An intelligent epidemic prediction system based on deep learning

Abstract

The application relates to the technical field of epidemic prediction, and discloses an intelligent epidemic prediction system based on deep learning. A multivariate data acquisition module of the system acquires real-time environmental meteorological data, population flow data, medical resource distribution data and historical epidemic transmission data related to an epidemic; a deep learning prediction module analyzes regional transmission dynamic characteristics through a space-time graph neural network according to the above-mentioned multi-source data; a risk area identification module constructs a regional risk level heat map according to the regional transmission dynamic characteristics, and identifies a high-risk area with a risk level exceeding a preset threshold; and a resource matching analysis module calculates the resource coverage range of each medical institution in combination with the medical resource distribution data and the spatial position of the high-risk area. The system can integrate multi-source data, accurately analyze the epidemic transmission trend, intelligently identify the high-risk area and realize reasonable matching of medical resources, and provide effective technical support for epidemic prevention and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of epidemic prediction, in particular to an intelligent epidemic prediction system based on deep learning. BACKGROUND

[0002] In the global public health security system, the outbreak and spread of epidemics will have a great impact on social order, economic development and public health. Timely and accurate prediction of epidemic transmission trends, identification of high-risk areas and rational allocation of medical resources are important links in prevention and control. Traditional epidemic prediction methods rely mainly on epidemiological survey data and simple statistical models, which have obvious limitations.

[0003] From the perspective of data utilization, traditional methods often only focus on historical epidemic transmission data, ignoring the impact of environmental and meteorological data, population flow data and medical resource distribution data on epidemic transmission. In fact, environmental and meteorological factors such as temperature, humidity and precipitation directly affect the survival, reproduction and activity of the transmission medium of the pathogen. For example, some viruses are more likely to survive and spread in low-temperature and humid environments. Population flow data reflects the cross-regional movement of personnel and is a key driving factor for the spread of epidemics across regions. In particular, during periods of large-scale population movement such as holidays, the risk of epidemic transmission will increase significantly. Medical resource distribution data is closely related to epidemic prevention and control capabilities. The number, type and distribution density of medical resources in a region will affect the speed and effectiveness of epidemic control.

[0004] In terms of prediction models, traditional statistical models are difficult to capture the spatio-temporal dynamic characteristics of epidemic transmission. Epidemic transmission has obvious spatio-temporal correlation, and the transmission situation of different regions at different time points is different, and the transmission between regions is mutually influenced. Traditional models cannot effectively integrate multi-source heterogeneous data, and it is also difficult to accurately describe the complex spatio-temporal dynamic relationship, resulting in insufficient accuracy and timeliness of the prediction results, and often failing to provide timely and effective reference for prevention and control decisions.

[0005] In the aspects of risk area identification and resource allocation, the traditional method mainly adopts manual analysis and experience judgment, lacking systematic and intelligent technical support. Manual analysis is not only inefficient, but also difficult to dynamically adjust the risk level division according to the real-time changing epidemic data, which is prone to the problem of lag in identifying high-risk areas. Meanwhile, in the process of matching medical resources, due to the inability to quickly and accurately calculate the resource coverage range of each medical institution, it often leads to the over-concentration of medical resources in some areas, while the shortage of resources in high-risk areas, affecting the orderly development of epidemic prevention and control work. With the frequent occurrence of epidemics in the global range in recent years, the traditional method has been difficult to meet the actual needs of epidemic prevention and control under the current complex situation, and an urgent need for a technical solution that can integrate multi-source data, has accurate prediction ability, can intelligently identify risk areas and realize reasonable matching of resources. SUMMARY

[0006] The purpose of the present application is to provide an intelligent epidemic prediction system based on deep learning to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides an intelligent epidemic prediction system based on deep learning, which comprises:

[0008] A multi-element data acquisition module is used to acquire real-time environmental meteorological data, population flow data, medical resource distribution data and historical epidemic transmission data related to the epidemic;

[0009] A deep learning prediction module is used to analyze the regional transmission dynamic characteristics through the collaborative architecture of spatio-temporal graph neural network and long short-term memory network according to the environmental meteorological data, population flow data, medical resource distribution data and historical epidemic transmission data;

[0010] A risk area identification module is used to construct a regional risk level heat map according to the regional transmission dynamic characteristics, and identify high-risk areas with risk levels exceeding a preset threshold. The risk area identification module is based on the output embedding vector of the deep learning prediction module to construct the regional risk level heat map, which includes: inputting the embedding vector into a classification model, which divides the risk level into three categories of low, medium and high by using support vector machine; the heat map is drawn by geographic information system, and each grid cell corresponds to a risk probability; the preset threshold is set to high risk above the median value, the system automatically identifies the high-risk areas with risk levels exceeding the threshold, and the output is raster data;

[0011] A resource matching analysis module is used to calculate the resource coverage range of each medical institution according to the spatial position of the medical resource distribution data and the high-risk area;

[0012] When the resource matching analysis module calculates the resource coverage range of each medical institution, the following operations are specifically performed:

[0013] In the first step, real-time resource capacity parameters of each medical institution are obtained, including the number of on-duty medical staff, the number of available beds, and the reserve amount of emergency equipment. The above parameters are standardized into a resource sufficiency index in the range of 0-1.

[0014] In the second step, a coverage radius basic model is established by combining population density data and traffic accessibility data of high-risk areas.

[0015] In the third step, the coverage radius is dynamically adjusted according to the resource sufficiency index: when the resource sufficiency index is ≥0.8, the coverage radius is expanded by 20% based on the basic model result; when 0.5≤resource sufficiency index<0.8, the coverage radius of the basic model is maintained; when the resource sufficiency index is <0.5, the coverage radius is reduced by 30% based on the basic model result.

[0016] In the fourth step, the adjusted coverage radius is checked for spatial boundaries to exclude natural geographical barrier areas such as rivers and mountains, and traffic control areas. Finally, the actual resource coverage range of each medical institution is determined, and the output is polygon vector data with geographic coordinates.

