. The pre-trained deep learning model is a model that has been trained in advance using a large amount of relevant data and has powerful data processing and feature extraction capabilities. The SEIR infection model is a classic infection transmission dynamics model that models the dynamic changes of susceptible (S), exposed (E), infected (I), and recovered (R) to estimate the infection's transmissibility indicator R0. R0 represents the average number of people an infected person will infect during the course of the disease, assuming no external intervention and a generally susceptible population. This R0 value set can reflect the potential transmission intensity of the current infection. This step processes the multi-source heterogeneous data through a deep learning model and combines it with the SEIR infection model to output the key R0 value set, providing important basic data for subsequent evaluation.

[0047] In step S2, a spatiotemporal graph network and a regional vulnerability score matrix are constructed. The R0 value set is used as the initial feature vector of the nodes of the spatiotemporal graph network through feature embedding or normalization formulas. The R0 value set is used as the initial feature vector of the nodes of the spatiotemporal graph network. The spatiotemporal graph network is a network structure that can simultaneously consider information in both time and space dimensions, and can effectively describe the spread of infection in different regions over time. A two-layer spatiotemporal graph network is constructed based on the dynamic flow risk matrix. The dynamic flow risk matrix is an infection risk matrix that reflects the flow of people between different regions by comprehensively considering factors such as population mobility and transportation. The regional vulnerability score matrix is generated by combining multi-source heterogeneous data. The regional vulnerability score matrix quantifies the vulnerability of different regions to the spread of infection through a comprehensive analysis of various data, providing a basis for the subsequent calculation of the infection risk index.

[0048] In step S3, the regional vulnerability score matrix and the R0 value set are linearly combined using a well-designed weighted calculation method to integrate the regional vulnerability score and R0 value. This linear combination method comprehensively considers multiple factors, including the region's own vulnerability and the potential for infection spread, to accurately calculate the global infection risk index, which comprehensively reflects the overall infection risk status of the entire region.

[0049] The preset infection threshold in step S4 is a risk warning line established based on historical data, expert experience, and other factors. A predicted infection density matrix is generated. This matrix, based on the global infection risk index and related models, further infers the possible future infection density distribution in different regions. A hierarchical prevention and control strategy, including prevention and control level parameters, is also generated. This strategy sets specific prevention and control measures based on different infection risk levels, such as restrictions on personnel movement, social distancing requirements, and testing frequency, to achieve precise prevention and control.

[0050] Step S5 calculates the spatial overlap index of the predicted infection density matrix and the measured infection density matrix. The spatial overlap index is used to measure the degree of consistency between the predicted results and the actual situation in terms of spatial distribution. When the spatial overlap is lower than the preset threshold, the symptom keyword weight is increased proportionally. The increase in the symptom keyword weight can enable the model to pay more attention to the data features related to the symptoms in subsequent calculations, thereby optimizing the model's assessment of the infection risk. When the consumption rate of medical resources continues to exceed the consumption threshold, the consumption rate of medical resources reflects the pressure on the medical system to cope with the epidemic. Reduce the dependency parameter of the associated area. By reducing the dependency parameter, the model's assessment of the correlation between different regions can be adjusted to be more in line with the actual infection risk changes under tight medical resource conditions. When the correction direction is consistent continuously, that is, multiple adjustments of the model are all in the same direction that is conducive to improving the accuracy of the prediction, the model is subjected to online incremental learning. By using new data and adjustment information to further optimize the model, the deep learning model is retrained, so that the model can continuously adapt to new situations such as changes in the epidemic situation, thereby improving the accuracy of the assessment.

[0051] The core innovation of this application lies in the construction of a closed-loop risk assessment system based on SEIR-space-time graph network coupling. Through dynamic fusion of multi-source data and feedback optimization mechanisms, it overcomes the spatiotemporal resolution limitations of traditional static models. It innovatively establishes a collaborative modeling system for virus transmission mechanisms and deep learning features, achieving bidirectional optimization for minute-by-minute infection risk assessment and adaptive generation of prevention and control strategies.

[0052] The working process and principle of this application are as follows: first, call the pre-trained deep learning model to process multi-source heterogeneous data, and output the real-time R0 value set through the improved SEIR infection model; use the R0 value set as the node feature of the spatiotemporal graph network to construct a two-layer dynamic network reflecting the virus transmission path; calculate the global infection risk index based on the regional vulnerability score matrix and real-time R0 data, and generate a predicted infection density map and graded prevention and control instructions when the index exceeds the dynamic threshold; dynamically correct the symptom semantic weights and resource dependency parameters by comparing the predicted results with the actual case data of the disease control system, and trigger the incremental learning of the model to optimize the prediction accuracy; continue to adopt the standard monitoring mode for low-risk areas, and maintain the dynamic balance between risk assessment and prevention and control execution through closed-loop feedback. This application significantly improves the response speed and handling accuracy of public health emergencies, and provides reliable technical support for smart city epidemic prevention and control.

[0053] Traditional methods rely on single-dimensional data analysis and fixed response thresholds, and are unable to cope with the spatiotemporal heterogeneity of virus transmission. This application innovatively establishes a multi-scale data fusion architecture, accurately capturing the micro-macro transmission patterns of epidemic spread through bidirectional modeling of venue-level contact networks and regional-level transmission networks. Unlike existing technologies, this application possesses dynamic feature reconstruction capabilities and model self-optimization characteristics, adapting to the complex prevention and control needs of different transmission scenarios.

[0054] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0055] S1. Obtain medical data such as the number of newly confirmed cases, suspected cases, and patient symptom descriptions in the past week from the information systems of major local hospitals; obtain daily passenger flow data of major transportation hubs (such as airports, railway stations, bus stations, etc.) and personnel flow trajectory data between different regions from the transportation department; obtain resident health monitoring data of each community from community health service centers, including the number of people with abnormal body temperature and respiratory symptoms.

[0056] These data in different formats and from different sources are input into a pre-trained deep learning model. The pre-trained deep learning model adopts a 12-layer Transformer architecture with an initial learning rate set to 0.001. The training data includes a multi-source epidemic data set of 200 cities around the world from 2019 to 2023. After training with a large amount of historical epidemic-related data and other relevant social data, it has powerful feature extraction and data fusion capabilities.

