Community epidemic risk assessment model construction method based on big data

Abstract

The invention relates to the field of artificial intelligence, and discloses a community epidemic risk assessment model construction method based on big data, and the method comprises the following steps: S1, collecting data to construct a heterogeneous space-time atlas; s2, performing node representation learning on the map by using a heterogeneous map attention mechanism in a map neural network, and extracting propagation association features between communities; s3, constructing a coupled cellular automaton model, and simulating a disease transmission process between communities; s4, performing structure mapping and adaptive optimization by adopting a structure transfer learning mechanism; and S5, in combination with a multi-agent reinforcement learning strategy, community-level prevention and control strategy recommendation and risk assessment output are carried out by taking the risk cost as an optimization target. The community propagation potential is modeled by combining population mobility, environmental meteorology and medical resource data, accurate description of a propagation chain in space and time dimensions is realized, the dynamic perception ability of a real propagation path is improved, and the problems of rough description of a propagation scene and large prediction deviation are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically to a method for constructing a community epidemic risk assessment model based on big data. Background Art

[0002] In existing community-level epidemiological monitoring and risk assessment research, mainstream methods are mostly based on traditional transmission model construction and statistical feature analysis. Infection trends are generally estimated using historical case data, population density information, and regional contact frequency. Some systems incorporate simplified graphical models to simulate transmission pathways between communities and assess infection risk based on this. Furthermore, some studies combine basic environmental factors and medical resource parameters to conduct a certain degree of scenario simulation to assist in policy formulation or recommend emergency response strategies. For areas with insufficient data, existing methods often achieve horizontal generalization of risk assessment through methods such as model parameter migration or regional classification.

[0003] However, with the increasing complexity of community structures and the intensification of population mobility, existing technologies have gradually revealed bottlenecks in processing high-dimensional heterogeneous information, dynamic transmission paths, and multi-regional coordinated prevention and control. On the one hand, due to the lack of modeling of deep relationships between heterogeneous data, the potential transmission mechanism between communities is difficult to be fully characterized, affecting the timeliness and accuracy of the overall prediction. On the other hand, current methods mostly focus on static analysis or manual rule setting, and are insufficiently responsive in terms of strategy optimization and responding to dynamic changes. Especially in data-sparse areas, model migration is mostly based on a coarse-grained method that is insensitive to structure, making it difficult to effectively reuse empirical information from existing regions, limiting the credibility and adaptability of risk assessment results. Overall, the adaptability, flexibility, and prediction depth of existing solutions in complex scenarios still need to be improved. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for constructing a community epidemic risk assessment model based on big data, which solves the problems of inaccurate characterization of community epidemic transmission patterns, untimely risk assessment, and unintelligent response of prevention and control strategies in a multi-source heterogeneous data environment.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for constructing a community epidemic risk assessment model based on big data, comprising the following steps:

[0006] S1, collects and integrates multi-source spatiotemporal data including population mobility data, environmental meteorological data, and medical resource data, and constructs a heterogeneous spatiotemporal graph containing nodes, edges, and node attributes;

[0007] S2, using the heterogeneous graph attention mechanism in the graph neural network to learn node representations on the graph and extract the communication correlation features between communities;

[0008] S3, based on the above node representation and transmission potential information, a graph-driven coupled cellular automaton model is constructed to simulate the disease transmission process between communities;

[0009] S4, adopts the structure transfer learning mechanism to perform structural mapping and adaptive optimization of the contact matrix between different regions;

[0010] S5, combined with multi-agent reinforcement learning strategy, uses risk cost as the optimization target to recommend community-level prevention and control strategies and output risk assessment.

[0011] Preferably, in step S1, the collection of multi-source spatiotemporal data includes:

[0012] The collection of population mobility data includes obtaining population migration trajectories between different communities at different times through operator mobile phone signaling data or map location service platforms, and constructing an inter-community migration matrix;

[0013] Collection of environmental meteorological data, including obtaining temperature, humidity, and wind speed for each community from the National Meteorological Service, and assigning these values ​​to the community level through spatial interpolation methods;

[0014] Collection of medical resource data, including statistics on the number of hospitals, total number of beds, and average response time in each community, and standardized processing as community attribute input;

[0015] The construction of the heterogeneous spatiotemporal graph includes the following steps:

[0016] Each community is defined as a graph node, and the node attributes include population density, medical resource indicators and environmental meteorological parameters;

[0017] Based on population migration data, determine whether there are edge connections between communities. If there is migration flow, establish graph edges and use the migration frequency or intensity as the edge weight;

[0018] The constructed graph is a heterogeneous graph, where the heterogeneity of node attributes is reflected in the multimodal input features, and the heterogeneity of edges is reflected in the diversity of edge weights, directionality, and temporal information;

[0019] The graph is used to model community communication correlation in subsequent graph neural network models.