[0017] Preferably, the resource matching analysis module comprises:

[0018] The resource coverage range of each medical institution is spatially overlaid and analyzed with the high-risk area to generate the overlapping area and the position of the overlapping area between the resource coverage range and the high-risk area.

[0019] According to the position data of the medical institution and the position of the overlapping area, the separation distance between the medical institution and the overlapping area is calculated.

[0020] The prevention and control potential coefficient of each medical institution is calculated by combining the overlapping area, the separation distance, and the real-time load data of medical resources.

[0021] Preferably, the deep learning prediction module performs:

[0022] The temperature and humidity parameters in the environmental meteorological data, the cross-regional migration intensity in the population flow data, and the bed turnover rate in the medical resource distribution data are taken as input features.

[0023] The spatial correlation of multi-source data is processed by a spatio-temporal graph neural network to provide accurate spatial feature basis for a long short-term memory network.

[0024] The spatio-temporal propagation feature vector is extracted by the long short-term memory network.

[0025] The propagation rate and the mutation risk probability in a future specified time window are predicted based on the spatio-temporal propagation feature vector.

[0026] Preferably, the system further comprises:

[0027] The prevention and control strategy generation module is configured to fuse the prevention and control potential coefficients of the medical institutions with the transmission rates and the mutation risk probabilities;

[0028] The multi-layer perception network outputs a prevention and control matching degree score of each region;

[0029] The prevention and control resource scheduling priority list is generated according to the prevention and control matching degree score.

[0030] Preferably, the system further comprises:

[0031] The dynamic decision module is configured to determine the prevention and control resource scheduling priority list and a real-time epidemic severity index;

[0032] When the real-time epidemic severity index exceeds a dynamically adjusted threshold, a hierarchical response mechanism is activated;

[0033] Based on a matching result of the prevention and control matching degree score and a preset response rule, a prevention and control parameter adjustment instruction is output.

[0034] Preferably, the prevention and control parameter adjustment instruction comprises:

[0035] A detection point density parameter is adjusted according to a high-risk area area change rate;

[0036] An isolation range parameter is adjusted according to a ratio of the transmission rate to the medical resource load;

[0037] A vaccine allocation weight parameter is adjusted according to the mutation risk probability.

[0038] Preferably, the system further comprises:

[0039] The execution feedback module is configured to collect actual transmission attenuation rates, resource usage deviation values, and new case distribution data after implementation of the prevention and control measures;

[0040] A first feedback coefficient is generated by difference comparison between the actual transmission attenuation rates and the predicted transmission rates;

[0041] A second feedback coefficient is generated by correlation analysis between the resource usage deviation values and the prevention and control matching degree scores.

[0042] Preferably, the system further comprises:

[0043] The model optimization module is configured to perform spatial error calculation on the new case distribution data and the predicted high-risk areas;

[0044] A loss function is constructed by fusing the first feedback coefficient and the second feedback coefficient;

[0045] The weight parameters of the spatio-temporal graph neural network are dynamically updated by a back propagation algorithm.

[0046] Preferably, the system further comprises:

[0047] An iterative early warning module for regenerating regional propagation dynamic features according to the updated deep learning prediction module;

[0048] Triggering a cross-regional collaborative early warning protocol when the regenerated mutation risk probability exceeds the historical peak value;

[0049] Updating the prevention and control resource scheduling priority list based on the recalculated prevention and control matching degree score.

[0050] Preferably, according to the deep learning-based intelligent epidemic prediction system described above, the system is deployed on a distributed computing platform, and a spatio-temporal database for storing the output of the multi-source data acquisition module is included.

[0051] The deep learning prediction module accelerates the matrix operation of the spatio-temporal graph neural network through a graphics processing unit.

[0052] The output instructions of the dynamic decision module are synchronized to the public health response terminal through an Internet of Things interface.

[0053] Compared with the prior art, the beneficial effects of the present application are:

[0054] The multi-source data acquisition module can obtain real-time epidemic-related environmental meteorological data, population flow data, medical resource distribution data, and historical epidemic transmission data, breaking the limitation of single data source in traditional prediction methods. This module realizes the comprehensive collection and real-time update of multi-source heterogeneous data, so that the system can fully consider various key factors affecting the spread of epidemics when performing prediction analysis. The introduction of environmental meteorological data allows the system to accurately grasp the environmental conditions for the survival and spread of pathogens, thereby more accurately judging the transmission risk in different environments; real-time acquisition of population flow data can timely reflect the trend of personnel flow, providing a data basis for capturing the dynamic of regional epidemic transmission; the inclusion of medical resource distribution data enables the system to have data support in the subsequent resource matching link, avoiding unreasonable resource allocation due to information loss; and historical epidemic transmission data provides a basis for model training and transmission rule summary, helping to improve the reliability of prediction.

[0055] The deep learning prediction module uses a spatio-temporal graph neural network to analyze regional transmission dynamics, which has a significant advantage over traditional statistical models in handling the spatio-temporal correlation of epidemic transmission. The spatio-temporal graph neural network can effectively integrate multi-source data, accurately depict the transmission situation of different regions at different time points and the mutual influence between regions by constructing a spatio-temporal correlation model. This model architecture can fully exploit the hidden information in the data and capture the dynamic changes in the process of epidemic transmission, thereby improving the accuracy and timeliness of the transmission trend prediction. Through the analysis of this module, the transmission range, transmission speed, and the number of infected people in the future can be predicted, providing scientific guidance for the prevention and control departments to develop prevention and control strategies and deploy prevention and control forces in advance, which helps to take timely intervention measures at the initial stage of the epidemic spread, delay or block the spread of the epidemic.

[0056] The risk area identification module constructs a regional risk level heat map based on the regional transmission dynamics and identifies high-risk areas, achieving the intelligent and visual identification of risk areas. The heat map can intuitively present the risk level distribution of different regions, making it easy for prevention and control personnel to quickly grasp the overall epidemic risk situation. At the same time, this module assesses the risk level based on real-time updated transmission dynamics, which can timely identify areas with rising risk levels and avoid the problem of lagging behind in identifying high-risk areas. By accurately identifying high-risk areas, prevention and control resources can be concentrated in the highest risk areas, improving the targeting and efficiency of prevention and control work, reducing unnecessary resource waste, and maximizing the use of limited prevention and control forces.