[0057] The processed data is analyzed based on the SEIR infection model, whose parameters are initialized based on local demographics, social behavior, and other factors. For a city with a population of 5 million, the initial proportion of susceptible individuals is set at 90%, the proportion of exposed individuals at 5%, the proportion of infected individuals at 3%, and the proportion of recovered individuals at 2%. The model calculates and outputs a real-time set of R0 values, representing the basic reproduction number (R0). In this example, the R0 value set is [2.2, 2.3, 2.1], indicating that the potential spread of the infection in the city is currently high.

[0058] S2. Use the R0 value set obtained in S1 as the initial feature vector of the nodes in the space-time graph network to construct a two-layer space-time graph network. The network is constructed based on the dynamic flow risk matrix, which comprehensively considers factors such as the frequency and distance of personnel flow, and the population density of different regions and places. For two regions with frequent personnel flow, close distances, and high population density, the corresponding values in the dynamic flow risk matrix are higher. Combining multi-source heterogeneous data to generate a regional vulnerability score matrix, areas with relatively scarce medical resources, a high proportion of elderly population, and poor community health conditions are given higher vulnerability scores in the regional vulnerability score matrix. After calculation, the regional vulnerability score of this area is 0.8, while the vulnerability score of another area with abundant medical resources and a large young population is 0.3.

[0059] S3. Calculate the global infection risk index based on a linear combination of the regional vulnerability score matrix and the R0 value set. For example, using a weighted average, assign a weight of 0.6 to the regional vulnerability score and a weight of 0.4 to the average of the R0 value set. This calculation yields a global infection risk index of 0.65 for the city, indicating an overall infection risk of moderate to high.

[0060] S4. When the global infection risk index is detected to exceed the preset infection threshold of 0.6, a predicted infection density matrix is generated. By analyzing historical epidemic spread data and current multi-source data, the infection density of each region is predicted for the next week. It is predicted that the infection density of a certain area may increase from the current 0.05 to 0.1. At the same time, a hierarchical prevention and control strategy containing prevention and control level parameters is generated. Different prevention and control measures are implemented for areas with different levels of infection risk, including restricting the flow of people, strengthening disinfection of public places, limiting large-scale gatherings, and focusing on daily health monitoring and publicity and education.

[0061] S5. Calculate the spatial overlap between the predicted infection density matrix and the measured infection density matrix. Assume the calculated spatial overlap is 0.4, while the preset threshold is 0.5. Since the spatial overlap is lower than the preset threshold, increase the weight of the symptom keywords proportionally. For example, increase the weight of symptom keywords such as fever and cough from 0.3 to 0.4, so that the model pays more attention to the data features related to these symptoms in subsequent evaluations.

[0062] When it is found that the consumption rate of medical resources in a certain area (such as the number of beds, number of medical staff, etc.) continues to exceed the consumption threshold (for example, the bed occupancy rate exceeds 80% for three consecutive days), the dependence parameter of the area with the surrounding high-infection risk areas will be reduced from the original 0.7 to 0.5 to reflect the impact of changes in the correlation between the area and the surrounding areas, such as the flow of personnel, on the infection risk when medical resources are tight.

[0063] When the correction direction is consistent for multiple times, for example, by adjusting the symptom keyword weights three times in a row, the spatial overlap index between the predicted infection density matrix and the measured infection density matrix is improved.

[0064] Conduct online incremental learning of the model, input newly collected multi-source heterogeneous data and relevant information during the model adjustment process as incremental data into the model, trigger retraining of the deep learning model, further optimize the model's parameters and structure, and improve the model's accuracy and adaptability to infection risk assessment.

[0065] Through the above specific embodiments, the application process of the dynamic infection risk assessment method based on deep learning in practice is demonstrated, which realizes the dynamic monitoring of infection risks and the formulation of precise prevention and control strategies.

[0066] like Figure 2 As shown, this application further proposes that multi-source heterogeneous data can be filtered out of abnormal data through pre-screening rules:

[0067] When the crowd density data exceeds the preset safety threshold, Z-Score standardization is implemented on the crowd density data, and outlier data is filtered based on the statistical anomaly detection algorithm;

[0068] Set up a dynamic sensitive vocabulary for symptom keywords, and mark them as high risk when the frequency of symptom keywords exceeds the baseline value threshold;

[0069] Set category association constraints for protective material inventory data, and only transmit category data related to respiratory disease prevention and control;

[0070] Train and optimize deep learning models, and regularly send incremental model parameters to edge nodes.

[0071] By building a hierarchical data quality control system and an adaptive model update framework, the accuracy of infection risk assessment in high-noise environments is significantly improved. On the data input side, hardware-level real-time filtering rules and dynamic constraint strategies are deployed for three core data streams: crowd density data, symptom keywords, and protective material inventory. Threshold-triggered Z-Score normalization is implemented for crowd density data, and outliers are removed in real time using an FPGA-accelerated anomaly detection algorithm. A dynamic sensitive vocabulary is constructed for symptom keywords, and a sliding window mechanism is used to calculate the frequency change rate. When the frequency of a specific symptom keyword exceeds the historical baseline threshold, a high-risk flag is automatically triggered and its weight coefficient in the feature fusion stage is increased. Category association constraints are imposed on protective material inventory data, and a predefined knowledge graph is used to screen material categories directly related to respiratory disease prevention and control (such as N95 masks and ventilator consumables). Non-relevant data is filtered before transmission. On the model optimization side, an edge-cloud collaborative training architecture is established. Deep learning model parameters are continuously optimized using an incremental learning algorithm, and the parameters are periodically transmitted to distributed edge nodes through an encrypted channel to ensure that the model adapts to the dynamic evolution of the epidemic in real time.

[0072] Specifically, crowd density data is collected jointly by smart cameras and mobile signaling, and is input into dedicated anomaly detection after Z-Score standardization. Symptom keywords are extracted from the hospital's electronic medical record system in real time, and a dynamic sensitive vocabulary is generated by analyzing historical epidemic data through a bidirectional LSTM model. When a sudden increase in the frequency of core keywords such as "fever" and "difficulty breathing" is detected, the vulnerability score of the relevant area is automatically increased. Through the built-in infection prevention and control knowledge graph, the graph neural network is used to identify the semantic correlation between supplies and disease prevention and control, and only category data with a correlation exceeding the preset correlation is allowed to enter the subsequent processing flow. During the model training phase, the cloud central server performs incremental learning once an hour, and sends it to the edge node after quantization and compression. The edge device uses a federated learning mechanism to complete model fine-tuning locally to avoid cross-domain transmission of sensitive data.