[0020] Preferably, in step S2, node representation learning includes:

[0021] For each community node in the graph, an initial feature vector is constructed based on its population, medical and meteorological attributes;

[0022] For different types of graph edges, the attention weights of neighbor nodes to the central node are calculated separately, and a learnable attention function is used to aggregate and weight different edge types.

[0023] The weighted neighbor features are concatenated and mapped to generate a node embedding representation vector that integrates multilateral relationship information as the propagation state encoding of the community;

[0024] The extraction of communication correlation features between communities includes:

[0025] Using the community embedding vector after completing node representation learning and combining it with the migration edge weight information in the graph, we can calculate the communication potential score between any two communities.

[0026] Quantify the potential risk of disease spread between communities by constructing a similarity matrix or a transmission potential matrix;

[0027] This transmission association feature is used in downstream transmission modeling and intervention strategy recommendations to improve the accuracy of risk assessment.

[0028] Preferably, in step S2, node representation learning includes:

[0029] For each community node in the graph, an initial feature vector is constructed based on its population, medical and meteorological attributes;

[0030] For different types of graph edges, the attention weights of neighbor nodes to the central node are calculated separately, and a learnable attention function is used to aggregate and weight different edge types.

[0031] The weighted neighbor features are concatenated and mapped to generate a node embedding representation vector that integrates multilateral relationship information as the propagation state encoding of the community;

[0032] The extraction of communication correlation features between communities includes:

[0033] Using the community embedding vector after completing node representation learning and combining it with the migration edge weight information in the graph, we can calculate the communication potential score between any two communities.

[0034] Quantify the potential risk of disease spread between communities by constructing a similarity matrix or a transmission potential matrix;

[0035] This transmission association feature is used in downstream transmission modeling and intervention strategy recommendations to improve the accuracy of risk assessment.

[0036] Preferably, in step S4, the structure mapping includes the following steps:

[0037] The alignment of inter-regional structures is completed through node rematching;

[0038] Use the structural similarity matrix to establish mapping rules between source and target nodes / edges;

[0039] The adaptive optimization includes:

[0040] Use attention mechanisms to adjust the weights of different structural components;

[0041] Introducing adversarial training to unbiasedly align the target region structure embedding with the source region;

[0042] A residual adaptation module is implemented to correct the structure mismatch error.

[0043] Preferably, in step S4, the structure mapping includes the following steps:

[0044] The alignment of inter-regional structures is completed through node rematching;

[0045] Use the structural similarity matrix to establish mapping rules between source and target nodes / edges;

[0046] The adaptive optimization includes:

[0047] Use attention mechanisms to adjust the weights of different structural components;

[0048] Introducing adversarial training to unbiasedly align the target region structure embedding with the source region;

[0049] A residual adaptation module is implemented to correct the structure mismatch error.

[0050] Preferably, the strategy optimization specifically includes the following steps:

[0051] Embed the contact matrix structure as the agent observation input to construct the local state vector;

[0052] Based on the state vector, the action space of the individual agent in each community node is defined, and the risk exposure cost corresponding to each action is set;

[0053] A centralized training and distributed execution architecture is adopted to enable each intelligent agent to iteratively optimize the policy function under the joint training framework and output community-level intervention strategies and overall risk assessment values.

[0054] Preferably, the propagation potential value is calculated using the following formula:

[0055] P(i,j,t)=β(i,j,t)×n(i,j,t)×(1-p(i,t));

[0056] Among them, β(i,j,t) represents the probability of infection, n(i,j,t) represents the environmental amplification coefficient, and p(i,t) represents the medical response capacity of the community.

[0057] Preferably, the attention weight α(i, j) between the nodes is calculated as follows:

[0058] α(i,j)∧r=softmaxj(LeakyReLU(a∧r·[W∧r·h(i)||W∧r·h(j)]));

[0059] Among them, r represents the edge type, a∧r and W∧r are parameters, and h(i) and h(j) are node features.