[0057] The resource matching analysis module calculates the resource coverage range of each medical institution according to the medical resource distribution data and the spatial position of the high-risk area, and provides intelligent support for the reasonable allocation of medical resources. When calculating the resource coverage range, the resource matching analysis module is not simply based on geographical distance division, but is dynamically adjusted through multi-dimensional parameters: first, the real-time resource capacity of the medical institution (such as on-duty medical staff, available beds, and emergency equipment) is collected and converted into a resource sufficiency index, and the population density and traffic accessibility of the high-risk area are combined to determine the basic coverage radius; then, the basic radius is expanded or reduced according to the resource sufficiency index, for example, the resource-sufficient institution can expand the service range, and the resource-stressed institution can reduce the range; finally, the natural obstacles and traffic control areas are excluded to ensure that the calculated coverage range meets the actual service capacity and geographical accessibility, avoiding the resource coverage evaluation deviation caused by single geographical distance calculation. This module can quickly and accurately analyze the range of areas covered by each medical institution and the supply of medical resources around different high-risk areas. Based on such analysis results, the prevention and control department can clearly understand which high-risk areas have sufficient medical resources around them and which areas have resource gaps, so as to develop resource allocation schemes accordingly and allocate medical resources from the sufficient areas to the gap areas. This resource matching method avoids the subjectivity and lag of manual allocation, ensures that medical resources can be quickly and accurately allocated to the areas that need them most, ensures that patients in high-risk areas can obtain medical treatment in a timely manner, and also avoids the waste caused by the overaccumulation of medical resources in some areas, improves the overall medical resource utilization efficiency, and provides strong support for medical treatment in epidemic prevention.

[0058] The modules of the whole system cooperate with each other to form a complete technical chain from data collection, trend prediction, risk identification to resource matching. The multi-dimensional data collection module provides comprehensive and real-time data support for the subsequent modules, the analysis results of the deep learning prediction module provide the basis for risk area identification, and the risk area identification results guide the resource matching analysis module to carry out accurate resource coverage range calculation. The modules cooperate with each other, so that the whole system has efficient and accurate epidemic prevention and control auxiliary ability. In practical application, the system can be widely applied to CDCs at all levels, health departments and other institutions to provide comprehensive technical support for epidemic prevention and control decision-making, help to improve the overall epidemic prevention and control level, and better protect public health and safety and the health of the public. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The timing diagram of the deep learning-based intelligent epidemic prediction system described in the present application;

[0060] Figure 2 The work flow chart of the resource matching analysis module;

[0061] Figure 3 Workflow diagram for the prevention and control strategy generation module;

[0062] Figure 4 Workflow diagram for the dynamic decision module. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0064] Please refer to Figure 1 The present application provides an intelligent epidemic prediction system based on deep learning, which includes a multi-element data acquisition module, a deep learning prediction module, a risk area identification module and a resource matching analysis module.

[0065] The multi-element data acquisition module performs real-time acquisition of environmental meteorological data related to the epidemic, population flow data, medical resource distribution data and historical epidemic transmission data. These data are integrated through a heterogeneous data interface. The environmental meteorological data is sourced from the real-time sensor network of the national meteorological center, including temperature, humidity, wind speed and other parameters. The population flow data is obtained through the mobile operator signaling platform and the traffic monitoring system, recording the cross-regional migration intensity. The medical resource distribution data is extracted from the health management department database, covering hospital bed number, equipment distribution and personnel establishment. The historical epidemic transmission data is downloaded from the historical record library of the disease control agency, including past case distribution and transmission cycle. The multi-element data acquisition module realizes data cleaning and standardization, adopts time stamp alignment technology to ensure time sequence consistency, and outputs a structured data set to the deep learning prediction module.

[0066] The deep learning prediction module receives the output of the multi-element data acquisition module, and analyzes the regional transmission dynamic characteristics through a spatio-temporal graph neural network. The spatio-temporal graph neural network architecture is composed of an input layer, a graph convolution layer and an output layer, wherein the graph convolution layer constructs a spatial adjacency matrix based on geographical zoning to process the interaction between regions. The input features include multi-dimensional vectors of environmental meteorological data, population flow data, medical resource distribution data and historical epidemic transmission data. The graph convolution layer applies an attention mechanism to dynamically weight the influence of neighbor nodes, and the output layer generates an embedding vector representing the regional transmission dynamic characteristics. The deep learning prediction module is deployed on a distributed cluster to realize efficient computation through batch processing.

[0067] The risk area identification module constructs a risk level heat map based on the output embedding vector of the deep learning prediction module. The process includes: inputting the embedding vector into a classification model, which divides the risk level into three categories of low, medium and high using a support vector machine; the heat map is drawn by a geographic information system, and each grid cell corresponds to a risk probability; the preset threshold is set to high risk above the median value, and the system automatically identifies the high risk area whose risk level exceeds the threshold, and outputs raster data.

[0068] The resource matching analysis module calculates the resource coverage of each medical institution using medical resource distribution data and the spatial location of high-risk areas. The coverage is defined as the maximum area radius that can be served within the buffer zone of the medical institution, and the buffer zone radius is dynamically adjusted based on the medical resource capacity formula, including bed density and personnel density. This module executes a spatial analysis algorithm to calculate the coverage radius in combination with regional population density and traffic network, and outputs polygon boundary data. The following further details the extended implementation of the system through multiple embodiments.