[0073] Compared with traditional methods, existing technologies usually use single threshold filtering or static vocabulary matching, resulting in a high misjudgment rate in high-volatility scenarios. This application reduces the misjudgment rate in the data preprocessing stage through hardware-accelerated real-time anomaly detection and dynamic semantic analysis. The existing model update mechanism relies on full retraining, and the parameter delivery delay is as long as several hours. This application shortens the model update delay through edge-cloud collaborative architecture and incremental learning technology. The existing protective material screening method uses a manual rule library, and the category misassociation rate is high. This application reduces the misassociation rate based on semantic association analysis of the knowledge graph.

[0074] The synergy of these advanced technical solutions has led to breakthroughs in three key areas: First, data quality: a layered filtering and pre-screening mechanism improves the signal-to-noise ratio of input data, reducing the penetration of abnormal data compared to traditional solutions. Second, model timeliness: incremental parameter delivery and instant edge fine-tuning reduce real-time errors in risk assessment results. Third, prevention and control precision: a dynamic sensitive vocabulary and vulnerability scoring mechanism integrate to improve the accuracy of high-risk area identification. These improvements together support a comprehensive enhancement system covering "data cleaning, feature optimization, and model iteration," providing a highly reliable technical infrastructure for large-scale infection prevention and control.

[0075] like Figure 3 As shown, the present application further proposes that the cross-modal alignment process includes:

[0076] Define default filling strategies based on preset protocols and design timing and spatial features;

[0077] According to the preset protocol, symptom keywords are extracted from the medical institution's case report database and medical record data, and sliding window normalization is performed to generate symptom density time series feature vectors;

[0078] One-hot encoding is performed on the census database according to the preset protocol to generate the distribution feature vector of susceptible population.

[0079] Specifically, cross-modal alignment processing is achieved through a dual-track mechanism of hardware-level time base synchronization and feature encoding normalization. For example, a high-precision clock synchronization module (error <1ms) is deployed at the data acquisition end to perform global clock calibration on the data acquisition equipment of the medical institution case database, census database, and traffic flow monitoring system. Nanosecond-level timestamp synchronization of multi-source sensors is achieved through the PTP protocol to eliminate timing misalignment caused by system response delays. For example, the collection terminals of a hospital's electronic medical record system and the city's population database achieve time base alignment through GPS timing, ensuring that the error in the start time of the collection of daily new case records and population flow data is controlled within 0.5ms.

[0080] Dedicated data transmission links are configured for medical text data and demographic data respectively. When extracting symptom keywords from the HIS system, a double-buffered DMA architecture is used to write directly to the memory to avoid transmission jitter in the traditional interrupt-driven mode. The symptom frequency is dynamically smoothed through sliding window normalization (for example, the window length is 24 hours and the step length is 1 hour) to generate a symptom density time series feature vector. For example, for core symptom words such as "fever" and "cough", the frequency of occurrence within each hour is normalized by Z-Score and mapped to the [-1,1] interval to eliminate the interference of sudden outliers. Discrete features such as age and vaccination status of the census database are converted into 64-dimensional vectors in real time through the One-Hot encoding engine, and the spatial distribution matrix is generated by linking the administrative division GIS coordinates. In addition, through built-in sparse processing, secondary features with low coverage are automatically filtered to improve data representation efficiency.

[0081] A spatiotemporal cross-attention mechanism is deployed at the fusion layer to map the time series feature vectors of symptom density and the feature vectors of susceptible population distribution to a unified spatiotemporal coordinate system. An LSTM time series alignment network compensates for differences in sampling frequencies between different data sources (e.g., medical data is updated every minute, while population data is updated daily), generating a synchronized timestamp sequence. A regional association topology is constructed based on the Delaunay triangulation algorithm, interpolating the feature vectors of discrete administrative units onto a continuous spatial grid. For example, when population data is missing for a region, the interpolation result is generated by weighting the feature vectors of three adjacent administrative districts (weight = 1 / distance²).

[0082] Traditional cross-modal alignment schemes rely on manually set time windows for post-processing and splicing (e.g., segmenting data by day or hour), which leads to excessive accumulation of temporal phase deviations. Spatial interpolation uses a simple linear averaging method, which is sensitive to non-uniform distributions. Existing feature fusion methods, which fail to consider semantic alignment, are prone to spurious correlations between "cough frequency" and "elderly population density" and do not involve the calculation of feature association weights using attention mechanisms.

[0083] Through the above technical solutions, this application further proposes a cross-modal spatiotemporal alignment framework based on multi-source heterogeneous data, defines a unified temporal-spatial feature fusion benchmark through a preset protocol, and constructs a standardized data processing channel to solve the information fusion distortion problem caused by the data acquisition timing deviation and feature semantic gap in traditional methods. The temporal overlap between the symptom density fluctuation curve and the migration trajectory of susceptible populations is improved. Through the feature gridding accuracy, high-risk areas at the community level (such as a sudden increase in vulnerability scores within 500 meters around a nursing home) can be identified. The risk assessment model is trained based on the aligned joint data set to improve the prediction accuracy (F1-Score) in scenarios such as influenza A outbreaks. A full-link closed loop from data acquisition to feature fusion is constructed to provide high-precision, low-latency cross-modal data support for infection risk assessment.

[0084] The present application further proposes that the generation of the dynamic liquidity risk matrix in step S2 includes:

[0085] Divide mobile phone signaling data into time-space grids with a predetermined number of digits and calculate the pedestrian density of each grid unit;

[0086] Statistics of traffic density change rate according to preset time window;

[0087] The current risk value is generated by the weighted formula:

[0088] Flow risk value = pedestrian density × moving speed + vehicle density change rate × road weight coefficient,

[0089] The road weight coefficient is preset according to the road grade.

