A method and system for predicting infectious disease infection based on environmental analysis

Abstract

The application provides an infectious disease infection prediction method and system based on environment analysis, and relates to the technical field of infectious disease infection prediction, and comprises the following steps: collecting multi-source data of each prediction area, aligning the multi-source data with area codes and time stamps, and forming a multivariate time series tensor; adopting an adaptive lag window mechanism to perform weighted aggregation on the historical time series, so as to obtain meteorological lag feature vectors of each area; modeling the meteorological data of each area as a probability distribution, and constructing a dynamic environment similarity graph; constructing a directed propagation graph; at the same time, a static geographical adjacency graph is constructed; a multi-relation graph neural network is adopted to perform spatial aggregation on node features, and then the aggregated time series features are encoded through a gated recurrent unit, so as to output prediction values of future infectious cases of each area. The application can improve the modeling capability of the prediction model for meteorological environment changes and spatial propagation dynamics, and provide more refined prediction basis for public health prevention and control decisions.

Description

Technical Field

[0001] This invention relates to the field of infectious disease infection prediction technology, and in particular to an infectious disease infection prediction method and system based on environmental analysis. Background Technology

[0002] The prevalence and spread of infectious diseases pose a continuous threat to public health security. Their transmission is driven by multiple factors, including meteorological environment, population movement, geographical distribution, and host susceptibility. Meteorological conditions (including temperature, humidity, and ultraviolet radiation) can exert a time-lag regulatory effect on regional infection risk by influencing the survival ability and transmission efficiency of pathogens in the environment. Meanwhile, cross-regional directional population movement is a crucial vehicle for the geographical spread of pathogens, resulting in significant spatial heterogeneity and asymmetry in infection pressure between different regions. With advancements in deep learning and multi-source data fusion technologies, combining meteorological environmental modeling with spatiotemporal transmission dynamics to construct refined methods for predicting infectious disease infection has become a significant research direction in the field of epidemiological forecasting.

[0003] In the prior art, Chinese invention patent CN118866395A discloses an influenza trend prediction method based on multimodal data. This method extracts three types of features: Baidu Index modality, meteorological data modality, and graph data modality. These features are processed by their respective encoders and then fused. The graph structure is then used to model the relationships between regions, ultimately for predicting influenza trends. While this approach introduces meteorological features and graph structure methods into infectious disease prediction tasks and has some practical significance, the correlation between meteorological data and the transmission graph structure relies primarily on data-level splicing and fusion. It does not model the physical mechanisms by which meteorological factors affect pathogen activity, nor does it consider the time lag effect of meteorological factors on transmission intensity and their regional distribution differences. This results in limited sensitivity of the model to changes in the meteorological environment and limited ability to characterize the directionality of cross-regional transmission. Summary of the Invention

[0004] In view of this, the present invention provides an infectious disease infection prediction method and system based on environmental analysis. By adaptively modeling the lag effect of meteorological factors, probabilistically measuring the differences in regional meteorological environment distribution, and decomposing the physical mechanism of cross-regional transmission processes, a multi-relation spatiotemporal graph structure with clear physical meaning is constructed. Combined with graph neural networks and gated recurrent units, multi-step prediction of infectious disease infection cases is realized, thereby improving the prediction model's ability to model changes in meteorological environment and spatial transmission dynamics, and providing more refined prediction basis for public health prevention and control decisions.

[0005] The technical solution of this invention is implemented as follows: On the one hand, this invention provides a method for predicting infectious disease infection based on environmental analysis, including: S1. Collect multi-source data for each prediction area. The multi-source data includes meteorological data, population flow data, historical infection cases and population statistics. Align the multi-source data with regional codes and timestamps to form a multivariate time series tensor. S2. For meteorological data in multivariate time series tensors, an adaptive lag window mechanism is used to weight and aggregate its historical time series to obtain meteorological lag feature vectors for each region. S3. Model the meteorological data of each region as a probability distribution, use the probability distribution distance to measure the environmental differences between regions, and combine the dynamic change rate of environmental differences between regions to generate edge weights and construct a dynamic environmental similarity graph. S4. The cross-regional transmission process is decomposed into three aspects: the pathogen output capacity of the originating region, the directed population flow between regions, and the acceptance capacity of the susceptible population in the destination region. The product of the three is used as the inter-regional transmission intensity to construct a directed transmission map. At the same time, a static geographical adjacency map is constructed based on the spatial adjacency relationship between each region. S5. The meteorological lag feature vector is concatenated with the historical number of infected cases and population statistics to form node features. The dynamic environment similarity graph, directed propagation graph and static geographic adjacency graph are used as graph structure inputs. The node features are spatially aggregated by a multi-relation graph neural network. Then, the aggregated temporal features are encoded by a gated recurrent unit to output the predicted number of future infected cases in each region.

[0006] Preferably, in step S2, the adaptive lag window mechanism is implemented as follows: For the m-th type of meteorological data in region i at time t, let the maximum lag length be L, and assign trainable weight vectors to historical observations over the past L time steps. Normalized weights are obtained after softmax normalization. , ; By applying normalized weights to the historical observations and then summing the results, the lagged aggregate value of this type of meteorological data is obtained: ; in For region i in The m-th type of meteorological data observation value at time m, This is the corresponding lagged aggregation value.