[0069] Embodiment 1: refer to Figure 2 The execution process of the resource matching analysis module is based on medical resource distribution data and spatial location data of high-risk areas. The medical resource distribution data is stored in vector polygon format, representing the resource coverage of each medical institution, where the coverage is defined as the continuous geographical area within a certain service radius that the institution can effectively provide medical support. The high-risk area data is derived from the raster heat map output by the risk area identification module, which is converted into a polygon layer with clear boundaries through binary processing. The system starts the spatial overlay analysis program and calls the spatial operation interface of the geographic information system engine. The program loads the medical institution coverage layer and high-risk area layer into the memory workspace and applies spatial indexing technology to accelerate data processing. The overlay analysis performs polygon intersection operation to judge the positional relationship of all geometric objects in the two layers. The calculation process iterates through the high-risk area set for each institution, calculates the geometric intersection area of the coverage polygon and the high-risk area polygon. The intersection area value is calculated in real time by analyzing the polygon vertex coordinates and using numerical integration algorithm. At the same time, the geographic identification code of each high-risk area that produces intersection is recorded, and the center point coordinates of its bounding rectangle are extracted. The calculation process dynamically generates the result data set, which includes the fields of medical institution ID, high-risk area ID, intersection area and intersection area center point coordinates.

[0070] After obtaining the intersection area, the system calculates the interval distance between the medical institution and the corresponding high-risk area. The reference point position of the medical institution is extracted from the medical resource distribution database, and the latitude and longitude coordinates are usually set as the positioning point of the main building of the institution. The interval distance is calculated using the geodetic formula, based on the spatial reference coordinate system conversion, the medical institution coordinate point and the intersection center point of the high-risk area are calculated by geodesic distance. The calculation engine automatically processes the projection conversion problem of the coordinate system to ensure the accuracy of the earth surface distance. For traffic optimization, the system accesses the real-time road network database, and when it detects that there is special terrain or traffic control between the high-risk area and the medical institution, it calls the path planning algorithm to output the actual travel distance instead of the straight-line distance. The operation result is appended to the result dataset in the form of distance value.

[0071] After the spatial overlay analysis and distance calculation are completed, the system integrates real-time load data of medical resources. Real-time load data is updated every ten minutes through an application programming interface, including dynamic indicators such as bed remaining capacity, ventilator usage rate, and the number of medical staff on duty. The load data processing module normalizes the original indicators to generate load pressure values ranging from 0 to 1. The synthesis process of the prevention and control potential coefficient adopts a multi-dimensional decision model, combining the intersection area, interval distance, and load pressure value as three core parameters. The model architecture sets the intersection area as a positive factor, the interval distance as a negative factor, and the load pressure value as an adjustment factor. The data processing process first normalizes the area value to a unified dimension, and the distance value is inverted to conform to the distance decay law. The three groups of parameters are combined to calculate the intermediate value according to the preset weight, and the weight is dynamically calibrated through historical event verification. The coefficient value calculation adopts scalar operation rules, and the result is rounded to three decimal places. An outlier detection mechanism is implemented during the calculation process, and when a parameter exceeds the reasonable threshold, the data review program is triggered. The final output data table contains institution ID, potential coefficient value, and timestamp label, which is written into the central database for subsequent module calling.

[0072] The technical implementation involved in the data flow includes: the geometric operation engine using an R-tree index structure to accelerate polygon queries; distance calculation employing a multi-threaded parallel processing mechanism; and load data accessing a message queue to ensure real-time performance. Spatial overlay analysis is performed in blocks based on administrative divisions to avoid bottlenecks in processing ultra-large-scale datasets. Each medical institution's overlay analysis task is encapsulated as an independent computing unit, with computing nodes allocated by a distributed task scheduler. The interval distance calculation module has a built-in caching mechanism, performing local storage reuse for high-risk areas that are repeatedly calculated. The entire process triggers a full calculation every sixty minutes, but incremental updates are immediately initiated when more than 10% of the boundary changes in a high-risk area are detected. Calculation results are output in a standardized JSON format, preserving complete spatial topology relationships and parameter metadata. The database table structure uses spatial fields to store geometric data, supporting dynamic geographic queries and visualization. Version information is recorded for all intermediate data during processing to meet audit and backtracking requirements. The system implements an automatic retry mechanism for calculation failures; consecutive failures trigger manual intervention.

[0073] Example 2: See Figure 3 The data processing flow of the deep learning prediction module begins with the input feature extraction stage. Temperature and humidity parameters from environmental meteorological data are input as floating-point numbers, and the source data is updated every sixty minutes. The cross-regional migration intensity index from population flow data originates from mobile device location signaling; this index is generated through base station handover frequency statistics, and the value represents the proportion of migrating population per unit time. The bed turnover rate from medical resource distribution data comes from the hospital information system and is calculated as the ratio of daily discharged patients to available beds. After being received by the input interface, the three feature parameters undergo standardization processing. The processing rule uses the historical mean and standard deviation of each parameter for z-score transformation. The transformed feature tensor forms a fixed-dimensional matrix structure, with rows corresponding to the time step and columns corresponding to the feature categories. The time step is configured as a seven-day period.

[0074] The matrix is ​​input into a Long Short-Term Memory (LSTM) network for feature extraction. The network structure consists of two hidden layers, each containing 128 neurons. The network operation employs a gating mechanism to handle temporal dependencies: a forget gate controls the retention rate of historical information; an input gate filters important features at the current time step; and an output gate modulates the output strength of the feature vector. The connections between network units utilize recurrent computation logic, where the hidden state from the previous time step and the current input jointly participate in the gating decision at the current time step. Through iterative computation over multiple time steps, the network captures the dynamic changes in regional propagation from time-series data. The hidden layer output is passed to a fully connected layer to compress the dimension, generating a fixed-length spatiotemporal propagation feature vector. The dimension of the feature vector is consistent with the number of geographical partitions, and each element of the vector corresponds to the dynamic state code of a specific region.

[0075] Based on the spatiotemporal propagation feature vector, the module performs two parallel prediction tasks. The propagation rate prediction branch outputs a continuous value quantifying the intensity of epidemic spread per unit time within a seven-day window using a linear regression layer. The variant risk probability prediction branch applies a sigmoid activation function to transform the output value, mapping the regression result to a probability estimate between 0 and 1. The time window is fixed at a seven-day period, and the prediction engine performs a full calculation every twenty-four hours. The prediction results are stored by region ID, and the output data structure includes the predicted value, timestamp, and confidence interval.