[0090] The spatial and temporal gridding uses high-precision geofencing technology to spatially discretize mobile phone signaling data (for example, dividing the target area into hexagonal honeycomb grids with a side length of 50 meters). The pedestrian density within each grid cell is calculated in real time using a mobile terminal spatial distribution kernel density estimation algorithm. The vehicle density change rate is based on floating vehicle trajectory data from the intelligent transportation system (for example, the gradient of vehicle frequency changes on each arterial road segment within a 5-minute time window) and normalized with the road capacity coefficient. The weighted calculation of the mobility risk value utilizes a heterogeneous data stream parallel processing architecture, allocating independent computation channels for pedestrian density data and applying sliding window mean filtering. A dedicated transmission link is established for vehicle data, performing Kalman filtering for noise reduction. Both data are time-aligned to the millisecond level through hardware-level clock synchronization. Road weight coefficients are dynamically configured (for example, highways, urban expressways, and arterial roads are assigned baseline weights of 0.8, 0.6, and 0.4, respectively, and dynamically fluctuate within a ±0.2 range based on the real-time accident rate).

[0091] Compared with traditional solutions, existing technologies usually use fixed grid division and static weight allocation mechanisms, which have problems such as insufficient spatiotemporal resolution (commonly used 100-meter grids) and fragmented risk factor calculations. This application reduces spatial positioning errors through grid design and combines a dual filtering mechanism to suppress noise in pedestrian density calculations. The dynamic road weight configuration system for traffic data effectively solves the technical defect that traditional fixed weights cannot adapt to sudden traffic incidents. Through deep learning models, the nonlinear relationship between pedestrian speed and traffic changes is explored to improve the accuracy of risk prediction.

[0092] This application further proposes a method for generating a dynamic flow risk matrix based on collaborative calculation of multi-source spatiotemporal data, which realizes the refined integration of pedestrian and vehicle flow risk factors by constructing a spatiotemporal grid cascade dynamic weighted model. Through the three-level collaboration of spatiotemporal grid optimization, heterogeneous data processing, and dynamic weight configuration, the sub-minute update capability of the flow risk value is achieved. The hardware-level data channel isolation design controls the variance of the collection and transmission delay of pedestrian and vehicle flow data, and the microsecond-level clock synchronization mechanism compresses the time reference error of multi-source data. In actual deployment, the dynamic flow risk matrix generation system has successfully shortened the warning time of epidemic spread in key areas of the city, and the grid-level risk value refresh frequency provides a high-spatiotemporal resolution decision-making basis for the formulation of precise prevention and control strategies.

[0093] This application further proposes that the two-layer spatiotemporal graph network is divided into a bottom-layer graph and an upper-layer graph. The bottom-layer graph represents the place contact network, and the edge weights are calculated by accumulating crowd density data and movement speed; the upper-layer graph represents the regional association network, and the edge weights are calculated by population migration rate and medical resource dependence.

[0094] Specifically, this application constructs a two-layer spatiotemporal graph network to dynamically couple micro-site contact risks with macro-regional transmission paths, achieving multi-scale and accurate assessment of infection risks. The underlying site contact network generates edge weights using real-time collected site-level crowd density and movement speed data combined with a dynamic flow risk matrix, specifically using a weighted formula for calculation:

[0095] W ij =ρ i ·v j +ΔC k ·ω r

[0096] Among them, ρ i represents the crowd density of place i (unit: people / square meter), v j Indicates the moving speed (m / s), ΔC k is the rate of change of traffic density, ω r The system uses a preset road weight coefficient (dynamically adjusted according to road grade). Data is collected in real time through multiple sensors (such as video surveillance and mobile phone signaling), and a hardware-level clock synchronization mechanism (such as a GPS timing module) is used to ensure that the time alignment error of different data sources is less than 1 millisecond, avoiding the timing deviation caused by operating system delays in traditional software-level timestamps.

[0097] The upper-level regional association network dynamically generates edge weights based on the cross-regional population migration rate (based on traffic checkpoint data statistics) and the medical resource dependency. The dependency calculation formula is:

[0098] D m =F m ·(1-Um )

[0099] Among them, F m is the historical frequency of medical visits in region m (based on hospital HIS system data), U m The current resource intensity (ratio of ICU bed occupancy rate to inventory consumption rate) is calculated. Medical resource data and population migration data are transmitted in parallel through independent DMA channels, ensuring that the data transmission delay variance is controlled within 0.5 seconds, avoiding resource contention issues in a shared bus architecture.

[0100] The underlying network and the upper network achieve cross-layer risk transmission through a preset propagation attenuation coefficient (γ). For example, when the exposure risk of a lower-level location exceeds a threshold, the risk value is mapped to the upper-level associated area at a decay rate of γ = 0.8, triggering adjustments to regional prevention and control strategies. Simultaneously, the medical resource dependency parameters of the upper-level network are fed back to the underlying network in real time, dynamically adjusting the flow restriction strategy for the location (for example, reducing the flow density threshold in the area surrounding the hospital by 50%).

[0101] Compared with traditional schemes, existing methods usually adopt a single-layer static network (such as based only on population density), which cannot distinguish between dynamic contacts at the place level and resource dependence differences at the regional level, resulting in excessively high risk assessment errors.

[0102] Through the above technical solution, the error in global infection risk prediction is reduced and the real-time strategy response time is shortened through the synergy of the two-layer network. The bottom graph is based on hardware synchronization and independent DMA channels to achieve millisecond-level alignment of pedestrian flow, vehicle flow, and medical data, avoiding the phase offset problem of traditional interpolation compensation. The micro-macro risk transmission is quantified by the attenuation coefficient. For example, when the flow of people in a shopping mall surges, it is automatically associated with the surrounding community and triggers an early warning. The upper-level graph’s medical resource tension (such as ICU occupancy rate >85%) corrects the upper-level network edge weights in real time, and reversely regulates the contact risk threshold of the underlying venues. This application establishes a spatiotemporal consistency benchmark at the source of data collection to improve the accuracy of infection risk prediction and support the precise triggering of multi-level prevention and control strategies.

[0103] This application further proposes that the calculation of medical resource dependence includes:

[0104] Obtain historical service coverage data for healthcare institutions;

[0105] Count the frequency of medical visits in each area within the preset period;

[0106] Generate medical resource dependency through dynamic formula:

[0107] Dependence = frequency of visits × (1-current resource shortage)

[0108] The resource tension is calculated based on the ratio of inventory consumption rate to the safety threshold.