[0007] Preferably, in step S2, the meteorological lag feature vector It is composed of the lagged aggregated values ​​of five types of meteorological data: temperature, relative humidity, ultraviolet radiation, precipitation, and wind speed. ; Subscript These correspond to temperature, relative humidity, ultraviolet radiation, precipitation, and wind speed, respectively.

[0008] Preferably, in step S3, the probability distribution distance metric uses the Wasserstein-2 distance: Meteorological data of region i at time t are modeled as multivariate Gaussian distributions. ,in Let be the mean vector of the multivariate Gaussian distribution of the meteorological state of region i at time t. Let be the covariance matrix of the multivariate Gaussian distribution of the meteorological state of region i at time t; Environmental differences between region i and region j for: ; in For vectors Norm, for The positive definite square root is calculated using eigenvalue decomposition. This is the trace operation of a matrix.

[0009] Preferably, in step S3, the edge weights are generated based on the environmental differences between region i and region j at the current time. and its change relative to the previous moment As input, trained parameters After weighting both factors, the edge weights of the dynamic environment similarity graph are obtained by mapping using the sigmoid function: ; in It is the sigmoid activation function. Let be the edge weights between region i and region j at time t.

[0010] Preferably, in step S4, the inter-regional transmission intensity is based on the pathogen export capacity of the starting region i. Directed population flow from region i to region j and the acceptance capacity of susceptible populations in region j. Definition of product: ; in The population flow from region i to region j per unit time is taken from population flow data in multi-source data. by Construct a directed propagation graph for the weights of the directed edges.

[0011] Preferably, the pathogen export capacity of starting region i Defined as: ; in Let i be the infection rate in region i at time t. Let i be the number of infected cases in region i at time t. Let be the total resident population of region i; The meteorological-pathogen activity function is constructed by multiplying the Arrhenius temperature term, the absolute humidity segment term, and the ultraviolet inactivation term by the hysteresis aggregated values ​​of temperature, relative humidity, and ultraviolet radiation, with the input of temperature, relative humidity, and ultraviolet radiation. The acceptance level of susceptible populations in region j Defined as: ; in Let J be the baseline reproduction number for region j. Let be the density of susceptible populations in region j at time t.

[0012] Preferably, in step S5, the spatial aggregation process of the multi-relationship graph neural network is as follows: Using node features as input, graph convolution is performed on dynamic environment similarity graph, directed propagation graph and static geographic adjacency graph as adjacency structures respectively to obtain a three-way node spatial aggregation representation; With trainable attention weights , , The weighted summation of the three-way node spatial aggregation representations yields the fused node spatial representation. ; in and Updated dynamically over time. The weights are time-independent static weights, and the three satisfy the normalization constraint.

[0013] Preferably, in step S5, the gated loop unit continuously loops within the backtracking window. The fusion node spatial representation sequence at each time step Temporal encoding is performed on the input, and the spatiotemporal hidden layer representation of each region is output. ; After mapping by a multilayer perceptron, the output region i is predicted to have a sequence of infected cases for the next H time steps. ; An end-to-end spatiotemporal prediction model composed of multi-relation graph neural networks, gated recurrent units, and multilayer sensing mechanisms is jointly trained using the mean square error over the entire region and the entire prediction step size as the loss function. ; in Let i be the predicted number of infected cases in region i at time t+h. Where N is the actual number of infected cases, H is the total number of regions, and H is the prediction step size. After training, the predicted sequence The risk thresholds are determined by summing and aggregating data at the county, city, and provincial administrative levels, and combining this with the average historical infection cases at each level, thus outputting the infection risk level for each region.

[0014] In addition, the present invention also provides an infectious disease infection prediction system based on environmental analysis to implement the above-described method, the system comprising: The data acquisition and alignment module is used to collect multi-source data for each prediction area. The multi-source data includes meteorological data, population flow data, historical infection cases and population statistics. The above multi-source data is aligned with regional codes and timestamps to form a multivariate time series tensor. The meteorological lag feature extraction module is used to perform weighted aggregation of historical time series meteorological data in multivariate time series tensors using an adaptive lag window mechanism to obtain meteorological lag feature vectors for each region. The dynamic environment similarity graph construction module is used to model the meteorological data of each region as a probability distribution, use the probability distribution distance to measure the environmental differences between regions, and combine the dynamic change rate of environmental differences between regions to generate edge weights and construct a dynamic environment similarity graph. The directed propagation graph construction module is used to decompose the cross-regional propagation process into three aspects: the pathogen output capacity of the originating region, the directed population flow between regions, and the acceptance capacity of the susceptible population in the destination region. The product of these three factors is used as the inter-regional propagation intensity to construct a directed propagation graph. At the same time, a static geographic adjacency graph is constructed based on the spatial adjacency relationships between regions. The spatiotemporal prediction module is used to concatenate meteorological lag feature vectors with historical infection case numbers and population statistics to form node features. It uses dynamic environmental similarity graphs, directed propagation graphs and static geographic adjacency graphs as graph structure inputs, and employs a multi-relation graph neural network to spatially aggregate the node features. Then, it encodes the aggregated temporal features through a gated recurrent unit and outputs the predicted number of future infection cases for each region.

[0015] The present invention has the following advantages over the prior art: (1) This invention introduces a trainable attention weight vector for each type of meteorological factor, and automatically learns the optimal lag response window of each factor through end-to-end training, avoiding the subjective bias of manually setting a fixed lag period, and making the temporal expression of meteorological characteristics more in line with the actual biological response law.