[0076] The input of the prevention and control strategy generation module loads three data sources: the propagation rate and variant risk probability data table output by the deep learning prediction module, and the prevention and control potential coefficient list generated by the resource matching analysis module. The data fusion process starts the feature alignment mechanism, matching related parameters by region ID. The aligned parameter group performs normalization scaling to eliminate dimensional differences and form a joint feature vector. This vector is input into a multi-layer perceptron network for pattern analysis, which includes three fully connected layers. The input layer receives a five-dimensional vector (region propagation rate, variant risk probability, and prevention and control potential coefficients associated with three medical institutions); the first hidden layer has 64 neurons, with ReLU as the activation function; the second hidden layer has 32 neurons, also using ReLU activation; the output layer is designed as a single neuron structure to handle score mapping. The weight parameters of each layer are initialized through historical operation data. The prevention and control matching degree score calculation follows the following mapping relationship: ;

[0077] In the formula: represents the prevention and control matching degree score, is the output layer weight parameter vector, is the joint feature vector, represents the sigmoid function. The output score is subjected to threshold truncation processing, and values exceeding the 0-1 range are forcibly reduced to the boundary value. The score result is stored in the cache database.

[0078] The logic for generating the prevention and control resource scheduling priority list is based on the matching degree score ranking. The sorting algorithm uses a maximum heap structure to implement a priority queue, and the queue nodes contain region IDs and matching degree score values. During heap sorting, the score difference between new input data and the top element of the heap is dynamically compared, and when the new score is higher than the top element, the structure is adjusted. The final output list entries are arranged in descending order of score, and each entry is attached with a time validity flag. The list update period is synchronized with the prediction module every twenty-four hours, but when a high-risk region is added, the calculation is immediately recalculated. The queue structure uses persistent storage design, and each update record operation is logged for reference. The data output format is compatible with the public health database, and the fields include region code, medical resource type weight allocation suggestion, and scheduling emergency level label.

[0079] The system is implemented at the container level. The long short-term memory network model publishes API interfaces through the TensorFlow Serving framework; the multi-layer perception model is loaded into the in-memory database to accelerate real-time inference. Apache Kafka message queue is used for data transmission between modules, and a separate control channel is set up to transmit score update instructions. The cache mechanism uses Redis database to temporarily store intermediate feature vectors, effectively reducing the frequency of model recalculation. The distributed task scheduler monitors the load of the computing nodes, and automatically transfers abnormal timeout tasks to the standby nodes. The log system records the control relationship between the predicted value and the actual case number, but does not involve the evaluation conclusion. The model version control integrates the Git repository management, and each update submits a network weight snapshot. The hardware layer connects the GPU accelerator through the PCIe channel to optimize the efficiency of matrix multiplication operations.

[0080] Embodiment 3: see Figure 4 The operation mechanism of the dynamic decision module starts with double data source input: prevention and control resource scheduling priority list and real-time epidemic severity index. The priority list is stored in the form of a sorted queue, containing the region identifier and the corresponding prevention and control matching score value. The real-time epidemic severity index is obtained through the multi-element data acquisition module, and the calculation logic integrates the current confirmed case growth rate, the proportion of severe cases, and the regional transmission coefficient in three dimensions. The index calculation uses a weighted summation model, and the weight distribution is based on the importance of epidemiological parameters. The calculation result is normalized to a floating point number between 0 and 1, and the data snapshot is updated every thirty minutes.

[0081] The core logic of the module includes a threshold detection mechanism. The system presets a dynamically adjusted threshold in the configuration library, which is set to different values according to historical epidemic stage classification. The detection program continuously compares the real-time epidemic severity index with the current effective threshold, and triggers state transition when the index exceeds the threshold for two periods. State transition activates the hierarchical response mechanism, which defines three response modes: primary, intermediate, and advanced. Mode switching is based on the index threshold amplitude decision: within 10% of the threshold, the primary mode is activated; 10%-25% activates the intermediate mode; and 25% or more activates the advanced mode. The mode selection result is written to the system state register.

[0082] After determining the response mode, the module performs matching analysis of the prevention and control matching score and the preset response rule. The preset response rule is stored in the rule table of the relational database, and each rule contains four fields: the response mode field defines the rule effective range; the score interval field defines the matching degree score effective threshold; the response action field encodes specific operation instructions; and the priority field solves rule conflicts. The rule engine loads the rule subset corresponding to the current response mode, and scans the matching degree score interval one by one. The matching process uses interval tree data structure to accelerate query, and generates rule hit markers for each region.

[0083] The generation of the prevention and control parameter adjustment instruction is based on the rule matching result. The instruction generator analyzes the response action field of the hit rule and dynamically synthesizes the executable control command. The instruction set contains three types of core parameter adjustment:

[0084] The detection point density parameter adjustment is based on the area change rate of the high-risk area. The area change rate is calculated by comparing the current and previous period high-risk area heat map, and the percentage change value is obtained by using the grid difference statistical method. The adjustment logic follows a linear response model: ;

[0085] In the formula: is the adjusted detection point density (unit: point / square kilometer), is the basic density configuration value, is the area change rate percentage, is the density adjustment coefficient (default value 100). The calculation result is rounded to generate the density parameter update instruction.

[0086] The isolation range parameter adjustment depends on the ratio of the propagation rate and the medical resource load. The propagation rate is obtained from the deep learning prediction module, and the medical resource load is the real-time bed occupancy rate. The ratio is calculated as the propagation rate divided by the load rate, and the result is input into a segment function converter. The converter presets the ratio threshold: below 1.0, the isolation radius maintains the baseline value; in the interval 1.0-1.5, the radius increases linearly; and above 1.5, the maximum isolation radius is enabled. The output instruction includes the geographic fence coordinate update data packet.

[0087] The vaccine allocation weight parameter adjustment is based on the mutation risk probability. The probability value is mapped to a weight allocation table, which defines the correspondence between probability intervals and regional priority. Probability values below 0.3 are allocated standard weights; in the interval 0.3-0.6, the weight is increased by 50%; and above 0.6, the weight is doubled. The adjuster outputs a list of vaccine allocation weight coefficients for each region, with a precision of two decimal places.