[0109] Among them, historical service coverage data extracts patient source distribution information through the electronic medical record system of medical institutions, and uses spatial clustering algorithms to construct regional service radius heat maps to quantify the medical service radiation intensity of each region within a preset period (such as the past 3 months). The frequency of medical visits statistics collects the registration volume of each department, the number of emergency admissions and the turnover rate of inpatient beds in real time through the hospital HIS system, and calculates the average daily number of medical visits per 10,000 people in combination with regional population density data to generate a medical pressure index in the spatiotemporal dimension. The dynamic formula calculation defines resource tension as the dynamic ratio of the current medical supplies inventory consumption rate to the safety inventory threshold, realizing real-time coupling of medical demand and resource supply.

[0110] Specifically, dynamic medical resource dependency calculation is achieved through the following collaborative mechanisms:

[0111] The hospital's material management system uploads inventory data of key materials such as protective clothing and medicines to the cloud database in real time, and regularly updates the consumption rate; regional population flow data is counted through mobile phone signaling positioning information, and the base number for calculating the frequency of medical visits is dynamically corrected.

[0112] Dynamic Modeling Layer: Historical service coverage data and real-time visit frequency are integrated through a spatiotemporal interpolation algorithm to generate a baseline for regional medical service demand. A sliding window mechanism (e.g., a 24-hour window length) is used to monitor inventory trends. For example, if the consumption rate exceeds 1.2 times the safety threshold for three consecutive hours, an exponential growth mode of resource tension is triggered (e.g., a coefficient jumps from 0.5 to 0.9).

[0113] When the dependence of a certain area exceeds the preset critical value, the association weight between the area and the surrounding high-mobility areas will be automatically reduced to suppress the potential risk of resource runs. At the same time, a resource warning signal will be sent to trigger cross-regional medical supplies dispatch instructions.

[0114] Compared with the traditional static weight allocation method, the traditional scheme relies on medical resource allocation records at a fixed period (such as monthly) and cannot perceive the surge in resource consumption caused by sudden epidemics (such as a 300% increase in daily visits).

[0115] Through the above technical solution, this application further proposes a dynamic calculation model for medical resource dependency based on spatiotemporal coupling. By constructing a closed-loop feedback mechanism that integrates historical service coverage data, real-time resource consumption rates, and regional medical demand, it generates dynamically evolving medical resource dependency parameters. This application transforms medical resource assessment from a passive response to an active prediction, significantly improving resource utilization efficiency during major public health events.

[0116] This application further proposes that the construction of a two-layer spatiotemporal graph network includes:

[0117] The underlying graph edge weights implement a dynamic update mechanism that periodically recalculates the exposure risk based on the crowd density;

[0118] The upper-level graph edge weights enforce the constraint of automatically reducing the population migration rate weight of the associated region when the resource tension exceeds the preset resource threshold;

[0119] The bottom-layer contact risk is mapped to the upper-layer propagation path through the preset propagation attenuation coefficient for cross-layer connection.

[0120] Specifically, the edge weight update mechanism of the underlying graph uses an array of IoT sensors deployed in public places across the city to collect real-time data on crowd density and movement speed. It generates a contact risk index based on the dynamic flow risk value calculation formula (flow risk value = crowd density × movement speed + traffic change rate × road weight), and triggers the reconstruction of edge weights across the entire graph using a unified time window (e.g., a 5-minute cycle). For example, if a business district detects a sudden increase in crowd density to 5 people per square meter and an average movement speed of less than 0.5 meters per second, the edge weight connecting adjacent areas will be increased to 1.8 times the baseline value in the next update cycle, reflecting the risk of clustered transmission in real time.

[0121] The edge weight constraint engine in the upper-level graph continuously accesses resource stress indicators such as bed occupancy rates and ventilator availability rates at medical institutions. (For example, when the ICU bed utilization rate in a certain area exceeds a preset threshold of 85%, the population migration weight of that area and surrounding areas is automatically reduced by 30%-50%), thereby suppressing the cross-regional transmission effect in areas with overloaded medical resources. By introducing a dynamic transmission attenuation coefficient (e.g., α = 0.1 for areas surrounding hospitals and α = 0.5 for transportation hubs), the underlying location-level exposure risk is mapped to the upper-level regional transmission path according to the exponential decay law.

[0122] Through real-time risk perception of the underlying place contact network and resource constraint modeling of the upper-level regional association network, combined with quantitative mapping of cross-layer risk transmission paths, an infection risk prediction system with linked time and space dimensions is constructed.

[0123] In a specific embodiment, when an epidemic cluster occurs in a stadium, the underlying graph network detects abnormal crowd flow and increases edge weights within 15 minutes. The upper-layer graph network simultaneously reduces the weight of outward migration in the area by 45%. The cross-layer module transmits the venue risk to the administrative district-level network with a coefficient of α = 0.4, and finally generates prevention and control instructions including the lockdown of surrounding communities and the pre-allocation of hospital beds within 30 minutes, which is 4 times faster than the response speed of traditional solutions.

[0124] Compared with traditional solutions, existing single-graph network modeling methods often statically solidify the risk transmission relationship between places and regions, and cannot capture the real-time impact of dynamic changes in human flow on the transmission path. This application improves the timeliness of place contact risk identification through the minute-level weight update mechanism of the underlying graph; the resource constraint response mechanism of the upper-level graph improves the efficiency of transmission suppression in areas with tight medical resources. Existing cross-layer mapping methods mostly use fixed transmission coefficients, which are difficult to adapt to the differences in transmission characteristics under different scenarios. The dynamic attenuation coefficient of this application can be adaptively adjusted according to historical transmission data to reduce the cross-layer risk prediction error. Existing systems often have warning lags in resource-constrained scenarios, and this application shortens the prevention and control response time in high-risk areas through real-time weight constraints.

[0125] This invention innovatively establishes a three-level closed-loop optimization system: dynamic site risk perception, regional resource constraint regulation, and cross-layer attenuation transmission. The underlying IoT sensing nodes continuously input dynamic human flow data, driving the real-time evolution of the contact network. The upper-level resource monitoring system dynamically adjusts regional association strengths through a threshold trigger mechanism. The cross-layer propagation model precisely quantifies micro-site risks and maps them to macro-propagation paths using attenuation coefficients. This three-layer mechanism achieves data synchronization and decision-making coordination through a unified time window, improving the accuracy of global infection risk index predictions.