[0016] (2) This invention models the regional meteorological state as a multivariate Gaussian distribution, uses Wasserstein-2 distance to measure the environmental differences between regions, and introduces the rate of change of environmental similarity as an auxiliary criterion for graph edge weights, so that the graph structure can make a forward-looking response to the convergence trend of the regional environment and enhance the real-time adaptability of the graph topology to the dynamic changes of the environment.

[0017] (3) This invention decomposes the cross-regional transmission process into three physical links: the pathogen output capacity at the source, the directed flow of population along the route, and the acceptance capacity of susceptible populations at the destination. Each link is modified by introducing a meteorological-pathogen activity function, and the product of the three constitutes the directed transmission intensity. This design makes the transmission intensity matrix naturally asymmetric, with clear physical meaning, and supports end-to-end training, allowing for adaptive optimization of the meteorological response parameters of each link from the data.

[0018] (4) This invention simultaneously constructs a dynamic environment similarity graph, a directed propagation graph, and a static geographical adjacency graph. Figure Three The graph structures characterize three dimensions of regional dependence: environment-driven similarity, physical propagation channel strength, and spatial proximity. The three graph structures are then fused with attention weighting through a multi-relationship graph neural network, which enables the spatial aggregation process to take into account the inter-regional associations of different types and avoids the limitation of a single graph structure in expressing propagation relationships incompletely.

[0019] (5) This invention inputs the spatial aggregation representation of the multi-relation graph neural network into the gated recurrent unit for temporal encoding, captures the temporal evolution of infection trends, and finally outputs a multi-step prediction sequence through a multi-layer perceptron. The sequence is aggregated at the county, city and provincial levels and combined with historical infection benchmarks to output graded risk assessment results, which can provide quantitative reference for disease control decision-making at different spatial scales and time spans. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the end-to-end spatiotemporal prediction model of the present invention; Figure 3 This is a schematic diagram of the three types of graph construction mechanisms of the present invention; Figure 4 This is a system framework diagram of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this invention provides a method for predicting infectious disease infection based on environmental analysis, comprising: S1. Collect multi-source data for each prediction area. The multi-source data includes meteorological data, population flow data, historical infection cases and population statistics. Align the multi-source data with regional codes and timestamps to form a multivariate time series tensor. S2. For meteorological data in multivariate time series tensors, an adaptive lag window mechanism is used to weight and aggregate its historical time series to obtain meteorological lag feature vectors for each region. S3. Model the meteorological data of each region as a probability distribution, use the probability distribution distance to measure the environmental differences between regions, and combine the dynamic change rate of environmental differences between regions to generate edge weights and construct a dynamic environmental similarity graph. S4. The cross-regional transmission process is decomposed into three aspects: the pathogen output capacity of the originating region, the directed population flow between regions, and the acceptance capacity of the susceptible population in the destination region. The product of the three is used as the inter-regional transmission intensity to construct a directed transmission map. At the same time, a static geographical adjacency map is constructed based on the spatial adjacency relationship between each region. S5. The meteorological lag feature vector is concatenated with the historical number of infected cases and population statistics to form node features. The dynamic environment similarity graph, directed propagation graph and static geographic adjacency graph are used as graph structure inputs. The node features are spatially aggregated by a multi-relation graph neural network. Then, the aggregated temporal features are encoded by a gated recurrent unit to output the predicted number of future infected cases in each region.

[0024] like Figure 2As shown, the overall technical concept of this invention is as follows: Addressing the problems of imprecise meteorological environment modeling and insufficient characterization of cross-regional transmission relationships in infectious disease infection prediction, this invention constructs an end-to-end prediction model with a multi-relationship spatiotemporal graph neural network at its core. This model consists of three interconnected functional modules: meteorological lag feature extraction, dynamic graph construction, and spatiotemporal prediction. All modules share a single set of parameters and are jointly trained end-to-end using a unified loss function, eliminating sub-models that undergo phased pre-training or independent training. Specifically, the model takes meteorological data, population flow data, historical infection case counts, and population statistics from each region as input. Step S2 generates meteorological lag feature vectors. Steps S3 and S4 construct a dynamic environment similarity graph, a directed propagation graph, and a static geographical adjacency graph, respectively. Step S5 uses a graph attention network (GAT) to perform spatial aggregation of the three graphs, followed by temporal encoding by a gated recurrent unit (GRU). Finally, a multilayer perceptron (MLP) outputs a multi-step prediction sequence of future infection cases. The meteorological-pathogen activity function parameters and edge weight generation parameters involved in steps S3 and S4, as well as the hysteresis window weights in step S2, are all used as trainable parameters within the model and are updated end-to-end through backpropagation along with the network weights of the graph neural network and GRU.

[0025] In one embodiment of the present invention, step S1 includes: Multi-source data (province, city, and county levels) is collected from various administrative units within the forecast area. For meteorological data, daily or hourly temperature data is collected. (Unit: Kelvin, K), Relative Humidity (Unit: %) Ultraviolet radiation intensity (Using the internationally accepted Ultraviolet Index (UVI)) Precipitation (Unit: mm) and wind speed (Unit: m / s), data sources include ground station observation data from the National Meteorological Administration and meteorological gridded reanalysis datasets (such as ERA5). Relative humidity... In subsequent steps, the absolute humidity is converted according to the standard meteorological formula. (Unit: g / m) 3 ), used for pathogen activity modeling.