[0088] The instruction transmission mechanism uses a lightweight message protocol encapsulation. Each instruction contains a timestamp, region code, parameter type, and new parameter value four-tuple structure. The Internet of Things interface establishes a dedicated message topic channel and transmits instruction packets by region. The transmission process implements a message acknowledgement and retransmission mechanism, and the public health response terminal returns an execution status code after receiving. The system sets up an instruction life cycle manager, which implements three re-pushes for unconfirmed instructions, and marks an abnormal state if there is no response after timeout.

[0089] The module fault-tolerant design contains a rule conflict detection algorithm. When multiple rules match the same area, the algorithm selects the highest level instruction according to the rule priority field; when the priority is the same, the rule with the latest update time is used. The state machine monitors the module running stage, and the abnormal event triggers the rollback operation: when the instruction generation fails, the last valid configuration is restored; when the network is interrupted, the instruction is cached to the local queue. The historical instruction version is stored in the blockchain storage library, supporting parameter adjustment traceability audit. The real-time monitoring panel visualizes the current effective parameters and area coverage state, assisting artificial decision intervention.

[0090] The system integration layer connects upstream and downstream modules through a service bus. Priority list change events trigger subscription updates, and epidemic index exceeding triggers asynchronous processing threads. The hardware layer uses FPGA to accelerate rule matching operations, and rule table hot loading supports dynamic policy adjustment. The log system records the complete decision link: each state change from threshold detection, rule matching to instruction generation generates a timestamped log item.

[0091] Example 4: Execute the operation flow of the feedback module, taking a certain city administrative division as an example. The module starts the data collection task, and retrieves three core indicators after the implementation of prevention and control measures from the public health monitoring system. The actual propagation attenuation rate field is derived from the daily case statistics report, and the calculation logic is the percentage change value of the number of new cases in the current week compared with the last week. The resource usage bias value field is extracted from the hospital resource management system, and the calculation method is the difference between the planned allocation of ventilators and the actual number of ventilators received. The new case distribution data is obtained from the time and space database of the municipal disease control center, including the latitude and longitude coordinates and diagnosis timestamp of each confirmed patient. The data collection period is fixed at zero-time batch processing every day, covering the full amount of records in the past 24 hours. The collection result is stored in the feedback data table, and the typical data structure is shown in Table 1.

[0092] Table 1: Data collection table of the feedback module

[0093]

[0094] The data processing stage executes parallel branches:

[0095] The first feedback coefficient generation branch loads the actual propagation attenuation rate data, and at the same time extracts the predicted propagation rate data of the same time window from the deep learning prediction module. After the system aligns the region identifier and time label, the numerical difference between the two is calculated. For example, the attenuation rate of a certain region is -15% (case decline), and the predicted propagation rate is 0.2 (diffusion trend), and the difference value is -0.35. This value is converted to the [-1, 1] interval by a normalization function, and the output is the first feedback coefficient. The conversion function has built-in boundary control logic, which triggers the data review process when the difference exceeds the historical maximum fluctuation range.

[0096] The second feedback coefficient generation branch obtains the resource use deviation value and the prevention and control matching degree score. The data matching is in units of regions, and the deviation value sequence and the score sequence are aligned on the time axis. The system applies a statistical correlation algorithm to calculate the correlation degree of the change of the resource use deviation value with the prevention and control matching degree score. The algorithm outputs the Pearson correlation coefficient, whose value range is [-1, 1], which is directly used as the output value of the second feedback coefficient. The calculation process excludes invalid samples, such as deviation data of periods when the matching degree score is not updated.

[0097] The model optimization module simultaneously starts spatial error analysis. The newly added case distribution data is processed by the geographic information engine: first, the case point data is converted into a density grid layer, with a resolution of 100-meter grid; then the historical high-risk area prediction polygon generated by the risk area identification module is loaded. The system performs spatial overlay analysis, using a grid-by-grid comparison strategy: when the case high-density grid falls within the predicted high-risk area, it is recorded as a successful match, and when it falls outside the area, it is recorded as a prediction deviation. The error quantification calculation includes three indicators: prediction omission rate (the proportion of case grids outside the predicted high-risk area), prediction overcoverage rate (the proportion of the area of the predicted area without cases to the total area of the predicted high-risk area), and regional overlap accuracy (the overlapping area ratio of the predicted high-risk area and the case area). The calculation results generate a spatial error report file.

[0098] The model optimization module fuses the above results to construct a loss function. The function structure adopts a multi-source input design: the first feedback coefficient reflects the time prediction deviation weight, the second feedback coefficient reflects the resource prediction deviation weight, and the spatial error report contributes to the geographical prediction deviation weight. The loss value calculation process does not depend on mathematical formula expression, but is realized through the configuration of weight parameter table: the time prediction weight parameter is initially set to 0.5; the resource prediction weight parameter is initially set to 0.3; the spatial error weight parameter is initially set to 0.2; the module integrates the three inputs in proportion to the weights, and outputs a total loss value in the range of 0-1. This value is input into the neural network trainer to activate the backpropagation algorithm to update the weight parameters of the spatio-temporal graph neural network.

[0099] The training process implements a phased strategy:

[0100] Freeze input layer weights: fine-tune spatial feature extraction layers according to the characteristics of newly added case distribution data;

[0101] Adjust the graph convolution layer: optimize the regional correlation parameters based on the spatial error report;

[0102] Update the fully connected layer: calibrate the prediction output according to the time series feedback data;

[0103] The training cycle is configured to be executed once every three days, and weight checkpoints are retained for each iteration. The update mechanism implements an incremental training mode: only the latest feedback data is loaded instead of the full historical data, saving computing resources. The model weights completed by training are automatically deployed to the production environment, and the version number is incremented for record.

[0104] The exception handling mechanism includes data quality monitoring: automatically switch to a backup data source when the actual propagation attenuation rate is missing for three consecutive days; suspend the coefficient calculation for that dimension when the resource usage deviation exceeds a reasonable threshold; trigger the coordinate system correction program when spatial analysis finds that the map coordinates are drifting. The operation log records the complete flow from feedback collection to model update, and the log items include the original data hash value, the loss weight configuration parameters, and the updated model performance baseline.