[0060] Preferably, the optimization objective function of the community-level prevention and control strategy is defined as:

[0061] R(i,t)=-α×I(i,t)-β×TestCost(i,t)-γ×ShutdownCost(i,t);

[0062] Among them, R(i,t) is the comprehensive risk loss function value of the i-th community at time t; I(i,t) is the number of infections or infection rate in the i-th community at time t; TestCost(i,t) is the testing cost incurred by the i-th community for nucleic acid testing at time t; ShutdownCost(i,t) is the economic and social cost caused by the implementation of lockdown measures in the i-th community at time t; α is the weight coefficient of infection loss; β is the weight coefficient of testing cost; γ is the weight coefficient of lockdown cost.

[0063] The present invention provides a method for constructing a community epidemic risk assessment model based on big data. It has the following beneficial effects:

[0064] 1. This invention utilizes a multi-source heterogeneous graph construction mechanism, combining population mobility, environmental meteorology, and medical resource data to model community transmission potential, achieving precise characterization of transmission chains in both spatial and temporal dimensions. Compared to existing methods that rely solely on single infection data or static contact matrices, this significantly enhances the dynamic perception of actual transmission paths, overcoming the problems of crude transmission scenario characterization and large prediction errors.

[0065] 2. This invention leverages the heterogeneous attention mechanism within graph neural networks to construct a high-dimensional representation model capable of identifying factors influencing community communication. This model not only accurately extracts the drivers of communication but also possesses cross-modal fusion capabilities. Traditional methods often splice together various data types without being able to deeply model their structural connections. This invention effectively bridges this modeling gap, enabling the communication-driving model to have stronger generalization capabilities.

[0066] 3. This invention uses a structural-level migration optimization mechanism to achieve the migration of epidemic risk prediction from known areas to data-scarce areas. Previous migration methods often remain at the parameter level or model initialization stage, making it difficult to adapt to differences in community structure. This invention can automatically map the inherent connections between graph structures, enabling the model to have a high degree of fit even in unsupervised areas, solving the problem of difficulty in modeling weakly labeled areas.

[0067] 4. This invention utilizes multi-agent reinforcement learning as the core unit of strategic control, enabling the risk assessment system to adapt and proactively control emergencies. This system transcends traditional "rule-based" emergency response and instead truly simulates strategic linkages between communities. In particular, this invention overcomes the limitations of manual scheduling and single-minded response, providing proactive, strategic-level prevention and control capabilities in terms of multi-regional resource allocation and the selection of prevention and control opportunities. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] Please see the attached Figure 1 The present invention provides a method for constructing a community epidemic risk assessment model based on big data, comprising the following steps:

[0071] S1, collects and integrates multi-source spatiotemporal data including population mobility data, environmental meteorological data, and medical resource data, and constructs a heterogeneous spatiotemporal graph containing nodes, edges, and node attributes;

[0072] Specifically, in this embodiment, multi-source spatiotemporal data are first collected and integrated, including population flow data, environmental meteorological data, and medical resource data, to establish a multimodal community information foundation and provide data support for subsequent communication map modeling.

[0073] Population mobility data can be collected through mobile phone signaling data from telecom operators or location-based service data from internet map platforms, obtaining information on population migration trajectories between communities over a continuous time series. Based on this information, a population migration intensity matrix can be constructed.

[0074] At the same time, meteorological parameters such as temperature, humidity, and wind speed are obtained from the national meteorological data platform or the urban meteorological sensor system. Considering that meteorological data is usually grid data, spatial interpolation methods are used to map it to the coordinates of each community. For example, the temperature value T(i, t) of community i at time t can be calculated as follows:

[0075] T(i,t)=∑[w k (i)×T k (t)];

[0076] Among them, T k (t) is the measured temperature at meteorological observation station k at time t; w k (i) represents the interpolation weight of community i relative to site k.

[0077] Medical resource data, including the number of hospitals, total beds, and medical response time within a community, constitute the medical resource indicator vector. All raw medical resource data undergoes minimum-maximum normalization to the [0, 1] interval to ensure comparability across variables of different dimensions.

[0078] After data collection is completed, a heterogeneous spatiotemporal graph G = (V, E, A) is established, where:

[0079] Among them, V represents the node set, corresponding to each community unit; E represents the edge set, which represents the population migration relationship, adjacency relationship, etc. between communities; A represents the attribute set, including time series variables such as population density, meteorological data, and medical resources.