[0026] Regarding population mobility data, a directed migration matrix reflecting the actual intensity of population migration between regions is collected. This refers to the number of people moving from region i to region j per unit of time. Data sources include operator signaling data, traffic big data, or flow data based on publicly available migration indices. This data is inherently directional. ), which is the key input for the subsequent construction of the asymmetric propagation intensity matrix.

[0027] Regarding historical COVID-19 data, the number of daily reported infection cases in each region was collected. The data is sourced from publicly available official data from health commissions at all levels and is used as a supervisory signal during the model training phase.

[0028] Regarding population statistics, the total resident population of each region is collected. and density of susceptible populations (Unit: people / km) 2 Considering the impact of factors such as immunization and previous infections on the size of the susceptible population, Dynamic calculation based on the product of total population density and the estimated susceptible proportion in the current period.

[0029] The above four types of data are uniformly aligned with regional codes and timestamps to form a multivariate time series tensor covering N regions and spanning T time spans. When the collected meteorological data comes from a gridded reanalysis dataset, the gridded data needs to be cropped and spatially aggregated according to administrative boundaries to match the regional spatial resolution. For missing values ​​in each type of data, linear interpolation or spatial interpolation between adjacent stations is used. For outliers, they are removed based on the 3σ criterion of the historical average for the same period.

[0030] In one embodiment of the present invention, step S2 includes: The impact of meteorological factors on the risk of infectious disease infection exhibits a time lag effect. The lag period for different meteorological factors varies depending on the type of pathogen and regional differences, making it difficult to adapt to changes in actual scenarios using a fixed lag period. Therefore, for region i, time t, and the m-th type of meteorological factor, we first construct a historical observation vector within the maximum candidate lag period L (e.g., L=21 days): ; Based on this, a trainable attention weight vector is introduced. Normalized weights are obtained after softmax normalization. , By summing the historical observations using normalized weights, we obtain the lag aggregate value of factor m in region i at time t: ; in For region i in The raw observation value of the m-th type of meteorological data at time m. This corresponds to the lagged aggregation value. For factor m in the lag period The trainable scalar weights are defined on the model. Softmax normalization ensures that the sum of the weights is 1 and all are positive. Physically, it represents a probability-weighted average of observations over the past L time steps. The lag period in which the weight peaks is located represents the optimal response window for that factor learned by the model. All weights As the model is trained and updated end-to-end, the weight vectors of different meteorological factors are independent of each other, allowing the model to learn different lag distribution patterns for each factor.

[0031] The lagged aggregated values ​​of five types of meteorological data—temperature, relative humidity, ultraviolet radiation, precipitation, and wind speed—are concatenated to form the meteorological lagged feature vector of region i at time t. : ; The subscripts T, RH, U, P, and W correspond to temperature, relative humidity, ultraviolet radiation, precipitation, and wind speed, respectively. This vector serves two purposes in subsequent steps: firstly, as part of the meteorological-pathogen activity function in step S4. The hysteresis aggregation value of each component Firstly, it serves as input; secondly, it is a component of the initial features of the spatiotemporal prediction model nodes in step S5.

[0032] The improvement of the adaptive lag window mechanism lies in the fact that, compared with the fixed lag period or the method of manually selecting the optimal delay days, the trainable attention weights can automatically search for the optimal response window of each meteorological factor to the risk of infection during the end-to-end optimization process, and the weights of different factors can be learned independently, thereby avoiding the bias caused by empirical settings and improving the temporal expression ability of meteorological characteristics.

[0033] In one embodiment of the present invention, step S3 includes: Single meteorological observations are insufficient to describe the spatial heterogeneity of meteorological conditions within a region. This invention models the meteorological state of each region i at time t as a multivariate Gaussian distribution. , where the mean vector It is composed of the mean of the raw meteorological observations of all meteorological stations (or grid points) in the region at time t, where d is the dimension of the meteorological factors; covariance matrix The sample covariance of the original observations from each station is estimated to capture the spatial dispersion of meteorological factors within the region and the correlation structure between factors. The distribution parameters are constructed using the original multidimensional observations from each station (rather than the single lag-aggregated scalar output from step S2) because the estimation of the covariance matrix depends on the spatial differences among multiple stations within the region, while lag aggregation has compressed each station into a single representative value for the region, which can no longer reflect the spatial heterogeneity between stations. To ensure the positive definiteness of the covariance matrix to support subsequent square root calculations, a regularization correction is applied: ; in The regularization coefficient (e.g., take) ), It is a d-order identity matrix, ensuring matrix invertibility and numerical stability of square root calculation.

[0034] Environmental differences between region i and region j Using the Wasserstein-2 distance metric, the following closed analytical expression exists for the multivariate Gaussian distribution: ; in For vectors Norm, For the trace operation of a matrix, for The positive definite square root (calculated by eigenvalue decomposition). It is the positive definite square root of the matrix product, i.e., the Bures metric. The smaller the value, the more similar the meteorological environment distributions of the two regions, and stronger connections should be established in the graph structure. The above analytical expression is closed and differentiable, supports gradient backpropagation, and can directly participate in optimization during end-to-end training.