[0105] The hardware layer is configured with independent feedback processing clusters: time series feedback processing is executed using CPU clusters; spatial error calculation is scheduled on GPU nodes to accelerate grid operations; model training is allocated dedicated AI acceleration cards. The network topology uses a separate architecture: the collection nodes are connected to the government network, the processing nodes are deployed on the internal computing network, and the update nodes are connected to the model repository.

[0106] Embodiment 5: The operation flow of the iterative early warning module starts with the model update event trigger. When the model optimization module completes the weight parameter update of the spatio-temporal graph neural network, the system automatically issues a model version upgrade notification. The iterative early warning module receives the notification, loads the latest weight file from the model repository into memory. The loading process implements a verification mechanism: calculate the hash value of the weight file and compare it with the repository record, and if the verification fails, revert to the previous stable version. After successful loading, the module calls the operation interface of the deep learning prediction module, inputs the current time's multi-element data set to regenerate the regional propagation dynamic features. The input data includes real-time environmental and meteorological monitoring values, 24-hour statistics of population flow, real-time snapshots of medical resources, and historical epidemic propagation rolling window data. The feature generation process reuses the original network architecture, but uses the updated weight parameters to perform forward propagation calculation. The output feature vector dimension is consistent with the geographical grid division, and each vector element corresponds to the propagation situation encoding of a specific grid.

[0107] After regenerating the regional propagation dynamic features, the module performs threshold monitoring of the variant risk probability. The system extracts the variant risk probability component from the feature vector, which represents the likelihood of the virus strain undergoing major variation. It also queries the historical database to retrieve the probability peak records for the past 365 days in the region. The comparison logic uses a sliding window algorithm: take the maximum value within a seven-day window and compare it with the historical peak value. When the newly generated variant risk probability exceeds the historical peak value for three consecutive window periods, the cross-regional collaborative early warning protocol is triggered. The trigger condition is supplemented with regional correlation verification: if there is a cross-administrative area population flow hotspot in the probability exceeding area, the early warning range is expanded to the adjacent three administrative areas.

[0108] The execution of the cross-regional coordination early warning protocol includes information packaging and routing distribution. The protocol defines a standardized early warning message structure: the message header includes the early warning ID, the release time, and the validity period; the message body records the list of over-standard regions, the probability over-standard amplitude, and the associated regional flow intensity value. After the message is generated, it is submitted to the distribution engine, which selects the transmission path according to the pre-configured recipient list. The provincial health department terminal connects directly through the government affairs special line; the municipal level terminal relies on the health emergency communication network for transmission; the cross-provincial collaboration region uses an encrypted Internet channel for transmission. The transmission protocol implements priority classification: real-time streaming is used for the provincial region, and batch compression transmission is enabled for the cross-provincial region. The receiving terminal returns a signed confirmation code, and unconfirmed messages are re-sent every five minutes until they time out.

[0109] After the early warning is triggered, the priority list of prevention and control resources is updated synchronously. The system inputs the newly generated regional propagation dynamic characteristics into the prevention and control strategy generation module, which recalculates the prevention and control matching score based on the updated propagation dynamic characteristics. The scoring calculation logic maintains the original multi-layer perception network structure, but the input feature vector is replaced with the latest data. The recalculated score value is input into the sorting subsystem, which maintains a global priority queue. The update operation uses an incremental refresh strategy: only the over-standard risk probability regions and their associated regions are recalculated, and the original scores of non-affected regions are retained. The queue sorting algorithm uses the minimum heap adjustment technique to dynamically update the queue position of the regions with score changes. The final output is the new version of the prevention and control resource scheduling priority list, which includes the region code, the latest score value, the version identifier, and the medical resource scheduling recommendation level.

[0110] The overall deployment architecture of the system is implemented based on a distributed computing platform. The platform physical layer includes multiple groups of computing node clusters: the multi-element data acquisition module runs on edge computing nodes, close to the data source to reduce transmission delay; the deep learning prediction module is deployed on a GPU accelerated cluster with high speed NVLink interconnection; the risk region identification and resource matching analysis module runs on a geographic information dedicated server. The spatiotemporal database uses a distributed columnar storage architecture, with data partitioned by time and spatial grid. The storage engine optimizes the compression algorithm for time series data and establishes R-tree indexes for spatial data.

[0111] The matrix operation acceleration of the deep learning prediction module is realized through CUDA kernel functions. The adjacency matrix operation of the graph convolution layer is decomposed into sparse matrix multiplication, which is processed in parallel by the TensorCore unit of the GPU. The time step loop of the long short-term memory network is unfolded into parallel thread blocks, which use shared memory to cache intermediate states. The gradient calculation in the training phase is optimized by the automatic differentiation engine, and the mixed precision strategy is used for backpropagation to reduce memory usage.

[0112] The output instruction of the dynamic decision module is transmitted relying on the Internet of Things interface layer. The interface layer implements a protocol conversion gateway: the internal prevention and control parameter adjustment instruction of the system is encoded into an MQTT protocol payload. The topic naming follows the regional hierarchical specification, such as province-level instruction publishing to / cn / province / instruction type topic. The public health response terminal subscribes to the corresponding topic, receives and decodes the execution parameter adjustment. The terminal execution state is returned through the telemetry channel, the gateway monitors the online state and redirects the offline terminal message. The security mechanism implements bidirectional certificate authentication, and the instruction payload is encrypted using a national encryption algorithm.

[0113] The system operation and maintenance layer implements full-link monitoring. The log collector aggregates various module running indicators: data collection delay, model inference time consumption, instruction transmission success rate, etc. The monitoring dashboard displays the regional risk heat map and resource scheduling path in real time, assisting operation and maintenance personnel in identifying and processing bottlenecks. The version release adopts a blue-green deployment mode, and the new model version is verified in the shadow environment before switching traffic.