[0080] The node attribute vector x(i,t) of community i is defined as:

[0081] x(i,t)=[Pop(i,t),T(i,t),H(i),B(i),τ(i),I(i,t)];

[0082] Where Pop(i,t) is the total population density of community i at time t; Pop(i,t) is the temperature of community i at time t; H(i), B(i), and τ(i) are the number of hospitals, number of beds, and response time in community i, respectively; I(i,t) is the number of infected people in community i at time t.

[0083] The rule for constructing edges in the graph is: if M(i,j,t)>0, then a directed edge is established between community>0 and j, and the unit population emigration rate is used as the edge weight w i j(t) is calculated as follows:

[0084] w i j(t)=M(i,j,t) / Pop(i,t);

[0085] Among them, w ij(t) represents the emigration rate from community i to community j at time t (a reference indicator of transmission potential); M(i,j,t) represents the number of population migrants; Pop(i,t) represents the population size of community i at time t.

[0086] In addition, to reflect the heterogeneity of the graph structure, multiple edge types are introduced into the graph, such as:

[0087] Population migration edge (directed, dynamic);

[0088] Geographic adjacency edges (undirected, static);

[0089] Medical support edge (directed, conditional).

[0090] The heterogeneous graph also has the following characteristics:

[0091] Node attribute heterogeneity: Nodes contain multimodal input features from demographic, medical, environmental, etc.

[0092] Edge structure heterogeneity: edges have different weights, directions, and time attributes;

[0093] Dynamic evolution of the graph: node attributes and some edge weights change over time, forming a dynamic graph structure.

[0094] The graph construction process is as follows:

[0095] First, a node set is established based on the community space distribution, and an initial attribute vector is assigned to each node;

[0096] Then, by analyzing the migration data in continuous time periods, we construct time-stratified migration edges and label them according to edge types;

[0097] Finally, the node attributes and temporal edge weights are fused to generate a dynamic graph G(t) that evolves over time.

[0098] Once the graph is constructed, it will serve as input for subsequent graph neural network models, supporting tasks such as node representation learning, propagation modeling, and control strategy formulation.

[0099] The heterogeneous spatiotemporal graph constructed in this embodiment has a unified representation capability and can map multiple data types and propagation mechanisms into the same modeling framework, thereby providing basic data support for the modeling of complex propagation processes and the generation of control strategies.

[0100] S2 uses the heterogeneous graph attention mechanism in the graph neural network to learn node representations on the graph and extract the communication correlation features between communities;

[0101] Specifically, in this embodiment, based on the heterogeneous spatiotemporal graph G = (V, E, A) constructed in step S1, the heterogeneous graph attention mechanism in the graph neural network is used to perform representation learning on the nodes, aiming to explore the communication correlation characteristics between communities.

[0102] First, a feature vector is initialized for each community node. This vector is constructed based on population density, meteorological parameters, medical resource indicators, and historical infection status.

[0103] Subsequently, attention is calculated for each edge type (such as population migration edge, adjacent edge, and medical support edge). In the heterogeneous graph attention mechanism, attention weights are calculated for the central node and its neighboring nodes of the same type, which are defined as follows:

[0104] α(i,j)∧r=softmaxj(LeakyReLU(a∧r·[W∧r·h(i)||W∧r·h(j)]));

[0105] Among them, r represents the edge type, a∧r and W∧r are parameters, and h(i) and h(j) are node features.

[0106] Through this mechanism, the propagation characteristics under different edge types can be modeled independently, enhancing the model's ability to distinguish heterogeneous propagation relationships.

[0107] Next, we perform weighted aggregation on the neighbor representations of node i under different edge types r to obtain the intermediate node representation under each relationship type:

[0108] h i ∧r=∑i∈N r (i)α i j∧r·W r h(j);

[0109] Among them, h i ∧r represents the neighbor representation of node i after aggregation under edge type r; N r (i) represents the set of all neighbor nodes of node i under edge type r.

[0110] Finally, the aggregation results of all edge types are combined to form the final node embedding representation z(i), which is defined as follows:

[0111] z(i)=σ(∑rh i ∧r)=σ(∑r∑i∈N r (i)α i j∧r·W r ·h(j));

[0112] Among them, z(i) is the final embedding representation of node i, which contains information about multiple propagation channels; represents the activation function, such as ReLU or ELU.

[0113] The above z(i) represents the overall state of community nodes in the multidimensional information space of structure, attributes, and communication relationships, and is an important input for subsequent disease transmission simulation, risk assessment, and strategy optimization.