[0035] Constructing graph edges solely based on current distance ignores the dynamic evolution of regional environmental similarities. Two regions currently far apart but rapidly converging may quickly develop similar high-risk meteorological conditions within the next few days; this forward-looking signal should be incorporated into the graph structure in advance. This invention defines regional pairs. The change in environmental differences at time t is: ; When converged When diverging Given the current environmental differences and its change As input, trained parameters After weighting both factors, the edge weights of the dynamic environment similarity graph are obtained by mapping using the sigmoid function: ; in The sigmoid activation function maps the edge weights to (0,1); It is a trainable scalar that controls the intensity of the suppression of edge weights by the current environment distance; the greater the distance, the lower the edge weight. For a trainable scalar, when the two regions are converging ( )hour, It has a positive effect on edge weights; the faster the convergence, the higher the edge weight. and These are all trainable parameters within the model, updated end-to-end with the overall model. Below the preset threshold When this occurs, the edge is not activated, achieving adaptive sparsity of the graph structure and avoiding noise interference from redundant connections. Finally, a dynamic environment similarity graph is constructed. , where the node set For all N regions, edge set and their weights It changes dynamically over time, reflecting the similar structure of the meteorological environment in different regions in real time.

[0036] like Figure 3 As shown, in one embodiment of the present invention, step S4 includes: The cross-regional transmission process is decomposed into three physically independent components: the pathogen export capacity of the originating region, the directed population flow along the route, and the acceptance capacity of the susceptible population in the destination region. The directed transmission intensity from region i to region j at time t is also considered. Defined as the product of three segments: ; in The pathogen output capacity of originating region i. Let represent the directed population flow from region i to region j. To assess the acceptance level of the susceptible population in region j. Since all three segments possess endogenous asymmetry, It emerges naturally from the physical structure, without the need for human intervention.

[0037] Source pathogen output capability Defined as the product of the current infection case rate and the meteorological-pathogen activity coefficient: ; in Let i be the infection rate in region i at time t. Let i be the number of infected cases in region i at time t. Let be the total resident population of region i. The temperature, relative humidity, and ultraviolet radiation output from step S2 are used as inputs. The reason for using the lag aggregated values ​​instead of the current period's original observation values ​​is that pathogen activity is also affected by meteorological conditions, which also has a time lag effect. The pathogen output capacity in the current period is the result of the combined effect of meteorological conditions over the past few days, which is consistent with the modeling logic of step S2. The meteorological-pathogen activity function, which integrates the effects of three meteorological factors on the survival and transmission ability of pathogens in the environment, is defined as follows: ; in The temperature hysteresis aggregation value of region i at time t (unit: K); The reference temperature (K) is usually taken as 298.15 K; The hysteresis polymerization value is determined by temperature and relative humidity. , The absolute humidity lag aggregation value (g / m³) obtained by conversion according to the standard meteorological formula 3 ); Let be the hysteresis aggregation value of the ultraviolet index for region i at time t; Let be the temperature sensitivity parameter, a trainable real number, corresponding to the ratio of activation energy to gas constant in the Arrhenius equation. ; Let be the ultraviolet radiation inactivation parameter, and be a trainable positive real number.

[0038] The temperature term is expressed in the reciprocal temperature form of the Arrhenius equation. Within the range of low to normal temperatures, increased temperature accelerates viral envelope degradation, reducing its survival rate in the environment. The ultraviolet radiation term is based on first-order photochemical inactivation dynamics; the greater the UV radiation intensity, the higher the inactivation rate of pathogens in outdoor environments.

[0039] Humidity item Designed to meet critical absolute humidity A piecewise exponential function is used as the boundary to reflect the experimentally supported U-shaped dependency between absolute humidity and virus survival rate: ; in Critical absolute humidity threshold (g / m 3 The initial value is set based on the crystallization relative humidity (ERH) mechanism, corresponding to the point with the highest virus survival rate (i.e., the bottom of the U-shape), and is used as a trainable parameter for fine-tuning during training. denoted as the decay rate of pathogen activity on the high humidity side as humidity increases, and denoted as a trainable positive real number. The decay rate of pathogen activity on the low-humidity side as humidity deviates from the threshold is represented by a trainable positive real number. The physical meaning of this segmented design is as follows: when absolute humidity is too high, large aerosol particles settle more rapidly, shortening the virus's suspension time in the air; when absolute humidity is extremely low, salt crystals in the aerosol destroy the virus particle structure; and near moderate humidity, the virus survival rate is highest, but the risk of transmission is greatest. Considering these three factors... All parameters These are all trainable parameters within the model, and are jointly optimized with the overall model.

[0040] Path flux This represents the population flow intensity from region i to region j, normalized by the maximum network traffic: ; in Let be the average daily directed population moving from region i to region j, and after normalization, take a value in (0,1).

[0041] The acceptance level of susceptible populations in region j It is composed of the product of the normalized susceptible population density and the weather-corrected local basic reproduction number: ; in Let be the normalized susceptible population density of region j at time t (normalized by the maximum density of the entire network, with a value in (0,1]). The local basic reproduction number, after weather correction, is modeled as follows: ; in The baseline reproduction number is a uniform, trainable positive real number across the entire network. The initial value is set with reference to the empirical values ​​of different infectious diseases (e.g., 1.3 for influenza and 2.5 for COVID-19) and is optimized during training. The destination-side meteorological-pathogen activity function has the same form as the source-side function. Completely identical (temperature Arrhenius term × humidity U-shaped segment term × UV inactivation term), also using the lagged aggregated values ​​of various meteorological factors at the destination as input, but with different parameter sets. With source parameter set They are independent of each other and do not share resources. (Source end) This describes the ability of pathogens to remain active and be exported from their origin's meteorological environment to their destination. It describes the efficiency of the spread of foreign pathogens in the local population under the current meteorological conditions in the target area. The biological processes corresponding to the two are different, and their response coefficients to meteorological factors such as temperature and humidity may not be the same. Therefore, they are parameterized independently, and the model learns the optimal parameters for each from the data.