[0114] It should be noted that, in this text, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between or among the entities or actions. Also, the terms "comprises", "comprising", or any other variations thereof are intended to cover non-exclusive inclusions, so that a process, method, article, or apparatus that comprises a list of elements does not only include those elements, but also includes other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0115] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, alternatives, and variations can be made thereto without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent epidemic prediction system, characterized in that, Comprise the following modules: Multivariate data acquisition module for real-time acquisition of environmental meteorological data related to epidemic, population flow data, medical resource distribution data and historical epidemic transmission data; Deep learning prediction module for analyzing regional transmission dynamic characteristics based on the environmental meteorological data, population flow data, medical resource distribution data and historical epidemic transmission data through the collaborative architecture of spatio-temporal graph neural network and long short-term memory network; Risk area identification module for constructing a regional risk level heat map based on the regional transmission dynamic characteristics, identifying high-risk areas with risk levels exceeding a preset threshold, and the risk area identification module based on the output embedding vector of the deep learning prediction module, constructing a regional risk level heat map, which includes: inputting the embedding vector into a classification model, which divides the risk level into low, medium and high three categories using support vector machine; The heat map is drawn by geographic information system, and each grid cell corresponds to a risk probability; The preset threshold is set to high risk above the median value, and the system automatically identifies high-risk areas with risk levels exceeding the threshold, and outputs raster data; Resource matching analysis module for calculating the resource coverage of each medical institution based on the medical resource distribution data and the spatial position of the high-risk area; The deep learning prediction module performs: Taking the temperature and humidity parameters in the environmental meteorological data, the cross-regional migration intensity in the population flow data, and the bed turnover rate in the medical resource distribution data as input features; Process the spatial correlation of multi-source data through spatio-temporal graph neural network to provide accurate spatial feature basis for long short-term memory network; Extracting spatio-temporal transmission feature vectors through long short-term memory network; Predicting the transmission rate and mutation risk probability in the future specified time window based on the spatio-temporal transmission feature vectors; When the resource matching analysis module calculates the resource coverage of each medical institution, it specifically performs the following operations: First, obtain the real-time resource capacity parameters of each medical institution, including the number of on-duty medical staff, the number of available beds, and the reserve amount of emergency equipment, and standardize the above parameters to resource sufficiency index in the interval of 0-1; Second, combine the population density data and traffic accessibility data of the high-risk area to establish a coverage radius basic model; Third, dynamically adjust the coverage radius according to the resource sufficiency index: when the resource sufficiency index is greater than or equal to 0.8, the coverage radius is enlarged by 20% on the basis of the basic model result; When 0.5≤resource sufficiency index<0.8, keep the coverage radius of the basic model; When the resource sufficiency index is less than 0.5, the coverage radius is reduced by 30% on the basis of the basic model result; Fourth, perform spatial boundary verification on the adjusted coverage radius to exclude natural geographical barrier areas such as rivers and mountains and traffic control areas, and finally determine the actual resource coverage of each medical institution, output as polygon vector data with geographic coordinates. 2.The deep learning based intelligent epidemic prediction system according to claim 1, wherein, The resource matching analysis module comprises: Spatial overlay analysis of the resource coverage of each medical institution and the high-risk area to generate the overlapping area and the position of the overlapping area of the resource coverage and the high-risk area; According to the location data of the medical institutions and the location of the overlapping area, the interval distance between the medical institutions and the overlapping area is calculated; In combination with the overlapping area, the interval distance and real-time load data of medical resources, the prevention and control potential coefficient of each medical institution is calculated. 3.The deep learning based intelligent epidemic prediction system according to claim 1, wherein, Further comprising: A prevention and control strategy generation module is configured to fuse the prevention and control potential coefficient of each medical institution with the transmission rate and the mutation risk probability; A multi-layer perception network is used to output the prevention and control matching degree score of each area; A prevention and control resource scheduling priority list is generated according to the prevention and control matching degree score. 4.The deep learning based intelligent epidemic prediction system according to claim 3, characterized in that, Further comprising: A dynamic decision module is configured to generate a dynamic adjustment threshold according to the prevention and control resource scheduling priority list and a real-time epidemic severity index; When the real-time epidemic severity index exceeds the dynamic adjustment threshold, a hierarchical response mechanism is activated; Based on the matching result of the prevention and control matching degree score and a preset response rule, a prevention and control parameter adjustment instruction is output. 5.The deep learning based intelligent epidemic prediction system according to claim 4, characterized in that, The prevention and control parameter adjustment instruction includes: Adjusting the detection point density parameter according to the high-risk area area change rate; Adjusting the isolation range parameter according to the ratio of the transmission rate to the medical resource load; Adjusting the vaccine allocation weight parameter according to the mutation risk probability. 6.The deep learning based intelligent epidemic prediction system according to claim 5, wherein, Further comprising: An execution feedback module is configured to collect actual transmission attenuation rate, resource usage deviation value and new case distribution data after the implementation of the prevention and control measures; A first feedback coefficient is generated by comparing the actual transmission attenuation rate with the predicted transmission rate; A second feedback coefficient is generated by correlation analysis of the resource usage deviation value and the prevention and control matching degree score. 7.The deep learning based intelligent epidemic prediction system according to claim 6, characterized in that, Further comprising: A model optimization module is configured to calculate the spatial error of the new case distribution data and the predicted high-risk area; The first feedback coefficient and the second feedback coefficient are fused to construct a loss function; The weight parameters of the spatio-temporal graph neural network are dynamically updated by a back propagation algorithm. 8.The deep learning based intelligent epidemic prediction system according to claim 7, characterized in that, Further comprising: An iterative early warning module is configured to regenerate the regional transmission dynamic characteristics according to the updated deep learning prediction module; When the regenerated mutation risk probability exceeds the historical peak value, a cross-regional collaborative early warning protocol is triggered; The prevention and control resource scheduling priority list is updated based on the recalculated prevention and control matching degree score. 9.The deep learning based intelligent epidemic prediction system according to any one of claims 1-8, characterized in that, The system is deployed on a distributed computing platform and includes a spatio-temporal database for storing the output of the multi-element data acquisition module; The deep learning prediction module accelerates the matrix operation of the spatio-temporal graph neural network through a graphics processing unit; The output instruction of the dynamic decision module is synchronized to a public health response terminal through an Internet of Things interface.

Images