[0114] After the node representation learning is completed, the communication potential relationship between communities is further calculated. The communication potential score P(i, j, t) of any two communities i and j at time t can be defined as follows:

[0115] P(i,j,t)=β(i,j,t)×η(i,j,t)×(1-p(i,t));

[0116] Among them, represents the probability of transmission from community j to community i; η(i,j,t) represents the environmental transmission enhancement factor, which is usually related to parameters such as wind speed and humidity; p(i,t) represents the medical response capability of community i at time t. The higher the value, the stronger the prevention and control capability.

[0117] p(i,t) can be further modeled as a propagation similarity function between community embeddings, for example using the weighted inner product between embedding vectors.

[0118] S3, based on the above node representation and transmission potential information, a graph-driven coupled cellular automaton model is constructed to simulate the disease transmission process between communities;

[0119] Specifically, in this embodiment, based on the community embedding representation z(i) and the propagation potential matrix obtained in step S2, a graph-driven coupled cellular automaton model is constructed to simulate the epidemic propagation process at the community level.

[0120] First, each community node is considered a cellular unit, forming a multicellular system. The state of each cell is defined as a discrete epidemic state set S∈{S,E,I,R}, which represents the susceptible (Susceptible), latent (Exposed), infected (Infectious), and recovered (Recovered) states respectively.

[0121] At the initial moment, the initial state distribution of each node is determined by inputting the infection rate of the current community. The initial state of the node is determined by its embedding vector and historical infection data.

[0122] Then, based on the adjacent edges and propagation potential information in the heterogeneous graph, a state coupling relationship between nodes is constructed. This coupling relationship makes the evolution of each cell state not only dependent on its own historical state, but also dynamically influenced by the propagation potential and infection state of neighboring cells.

[0123] The coupling relationship works as follows: if at a certain moment a neighboring node is infected and has a high transmission potential score, the probability of the node transitioning from susceptible to latent increases. Accordingly, the dynamic adjustment of the transmission path is controlled by the changes in edge weights and the attention allocation mechanism in the graph.

[0124] The transmission potential value is calculated using the following formula:

[0125] P(i,j,t)=β(i,j,t)×n(i,j,t)×(1-p(i,t));

[0126] Among them, β(i,j,t) represents the probability of infection, n(i,j,t) represents the environmental amplification coefficient, and p(i,t) represents the medical response capacity of the community.

[0127] In the transmission modeling process, an improved SEIR (Susceptible-Latent-Infected-Recovered) modeling framework is adopted. The transition process between each state is as follows:

[0128] Susceptible S → Latent E: The transfer rate is regulated by the infection status of neighboring nodes and the common;

[0129] Latent E→Infected I: Time advances according to the set latent period parameters;

[0130] Infected I→Recovered R: State transition is performed based on the recovery rate parameter.

[0131] The state evolution process is iteratively updated in discrete time steps. In each time step, the system advances according to the following logic:

[0132] Count the infected neighbor set of each node at the current time step;

[0133] The probability of infection transmission is calculated in combination with the transmission potential score;

[0134] The status of each node is updated synchronously to form a new round of system status.

[0135] The transmission potential score plays a regulatory role in this process, enabling the transmission modeling to take into account both community structural differences and the heterogeneity of environment and medical resources, thus achieving dual control of structure-driven and attribute-driven approaches.

[0136] To improve the controllability and stability of the model simulation, we set transmission cutoff conditions during simulation iterations, including a maximum time step limit and an infection growth rate threshold. When the system meets the termination conditions, the model outputs the epidemic spread trend between communities, infection time series, and potential transmission paths.

[0137] The graph-driven coupled cellular automaton model uses heterogeneous graphs as the foundation for the propagation space, community embeddings as the expression of propagation characteristics, and a propagation potential matrix as the state coupling control signal. These three components form a complete propagation dynamics mechanism. This mechanism enables high-precision simulation of community-level epidemics, with high structural sensitivity and strong state expression capabilities.

[0138] S4, adopts the structure transfer learning mechanism to perform structural mapping and adaptive optimization of the contact matrix between different regions;

[0139] Specifically, in this embodiment, to address the problem of lack of complete historical epidemic data in the target area, a structural transfer learning mechanism is introduced to achieve structural mapping and adaptive optimization of the contact matrix between different regions, so as to improve the generalization ability of the propagation model.