[0042] by Construct a directed propagation graph for the weights of directed edges. ,in The (i,j)th element is It is an asymmetric matrix. Simultaneously, a static geographic adjacency map is constructed based on the spatial adjacency relationships of each administrative region. The edge weights are normalized geographical adjacency strengths, constructed once during system initialization and not updated throughout the system's lifecycle. This forms a dynamic environment similarity graph. Dynamic directed propagation graph and static geographic adjacency map The three types of diagrams respectively depict regional similarity connections, physical propagation channel strength, and spatial proximity relationships driven by meteorological environment.

[0043] In one embodiment of the present invention, step S5 includes: The meteorological lag feature vector obtained in step S2 Compared with the characteristics of the current epidemic (infection case rate) (e.g., 7-day moving average) and demographic characteristics (susceptible population density) The features are concatenated to form the node input features of region i at time t: ; in This represents the input feature dimension after concatenation.

[0044] The spatial aggregation process of the multi-relationship graph neural network takes node features as input and performs graph convolution on dynamic environment similarity graphs, directed propagation graphs, and static geographical adjacency graphs as adjacency structures to obtain three-way node spatial aggregation representations. Specifically, for the three types of graph structures, their edge weight matrices drive the corresponding graph attention aggregation operation (using the multi-head attention aggregation method of the Graph Attention Network GAT) to obtain three sets of node neighborhood aggregation representations. Trainable attention weights are used... , , The weighted summation of the three-way node spatial aggregation representations yields the fused node spatial representation. : ; in For the node feature matrix of all regions, This is the fused node space representation matrix. For hidden layer dimensions. and It is dynamically updated over time (predicted by a linear layer from the global state features at the current time step). The three weights are time-independent static weights, satisfying normalization constraints, and together reflect the relative importance of different graph relationships to infection transmission at different times.

[0045] Gated Recurrent Unit (GRU) spatially represents the sequence using fused nodes for a consecutive \tau+1 time steps within the backtracking window. Time-series encoding is performed on the input to capture the temporal evolution and periodic patterns of infection risk (backtracking window length). (For example, taking 14 days). The time-series encoding process for region i is as follows: ; in This represents the spatial fusion of region i at time t. This represents the hidden state of the GRU at the previous time step. This represents the hidden layer representation of the region at the current moment after spatiotemporal joint encoding. The GRU's forget gate automatically reduces seasonal background noise, while the update gate retains recent outbreak signals, aligning well with the dynamic characteristics of infectious disease outbreaks. Choosing GRU as the temporal encoder, compared to temporal convolutional networks (TCNs), demonstrates stronger generalization ability in scenarios with relatively limited sample sizes, such as regional infectious disease prediction. Furthermore, its node-level progressive hidden state update logic naturally corresponds to the node representation framework of graph neural networks, facilitating the implementation of spatiotemporal joint encoding.

[0046] Characterizing the spatiotemporal hidden layer After mapping using a multilayer perceptron (MLP, a two-layer fully connected network, with ReLU activation in the first layer and linear output in the second layer), the predicted sequence of the number of infected cases in region i over the next H time steps is output: ; Where H is the prediction step size (e.g., 7 days or 14 days). This is the predicted number of infections for region i over the next H time steps.

[0047] This invention is a unified end-to-end spatiotemporal prediction model, consisting of three interconnected parts: a multi-relation graph neural network, a gated recurrent unit, and a multilayer perceptron. It is jointly trained using the mean squared error over the entire region and the entire prediction step size as the loss function. ; in Let i be the predicted number of infected cases in region i at time t+h. Here, N represents the actual number of infected cases, N is the total number of regions, and H is the prediction step size. All trainable parameters—including those from the source—are included. function parameters Destination end function parameters Baseline Reproduction Number Lag window weight Graph fusion attention weights , generate parameters The graph attention network and all network weights of the GRU—all parameters of the same model—are optimized using the Adam optimizer with a unified loss function. Backpropagation is performed, and joint optimization is carried out in one go. There is no independent training process in stages or modules.

[0048] After training, the predicted sequence The data is aggregated level by level, from county to city to province, and then combined with the total population of each region. Calculate the infection rate for each region at each predicted time step. Based on the predicted infection rate, risk thresholds are defined by combining the average historical infection cases at each level. The system outputs the corresponding infection risk level (e.g., low risk, medium risk, high risk, very high risk) for each region at each time step according to preset grading thresholds, forming an infection risk prediction map with multiple time steps and spatial scales. Each grading threshold can be flexibly configured by the user according to actual management needs, without being bound to any fixed early warning triggering logic.

[0049] like Figure 4 As shown, in another embodiment of the present invention, an infectious disease infection prediction system based on environmental analysis is also provided to implement the above method. The system includes a data acquisition and alignment module, a meteorological lag feature extraction module, a dynamic environment similarity graph construction module, a directed propagation graph construction module, and a spatiotemporal prediction module. Each module has independent functions and clear interfaces, and they are connected in series through a unified data tensor format.