[0126] This application further proposes that the regional vulnerability scoring matrix is based on the attention-weighted fusion of the symptom density time series feature vector and the susceptible population distribution feature vector. The attention-weighted fusion includes:

[0127] The TF-IDF algorithm is used to generate keyword weights for symptom keywords;

[0128] Normalize the susceptible population distribution feature vector to the interval value of [0,1] according to the grid unit;

[0129] Generate fusion features through the product formula:

[0130] Regional vulnerability score = ∑(symptom keyword weight × proportion of susceptible population) × time decay factor,

[0131] The time decay factor is calculated according to a preset exponential function.

[0132] The symptom density time series feature vector is a semantic representation extracted from medical records using natural language processing technology. A modified TF-IDF algorithm is used to weight symptom keywords, and a sliding window mechanism is used to generate a time series evolution curve reflecting the intensity of symptom clustering. The susceptible population distribution feature vector is a spatial risk indicator constructed based on census data and real-time mobile trajectories. Specifically, through geographic gridding and kernel density estimation, risk factors such as age structure and the proportion of underlying diseases are normalized into a spatial distribution matrix in the [0,1] interval. Weighted attention fusion establishes a dynamic correlation model between semantic and spatial features. It uses a multiplication operation to capture the nonlinear coupling relationship between symptom clustering and susceptible population distribution, and introduces a time decay factor to eliminate the outdated effects of historical data. The time decay factor is an exponential decay function designed based on the dynamics of epidemic transmission. The time window weight coefficient is dynamically adjusted using the real-time transmission rate parameter output by the SEIR infection model, ensuring that the assessment results are more consistent with the actual evolution of the epidemic.

[0133] Specifically, when the system detects a sudden increase in the frequency of fever symptom keywords in a certain area, the TF-IDF algorithm automatically increases the semantic weight of the corresponding keywords. Simultaneously, it combines the data on the proportion of elderly people within that grid cell and amplifies the local risk signal through a multiplication operation. The propagation path predictions output in real time by the spatiotemporal graph network provide a basis for dynamic adjustment of the time decay factor: if the prediction indicates that the area is at the forefront of virus spread, the decay rate is reduced to prolong the impact of historical data; if it is in the propagation decay phase, the decay process is accelerated to focus on recent data features. The fused regional vulnerability score is spatially smoothed using a convolutional neural network to generate a risk assessment matrix overlaid with a heat map of medical resource distribution, providing a quantitative decision-making basis for tiered prevention and control strategies. In the closed-loop feedback phase, the system spatially and temporally aligns the actual infection case distribution data with the predicted vulnerability score. The backpropagation algorithm dynamically adjusts the TF-IDF weight coefficient and kernel density estimation bandwidth parameter, allowing the fusion model to continuously adapt to changes in feature distribution caused by virus mutations.

[0134] Compared with existing technologies, traditional vulnerability assessment methods mostly use linear weighted fusion strategies, which cannot effectively capture the nonlinear correlation between symptom semantic features and spatial risk factors. This application improves the spatial resolution of regional risk assessment results through the collaborative design of a multiplicative attention mechanism and a dynamic attenuation factor. In existing methods, the time decay parameter is usually fixed, resulting in continuous interference of early epidemic data on the assessment results. This application uses a dynamic attenuation function driven by the SEIR infection model to adjust the time window weight in real time in conjunction with the virus transmission rate, thereby reducing historical data interference errors. Existing assessment models lack the ability to deeply couple semantic-spatial features, while this application controls the deviation rate through an attention weighted fusion mechanism.

[0135] Through the above technical solutions, this application has constructed an intelligent evaluation system that integrates semantics, space, and time. The dynamic product fusion mechanism breaks through the limitations of the traditional linear weighted expression capabilities, and the real-time linkage design of the time attenuation factor and the virus transmission dynamics enables the evaluation model to have adaptive evolution capabilities. The closed-loop feedback system reversely optimizes the feature weight parameters through actual infection data, significantly improving the robustness of the evaluation under complex transmission scenarios. This application improves the prediction accuracy of regional vulnerability scores and provides core data support for precise epidemic prevention and control. In particular, in clustered epidemics caused by mutant strains, it can identify high-risk areas 24-48 hours in advance, greatly improving the effectiveness of prevention and control responses.

[0136] like Figure 4 As shown, this application further proposes that the hierarchical prevention and control strategy in step S4 includes:

[0137] Level 1 response: When the risk index is in the first preset range, flow control measures in key places are triggered;

[0138] Second-level response: When the risk index is in the second preset range, a resource allocation instruction is sent to the associated regions and a cross-regional resource allocation plan is generated;

[0139] Level 3 response: When the risk index is in the third preset range, a community-level lockdown order is generated and an emergency plan is activated;

[0140] Among them, the boundary value of each preset interval is dynamically adjusted according to the real-time medical resource occupancy rate.

[0141] Specifically, the tiered prevention and control strategy intelligently triggers different levels of prevention and control measures based on the preset range of the global infection risk index. This is achieved through the collaborative operation of a dynamic threshold adjustment algorithm and a resource scheduling model, ensuring that the intensity of prevention and control measures is precisely matched to the risk of epidemic transmission. The first-level response automatically activates a venue-level flow control plan when the risk index is in the low-risk range. This plan utilizes a key venue crowd flow prediction model based on the node characteristics of a spatiotemporal graph network to generate dynamic control instructions, including flow restriction ratios, opening hours, and testing frequency. The second-level response triggers a cross-regional resource coordination protocol when the risk index rises to the high-risk range. This plan uses a mixed integer programming algorithm to solve the optimal resource allocation path and, combined with real-time traffic data, generates an emergency plan, including transport vehicle scheduling, medical staff deployment, and negative pressure ward activation instructions. The third-level response implements a community-level lockdown and emergency medical facility deployment plan when the risk index exceeds a critical threshold. This plan uses a convolutional neural network to identify the spatial distribution characteristics of high-risk communities and integrates with the geographic information system to generate a comprehensive prevention and control plan, including lockdown boundaries, nucleic acid screening point layout, and modular hospital construction planning. Among them, dynamic interval boundary adjustment refers to reconstructing the risk level classification standards through fuzzy control algorithms based on real-time parameters such as hospital bed occupancy rate and protective material consumption rate. Specifically, a sliding window mechanism is used to monitor the time series change characteristics of the medical resource status and dynamically correct the trigger thresholds of each response level.