[0140] First, the source region and the target region are modeled as heterogeneous graphs G s ource and G t arget, the graph structure consists of community nodes, connection relationships and node attributes. The data source is consistent with step S1. To unify the input form, a graph neural network is used to extract the structural embedding vector of each graph:

[0141]

[0142] Among them, H s ource represents the structural feature matrix of the source region; H t arget represents the structural feature matrix of the target area;

[0143] f(.) is a graph neural network encoder with shared parameters, which is used to extract structural representation.

[0144] The above embedding vector is used to guide the mapping of the source region contact matrix to the target region. s Ource reflects the transmission coupling strength between communities and is generated by the transmission potential matrix and the cellular automaton model. The mapping function corresponds to a structural alignment mechanism and is defined as follows:

[0145] C t arget=M(C s ource,H s ource,H t arget);

[0146] Among them, C t arget is the contact matrix of the target region after migration; M(.) is a trainable structure mapping function used for migration fitting across structures.

[0147] The mapping function M(.) can be composed of a multi-layer perceptron, with input Cs The comparison result of ource and structure embedding vector is output as the adjusted C t arget value, minimize the following structural alignment loss function during the fitting process:

[0148] L a lign=||C t arget-M(C s ource,H s ource,H t arget)|| 2 ;

[0149] where represents the structural mapping error; ||.|| 2 Represents the Frobenius norm, which is used to measure the overall distance between matrices.

[0150] During the structural alignment process, three types of optimization mechanisms are further used to improve migration accuracy:

[0151] Node rematching mechanism: Establishes a remapping relationship between source and target communities based on the similarity of graph node features, improving cross-graph structural correspondence;

[0152] Adversarial training mechanism: Introducing a discriminator makes it impossible to identify the source or target region in the target region’s structural embedding, thus achieving spatial alignment of the embeddings.

[0153] Residual adaptive module: performs residual compensation correction on local structural errors in the image to adapt to the differences in structural details between regions.

[0154] In order to improve the convergence stability of the model, gradient regularization, dropout, structural consistency constraints and other methods can be used during the training process to control the boundaries of the contact matrix variation range to avoid overfitting or structural distortion.

[0155] The final output represents the contact relationship matrix in the target area after transfer learning optimization, which can be used as the basic input for propagation modeling in the target area to achieve high-confidence propagation prediction in unlabeled areas.

[0156] This structural migration mechanism enables unsupervised reasoning of cross-regional transmission structures through shared structural feature representation, mapping function fitting, and adversarial optimization. This mechanism offers strong migration flexibility and low data dependency, enabling scalable deployment of epidemic prediction models across multiple regions and structures.

[0157] S5, combined with multi-agent reinforcement learning strategy, uses risk cost as optimization target to recommend community-level prevention and control strategies and output risk assessment;

[0158] Specifically, in this embodiment, based on the propagation simulation results of step S3 and the target area contact matrix C obtained by optimization in step S4, t arget, builds a multi-agent reinforcement learning environment to achieve dynamic optimization and risk assessment of community-level prevention and control strategies.

[0159] Therefore, the state vector form is expressed as:

[0160] s(i,t)=[I(i,t),R(i),C t arget(i,·),z(i),u(i,t-1)];

[0161] The above status represents the community's current transmission risk, resource availability, transmission connection structure and behavioral history.

[0162] Subsequently, an independent agent is designed for each community node, which decides the current prevention and control strategy based on the status. The agent action space includes but is not limited to:

[0163] Whether nucleic acid screening is enabled (Boolean value);

[0164] Screening frequency (enumeration values: daily, every other day, weekly);

[0165] Regional lockdown level (enumerated values: no lockdown, partial lockdown, full lockdown);

[0166] Medical resource allocation ratio (floating point value).

[0167] Each type of action can be mapped into a vector a(i,t) to control the prevention and control status of the community.

[0168] To achieve collaborative optimization among agents, a joint reward function is constructed with the “risk-cost” balance as the optimization objective. For any community i, the instantaneous reward R(i,t) at time t is defined as follows:

[0169] R(i,t)=-α·I(i,t)-β·TestCost(i,t)-γ·ShutdownCost(i,t);

[0170] Among them, I(i,t) is the number of infections (reflecting the risk); TestCost(i,t) is the testing cost incurred by implementing nucleic acid testing; ShutdownCost(i,t) is the loss caused by the lockdown to the economy and society; α, β, and γ are positive weight coefficients reflecting the importance of each factor.