[0050] The data acquisition and alignment module is responsible for continuously acquiring multi-source input data from various external data sources, and completing data cleaning, alignment, and tensor quantization. This module has two sub-layers: a data access layer and a preprocessing layer. The data access layer interfaces with meteorological data (such as meteorological bureau APIs or ERA5 dataset download pipelines), population flow data (such as operator signaling data and transportation big data interfaces), public epidemic data interfaces from the National Health Commission, and population census statistics databases. It supports both timed polling and incremental update acquisition modes to ensure data timeliness. The preprocessing layer performs operations such as missing value imputation (time-step linear interpolation or adjacent station spatial interpolation), outlier removal (3σ truncation based on historical averages), and dimensional normalization (min-max or z-score normalization) on each data source. It also performs field-by-field alignment of the four types of data using a unified regional code (administrative division code) and timestamp (Unix millisecond timestamp). The aligned data is then packaged and stored in a multivariate time-series tensor with dimensions of number of regions × number of time steps × number of features for downstream modules to read. When the collected meteorological data comes from a gridded reanalysis dataset, the preprocessing layer is also responsible for cropping and aggregating the gridded data according to administrative boundaries to match the spatial resolution at the regional level.

[0051] The meteorological lag feature extraction module reads meteorological components from a multivariate time series tensor, performs weighted aggregation of the historical time series of each meteorological factor based on an adaptive lag window mechanism, and outputs the meteorological lag feature vectors for each region. The module is structurally composed of three parts: a lag candidate sequence construction unit, a trainable weight storage unit, and a weighted aggregation unit. The lag candidate sequence construction unit slides backward along the time axis, extracting the complete historical original observation sequence for each region and each meteorological factor within the maximum candidate lag period, forming a three-dimensional candidate sequence tensor (region × meteorological factor × candidate lag step number). The historical observation values ​​of each candidate lag step are directly indexed from the multivariate time series tensor. The trainable weight storage unit maintains a separate set of lag step weight parameters for each type of meteorological factor, continuously updating them through backpropagation during model training. After training, the weights are fixed for use in the inference phase. For multi-pathogen or multi-seasonal scenarios, this unit supports storing weight reassemblies separately by pathogen type or seasonal interval. The weighted aggregation unit performs factor-wise weighted summation on the candidate sequence input at the current time, using softmax-normalized weights, to obtain the lag aggregation value of each meteorological factor. The aggregation results of the five meteorological factors are then concatenated according to the factor dimension to obtain the meteorological lag feature vector of each region at the current time. Write it to the feature buffer for use in step S4 of the dynamic environment similarity graph construction module. Function calls and spatiotemporal prediction module node feature assembly and reading.

[0052] The dynamic environment similarity graph construction module takes the original multi-site meteorological observation data of each region as input and completes three sub-tasks: distributed modeling of regional meteorological states, measurement of environmental differences between regions, and generation of dynamic graph topology, outputting the dynamic environment similarity graph at the current time. This module maintains a regional distribution parameter buffer to store the mean vector and covariance matrix of the multivariate Gaussian distribution of the meteorological states of each region at the current and previous times. At each time step, the distribution estimation sub-unit reads the original multidimensional meteorological observation data of all meteorological stations (or grid points) in each region from the multivariate time series tensor, calculates the sample mean and sample covariance matrix of the features of each station in the region, applies regularization correction to the covariance matrix, writes the updated distribution parameters into the buffer, and retains the distribution parameters from the previous time step. The distance calculation sub-unit calculates the environmental differences for each pair of regional distribution parameters at the current time based on the closed Bures metric analytical expression of the Wasserstein-2 distance, obtaining the environmental difference matrix; simultaneously, it extracts the environmental difference matrix stored at the previous time step and subtracts it element-wise from the difference matrix at the current time step to obtain the environmental difference change matrix. To control computational load, in actual deployment, Wasserstein-2 distance calculation can be enabled only for regions with geographical distances within a set threshold range, while distant regions are set to zero, trading limited precision for linearized computational overhead. The edge weight generation subunit maps the current environmental difference value and its corresponding change to edge weights using trainable parameters and a sigmoid function. Edge connections are pruned for regions with weights below the sparsity threshold, ultimately generating a dynamic environmental similarity graph for the current moment, stored as a sparse adjacency matrix, retaining only the indices and weights of active edges.

[0053] The directed propagation graph construction module is responsible for calculating the directed propagation intensity of each region pair based on a three-stage physical decomposition framework, integrating population flow data, infection case data, susceptible population statistics, and meteorological lag features output from step S2, while maintaining a static geographic adjacency graph. This module includes a source-end calculation subunit, a path flux normalization subunit, a destination-end calculation subunit, and a graph assembly subunit. The source-end calculation subunit reads the number of infection cases and the number of permanent residents in each region at the current moment from historical epidemic data, calculates the infection rate of each region, and then reads the lag aggregation values ​​of each region output by the meteorological lag feature extraction module from the feature buffer. After converting to absolute humidity, input the meteorological-pathogen activity function. Calculate the pathogen export capacity at the source on a region-by-region basis. ; All parameters are stored in the model parameter registry and updated during overall model training. The path flux normalization subunit reads the directed population flow of each region pair at the current time from the multivariate time series tensor, traverses the entire network to find the maximum flow, and performs network-wide normalization on the flow of all region pairs to obtain the normalized path flux matrix. The destination-side computation subunit uses the normalized susceptible population density of each region as a basis, and similarly calls the lagged aggregate value of each region to substitute into the destination-side meteorological-pathogen activity function. Based on the baseline reproduction number, calculate the acceptance capacity of susceptible populations in each region. ; and While sharing a function form, each unit uses an independent parameter set, with both sets of parameters stored separately in the model parameter registry. The graph assembly subunit merges the source capability vector, normalized path flux matrix, and destination acceptor capability vector element-wise according to a three-segment product rule, generating a directed propagation strength matrix, which is stored in a directed sparse graph format. For region pairs with propagation strength below a set lower limit, edge connections are pruned to maintain graph sparsity. The static geographic adjacency graph is constructed once from administrative boundary data during system initialization and stored as a fixed sparse adjacency matrix, remaining unchanged throughout the entire runtime lifecycle.