[0142] Specifically, when the real-time medical resource occupancy rate continues to exceed the safety threshold, the prevention and control strategy generation module automatically compresses the range of the first response interval, lowers the secondary response trigger threshold, and optimizes the priority weight parameters in the resource allocation algorithm. The propagation path prediction data continuously output by the spatiotemporal graph network provides a dynamic path planning basis for cross-regional resource scheduling. For example, when it is predicted that a certain transportation hub will become a transmission hotspot, the protective material reserve threshold of the adjacent medical institutions is automatically adjusted, and a targeted reinforcement order is generated. During the execution stage of the community lockdown order, the closed-loop feedback system collects new case data and material consumption data in the lockdown area in real time. By comparing the spatiotemporal overlap between the predicted infection density matrix and the actual case distribution, it dynamically adjusts the lockdown range and nucleic acid screening frequency parameters, and feeds the correction amount back to the edge weight calculation of the spatiotemporal graph network, forming a complete data loop from strategy execution to model optimization.

[0143] Compared with the existing technology, the traditional hierarchical prevention and control scheme relies on fixed thresholds and static response rules, which cannot adapt to the dynamic changes in the spread of the epidemic and the status of medical resources. This application innovatively realizes the flexible adjustment and execution closed loop of the prevention and control strategy by establishing a real-time mapping relationship between risk index-resource status-prevention and control instructions. In the existing technology, the generation of resource scheduling instructions usually lags behind the development of the epidemic by 12-24 hours. This application shortens the response delay through the minute-level prediction capability of the spatiotemporal graph network and the real-time solution optimization of the mixed integer programming algorithm. Existing lockdown decisions mostly rely on manual experience to define boundaries, and there are over-controlled or missed areas. This application uses the community risk feature map extracted by the convolutional neural network to improve the spatial matching between the lockdown area and the transmission hotspot.

[0144] Through the above technical solutions, this application has built an intelligent decision-making chain from risk perception to prevention and control execution. The dynamic threshold adjustment mechanism effectively solves the problems of response lag and resource mismatch in traditional solutions. The hierarchical design of the three-level response strategy achieves precise adaptation of prevention and control efforts and transmission risks. The closed-loop feedback system continuously optimizes the spatiotemporal graph network parameters and resource scheduling models, enabling the entire prevention and control system to have dynamic evolution capabilities. This application improves the decision-making response speed of epidemic prevention and control, improves the accuracy of key area control, and provides technical support for responding to public health emergencies that is both timely and scientific.

[0145] This application further proposes that the measured infection density matrix is obtained in real time through a public health data interface and aligned according to a preset grid division standard; the execution terminal includes a hospital resource scheduling system and a community access control system; and online incremental learning includes:

[0146] Freeze the underlying graph structure parameters of the two-layer spatiotemporal graph network;

[0147] Perform gradient backpropagation to update the edge weight calculation function of the upper graph;

[0148] A sliding window mechanism is used to retain the training data of the most recent preset length.

[0149] This application is based on the infection density monitoring and dynamic learning mechanism of the real-time public health data interface, and realizes the closed-loop optimization of the infection risk assessment system by constructing a standardized grid data stream, a multi-terminal collaborative execution architecture and an incremental model update system. Among them, after the measured infection density matrix is obtained in real time through the public health data interface, the preset grid division standard is used for spatial alignment processing. Specifically, the case report data from different sources are uniformly mapped to the geographic grid through the grid coding mapping algorithm of the geographic information system (GIS), eliminating the data boundary offset problem caused by administrative division differences. The execution terminal is deployed as a distributed architecture, which includes a dual-channel response of the hospital resource scheduling system and the community access control system. The hospital system dynamically adjusts the ICU bed allocation priority and medical staff scheduling strategy based on the infection density matrix, and the community system triggers the hierarchical management of access control permissions through the real-time risk index. The two achieve millisecond-level command synchronization through the central command distribution platform.

[0150] The online incremental learning mechanism adopts a hierarchical parameter update strategy: first, the structural parameters of the bottom-level place contact network in the two-layer spatiotemporal graph network (including the cumulative calculation logic of pedestrian density and movement speed) are frozen to ensure the stability of the basic transmission path characteristics; then, the edge weight calculation function of the upper-level regional association network is gradient back-propagated to update, focusing on optimizing the dynamic association model between medical resource dependence parameters and population migration rate, and converting the spatial overlap error between the measured infection density and the predicted value into a weight correction through the back-propagation algorithm; at the same time, a sliding window mechanism is used to retain the training data of the most recent preset time length (for example, the last 72 hours), and the L-BFGS optimization algorithm is combined to perform rolling fine-tuning on the model, so that the incremental learning process can not only maintain the memory of historical transmission laws, but also quickly adapt to the sudden epidemic situation.

[0151] Compared with existing technologies, traditional infection monitoring systems have positioning deviations in cross-regional case data due to differences in administrative division coding, resulting in insufficient spatial resolution of the infection density matrix; the prevention and control command execution terminals lack a coordination mechanism, and there are response delays in hospital resource scheduling and community management; model updates rely on full data retraining, and each iteration consumes time and computing resources.