[0171] This reward function is adjustable and can be flexibly configured according to the policy preferences of different cities or stages. In order to strengthen the synergy between communities, the neighbor influence term is introduced to form the total reward:

[0172] R t otal(i,t)=R(i,t)-λ×∑j∈N(i)C t arget(i,j)×I(j,t);

[0173] Among them, λ is the penalty factor of neighborhood infection; λ is the set of communities adjacent to community i; C t arget(i,j) represents the spread influence of community j on community i.

[0174] The agent's goal is to maximize its long-term cumulative reward, and the strategy optimization objective function is:

[0175]

[0176] Among them, π(i) ★ represents the strategy function of community i; γ is the discount factor; Represents the expected value under environmental interaction.

[0177] Strategy learning uses a centralized training and distributed execution framework. During the training phase, the experience trajectories of all community agents are collected to uniformly optimize the policy network parameters. During the execution phase, each community independently makes policy decisions and adapts to local conditions.

[0178] Strategy learning can adopt multi-agent reinforcement learning algorithms, such as Multi-Agent Deep Deterministic Policy Gradient (MADDPG) and Multi-Agent Proximal Policy Optimization (MAPPO).

[0179] After each round of strategy training is completed, the current optimal strategy is applied to the simulation environment for deduction to obtain the optimal control strategy set and risk dynamic change trend within the corresponding time window of each community.

[0180] The final output includes:

[0181] The sequence of prevention and control strategies recommended by each community (frequency of nucleic acid testing, level of lockdown, etc.);

[0182] Risk level distribution and future infection trends in each community;

[0183] Evaluation of global infection volume and resource allocation efficiency.

[0184] The above strategy recommendation results can be used to provide a quantitative basis for actual epidemic prevention and control decision-making, and realize the deployment of intelligent prevention and control systems based on big data and multi-agent models.

[0185] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a community epidemic risk assessment model based on big data, characterized in that: The following steps are involved: S1, collects and integrates multi-source spatiotemporal data including population mobility data, environmental meteorological data, and medical resource data, and constructs a heterogeneous spatiotemporal graph containing nodes, edges, and node attributes; S2, using the heterogeneous graph attention mechanism in the graph neural network to learn node representations on the graph and extract the communication correlation features between communities; S3, based on the above node representation and transmission potential information, a graph-driven coupled cellular automaton model is constructed to simulate the disease transmission process between communities; S4, adopts the structure transfer learning mechanism to perform structural mapping and adaptive optimization of the contact matrix between different regions; S5, combined with multi-agent reinforcement learning strategy, uses risk cost as the optimization target to recommend community-level prevention and control strategies and output risk assessment.

2. The method for constructing a community epidemic risk assessment model based on big data according to claim 1, characterized in that: In step S1, the collection of multi-source spatiotemporal data includes: The collection of population mobility data includes obtaining population migration trajectories between different communities at different times through operator mobile phone signaling data or map location service platforms, and constructing an inter-community migration matrix; Collection of environmental meteorological data, including obtaining temperature, humidity, and wind speed for each community from the National Meteorological Service, and assigning these values ​​to the community level through spatial interpolation methods; Collection of medical resource data, including statistics on the number of hospitals, total number of beds, and average response time in each community, and standardized processing as community attribute input; The construction of the heterogeneous spatiotemporal graph includes the following steps: Each community is defined as a graph node, and the node attributes include population density, medical resource indicators and environmental meteorological parameters; Based on population migration data, determine whether there are edge connections between communities. If there is migration flow, establish graph edges and use the migration frequency or intensity as the edge weight; The constructed graph is a heterogeneous graph, where the heterogeneity of node attributes is reflected in the multimodal input features, and the heterogeneity of edges is reflected in the diversity of edge weights, directionality, and temporal information; The graph is used to model community communication correlation in subsequent graph neural network models.

3. The method for constructing a community epidemic risk assessment model based on big data according to claim 1, characterized in that: In step S2, node representation learning includes: For each community node in the graph, an initial feature vector is constructed based on its population, medical and meteorological attributes; For different types of graph edges, the attention weights of neighbor nodes to the central node are calculated separately, and a learnable attention function is used to aggregate and weight different edge types. The weighted neighbor features are concatenated and mapped to generate a node embedding representation vector that integrates multilateral relationship information as the propagation state encoding of the community; The extraction of communication correlation features between communities includes: Using the community embedding vector after completing node representation learning and combining it with the migration edge weight information in the graph, we can calculate the communication potential score between any two communities. Quantify the potential risk of disease spread between communities by constructing a similarity matrix or a transmission potential matrix; This transmission association feature is used in downstream transmission modeling and intervention strategy recommendations to improve the accuracy of risk assessment.