[0152] The technical solution of this application provides high-precision input for the model through real-time gridded data streams, dual-channel execution terminals ensure the spatiotemporal consistency of prevention and control strategies, and hierarchical incremental learning achieves dynamic calibration of model parameters. The grid division module quickly matches case locations with preset grids through GIS spatial indexing to generate a standardized infection density heat map. The hospital dispatch system predicts the trend of severe case conversion based on the gradient changes of the heat map and initiates the allocation of ventilator resources in advance. The community access control system implements dynamic access rules based on the grid risk level, and high-risk grids activate iris recognition and body temperature joint inspection mechanisms. During the incremental learning process, the underlying network freezes to ensure the stability of the physical laws of venue-level contact risks, while the upper network dynamically updates to capture the time-varying characteristics of regional correlations. The sliding window mechanism selects key training samples through temporal attention weights, allowing the model to quickly adapt to sudden situations such as virus mutations while retaining long-term transmission patterns. This "data-execution-model" ternary collaborative architecture shortens the response delay of global infection risk prediction and improves the accuracy of major epidemic warnings.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A dynamic infection risk assessment method based on deep learning, characterized by: The following steps are involved: S1. Call the pre-trained deep learning model to acquire multi-source heterogeneous data in real time, perform cross-modal alignment processing, and output the real-time basic reproduction number (R0) R0 value set based on the SEIR infection model; S2. Using the R0 value set as the node initial feature vector of the space-time graph network, constructing a two-layer space-time graph network based on the dynamic flow risk matrix, and generating a regional vulnerability score matrix in combination with the multi-source heterogeneous data; S3. Calculate the global infection risk index based on a linear combination of the regional vulnerability score matrix and the R0 value set; S4. When it is detected that the global infection risk index exceeds a preset infection threshold, a predicted infection density matrix and a hierarchical prevention and control strategy including prevention and control level parameters are generated; S5. Compare the predicted infection density matrix with the measured infection density matrix, dynamically update the weight parameter according to the comparison result, and output it to the execution terminal, wherein: Calculating a spatial overlap index between the predicted infection density matrix and the measured infection density matrix; When the spatial overlap is lower than a preset threshold, the weight of the symptom keyword is increased proportionally; When the consumption rate of medical resources continues to exceed the consumption threshold, the dependency parameter of the associated area is reduced; When the correction direction is consistent, online incremental learning of the model is performed to trigger retraining of the deep learning model.

2. The deep learning-based dynamic infection risk assessment method according to claim 1, characterized in that: The multi-source heterogeneous data is filtered out of abnormal data through pre-screening rules: When the crowd density data exceeds the preset safety threshold, Z-Score normalization is performed on the crowd density data, and outlier data is filtered based on a statistical anomaly detection algorithm; A dynamic sensitive word library is set for symptom keywords, and when the word frequency of the symptom keyword exceeds the baseline value threshold, it is marked as high risk; Set category association constraints for protective material inventory data, and only transmit category data related to respiratory disease prevention and control; The deep learning model is trained and optimized, and incremental model parameters are regularly sent to edge nodes.

3. The method for dynamic infection risk assessment based on deep learning according to claim 1, characterized in that: The cross-modal alignment process includes: Define default filling strategies based on preset protocols and design timing and spatial features; Extracting the symptom keywords from the medical institution's case report database and medical record data according to the preset protocol, performing sliding window normalization to generate a symptom density time series feature vector; According to the preset protocol, the population census database is obtained and One-Hot encoding is performed to generate a susceptible population distribution feature vector.

4. The method for dynamic infection risk assessment based on deep learning according to claim 1, characterized in that: The generation of the dynamic liquidity risk matrix in step S2 includes: Divide mobile phone signaling data into time-space grids with a predetermined number of digits and calculate the pedestrian density of each grid unit; Statistics of traffic density change rate according to preset time window; The current risk value is generated by the weighted formula: Flow risk value = pedestrian density × moving speed + vehicle density change rate × road weight coefficient, where the road weight coefficient is preset according to the road grade.

5. The method for dynamic infection risk assessment based on deep learning according to claim 3, characterized in that: The two-layer spatiotemporal graph network is divided into a bottom layer graph and an upper layer graph. The bottom layer graph represents the place contact network, and the edge weights are calculated by accumulating the crowd density data and movement speed; the upper layer graph represents the regional association network, and the edge weights are calculated by population migration rate and medical resource dependence.

6. The method for dynamic infection risk assessment based on deep learning according to claim 5, characterized in that: The calculation of the medical resource dependence includes: Obtain historical service coverage data of the medical institution; Count the frequency of medical visits in each area within the preset period; The medical resource dependency is generated by a dynamic formula: Dependence = frequency of visits × (1-current resource shortage), The resource tension is calculated based on the ratio of the inventory consumption rate to the safety threshold.

7. The deep learning-based dynamic infection risk assessment method according to claim 6, characterized in that: The construction of the two-layer spatiotemporal graph network includes: The underlying graph edge weights implement a dynamic update mechanism that periodically recalculates the exposure risk based on the crowd density; The upper-level graph edge weight executes a constraint condition of automatically reducing the population migration rate weight of the associated area when the resource tension exceeds a preset resource threshold; The bottom-layer contact risk is mapped to the upper-layer propagation path through the preset propagation attenuation coefficient for cross-layer connection.

8. The method for dynamic infection risk assessment based on deep learning according to claim 3, characterized in that: The regional vulnerability scoring matrix is subjected to attention weighted fusion based on the symptom density time series feature vector and the susceptible population distribution feature vector, and the attention weighted fusion includes: Generating the keyword weights using the TF-IDF algorithm for the symptom keywords; Normalizing the susceptible population distribution feature vector to a [0,1] interval value according to the grid unit; Generate fusion features through the product formula: Regional vulnerability score = ∑ (symptom keyword weight × proportion of susceptible population) × time decay factor, where the time decay factor is calculated according to a preset exponential function.

9. The method for dynamic infection risk assessment based on deep learning according to claim 1, characterized in that: The hierarchical prevention and control strategy in step S4 includes: Level 1 response: When the risk index is in the first preset range, flow control measures in key places are triggered; Second-level response: when the risk index is within the second preset range, sending a resource allocation instruction to the associated region and generating a cross-regional resource allocation plan; Level 3 response: When the risk index is in the third preset range, a community-level lockdown order is generated and an emergency plan is activated; Among them, the boundary value of each preset interval is dynamically adjusted according to the real-time medical resource occupancy rate.

10. The method for dynamic infection risk assessment based on deep learning according to claim 1, characterized in that: The measured infection density matrix is obtained in real time through the public health data interface and aligned according to a preset grid division standard; The execution terminal includes a hospital resource scheduling system and a community access control system; the online incremental learning includes: Freezing the underlying graph structure parameters of the dual-layer spatiotemporal graph network; Performing gradient backpropagation update on the edge weight calculation function of the upper graph; A sliding window mechanism is used to retain the training data of the most recent preset length.

Images