4. The method for constructing a community epidemic risk assessment model based on big data according to claim 1, characterized in that: In step S3, the construction of the graph-driven coupled cellular automaton model includes the following steps: The community nodes are regarded as cells, and their status is defined as susceptible, latent, infected or recovered; Use the node representation vector in the graph as the input feature of the initial state of the cell; The inter-cell coupling relationship is constructed based on the edge weights and propagation potential information in the graph, so that the state transition of each cell is not only affected by its own state, but also modulated by the states of adjacent cells and edge weights. Establish state transition rules and time step update mechanism to form a propagation evolution system with graph structure constraints; The process of disease transmission between simulated communities is simulated in the following ways: At each moment, the state update probability of each community node is calculated based on its transmission potential value with neighbor nodes, neighbor infection status and current status; The state transfer process follows the SEIR model framework, and the propagation coefficient is dynamically adjusted by the edge weights and attention mechanism in the graph structure; The simulation process proceeds in discrete time steps until the disease spread stabilizes or the termination condition is met; The simulation results are used to assess the spread trend of the epidemic and provide input for the subsequent optimization of prevention and control strategies.

5. The method for constructing a community epidemic risk assessment model based on big data according to claim 1, characterized in that: In the S4 step, the structure mapping includes the following steps: The alignment of inter-regional structures is completed through node rematching; Use the structural similarity matrix to establish mapping rules between source and target nodes / edges; The adaptive optimization includes: Use attention mechanisms to adjust the weights of different structural components; Introducing adversarial training to unbiasedly align the target region structure embedding with the source region; A residual adaptation module is implemented to correct the structure mismatch error.

6. The method for constructing a community epidemic risk assessment model based on big data according to claim 1, characterized in that: In the S5 step, the multi-agent reinforcement learning strategy constructs the environment state space and action space based on the structural embedding representation and historical behavior trajectory of each region, obtains the community-level intervention strategy through an interactive learning process, and optimizes the strategy using the risk cost function as the reward signal.

7. The method for constructing a community epidemic risk assessment model based on big data according to claim 6, characterized in that: The strategy optimization specifically includes the following steps: Embed the contact matrix structure as the agent observation input to construct the local state vector; Based on the state vector, the action space of the individual agent in each community node is defined, and the risk exposure cost corresponding to each action is set; A centralized training and distributed execution architecture is adopted to enable each intelligent agent to iteratively optimize the policy function under the joint training framework and output community-level intervention strategies and overall risk assessment values.

8. The method for constructing a community epidemic risk assessment model based on big data according to claim 4, characterized in that: The propagation potential value is calculated using the following formula: P(i,j,t)=β(i,j,t)×n(i,j,t)×(1-p(i,t)); Among them, β(i,j,t) represents the probability of infection, n(i,j,t) represents the environmental amplification coefficient, and p(i,t) represents the medical response capacity of the community.

9. The method for constructing a community epidemic risk assessment model based on big data according to claim 3, characterized in that: The attention weight α(i,j) between the nodes is calculated as follows: α(i,j)∧r=softmaxj(LeakyReLU(a∧r·[W∧r·h(i)||W∧r·h(j)])); Among them, r represents the edge type, a∧r and W∧r are parameters, and h(i) and h(j) are node features.

10. The method for constructing a community epidemic risk assessment model based on big data according to claim 1, characterized in that: The optimization objective function of the community-level prevention and control strategy is defined as: R(i,t)=-α×I(i,t)-β×TestCost(i,t)-γ×ShutdownCost(i,t); Among them, R(i,t) is the comprehensive risk loss function value of the i-th community at time t; I(i,t) is the number of infections or infection rate in the i-th community at time t; TestCost(i,t) is the testing cost incurred by the i-th community for nucleic acid testing at time t; ShutdownCost(i,t) is the economic and social cost caused by the implementation of lockdown measures in the i-th community at time t; α is the weight coefficient of infection loss; β is the weight coefficient of testing cost; γ is the weight coefficient of lockdown cost.

Images