[0054] The spatiotemporal prediction module aggregates the outputs of all the above modules, completing node feature initialization, multi-relation graph spatial aggregation, GRU temporal encoding, and multi-step prediction output. It is the core inference module of the system. The node feature assembly subunit assembles the meteorological lag feature vector at each time step. Features such as the current infection rate and its recent moving average, and the normalized density of susceptible populations are concatenated along the regional dimension to form a node feature matrix for all regions. The spatial aggregation subunit is based on a graph attention network (GAT) structure, using a dynamic environment similarity graph, a directed propagation graph, and a static geographic adjacency graph as adjacency structures, and independently performs three-way multi-head attention graph convolutions. The output of each path is a node neighborhood aggregation representation under the corresponding graph relationship. The graph fusion subunit reads the dynamic fusion weights at the current time step from the model parameter registry (environment graph weights and propagation graph weights are predicted by a linear layer from global state features, and geographic graph weights are static trainable scalars), performs a weighted summation on the three-way aggregation representations after softmax normalization, outputs the fusion node spatial representation matrix at the current time step, and writes it into the GRU hidden state update queue. The temporal coding subunit consists of independent GRU units for each region. At each time step, it inputs the fused node spatial representation at the current moment and the GRU hidden state of each region at the previous moment into the standard GRU update equation, outputting a new hidden state as the spatiotemporal hidden layer representation of each region at the current moment. For historical hidden states that exceed the backtracking window, a truncated backpropagation (TBPTT) strategy is used to balance long-range temporal dependencies and training efficiency. The prediction output subunit maps the spatiotemporal hidden layer representation of each region at the current moment into a predicted sequence of infection cases for the next H time steps through a two-layer fully connected network (first layer ReLU activation, second layer linear output). The predicted sequence is then aggregated level by level according to the three administrative levels of county, city, and province, combined with historical infection benchmarks to output a graded risk level, and written into the system's prediction result storage area for external display interfaces or decision support systems to access.

[0055] The system operates according to the following process during actual operation. In the offline training phase, the data acquisition and alignment module extracts multi-source data from the historical database in batches using a sliding window. After constructing a multivariate time-series tensor, it sequentially drives the meteorological lag feature extraction module, the dynamic environment similarity graph construction module, and the directed propagation graph construction module. At each training time step, various graph structures and node features are gradually generated and input into the spatiotemporal prediction module for forward propagation. The predicted number of cases and the actual number of cases are substituted into the mean squared error loss function, and the result is processed by the Adam optimizer (the learning rate can be set to...). (Combined with cosine annealing scheduling) Backpropagation is performed on all trainable parameters, iterating until the model converges. The parameters of each module are then fixed and stored after training. During the online inference phase, the data acquisition and alignment module automatically pulls the latest multi-source data according to the set acquisition cycle (e.g., once a day). After preprocessing and alignment, it triggers the downstream modules to update sequentially: the meteorological lag feature extraction module drives weighted aggregation with the latest meteorological data and outputs the latest meteorological lag feature vector; the dynamic environment similarity graph construction module updates the regional distribution parameters, calculates the new Wasserstein-2 distance matrix and generates new edge weights, and refreshes the dynamic environment similarity graph at the current time; the directed propagation graph construction module updates the population flow flux, source capacity and destination acceptance capacity, and refreshes the directed propagation graph at the current time; after completing spatial aggregation and GRU hidden state update, the spatiotemporal prediction module outputs the predicted sequence of county, city and province-level infection cases and corresponding risk levels for several future time steps, writes the results to the result storage area and notifies the external caller.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting infectious disease infection based on environmental analysis, characterized in that, include: S1. Collect multi-source data for each prediction area. The multi-source data includes meteorological data, population flow data, historical infection cases and population statistics. Align the multi-source data with regional codes and timestamps to form a multivariate time series tensor. S2. For meteorological data in multivariate time series tensors, an adaptive lag window mechanism is used to weight and aggregate its historical time series to obtain meteorological lag feature vectors for each region. S3. Model the meteorological data of each region as a probability distribution, use the probability distribution distance to measure the environmental differences between regions, and combine the dynamic change rate of environmental differences between regions to generate edge weights and construct a dynamic environmental similarity graph. S4. The cross-regional transmission process is decomposed into three aspects: the pathogen output capacity of the originating region, the directed population flow between regions, and the acceptance capacity of the susceptible population in the destination region. The product of the three is used as the inter-regional transmission intensity to construct a directed transmission map. At the same time, a static geographical adjacency map is constructed based on the spatial adjacency relationship between each region. S5. The meteorological lag feature vector is concatenated with the historical number of infected cases and population statistics to form node features. The dynamic environment similarity graph, directed propagation graph and static geographic adjacency graph are used as graph structure inputs. The node features are spatially aggregated by a multi-relation graph neural network. Then, the aggregated temporal features are encoded by a gated recurrent unit to output the predicted number of future infected cases in each region.

2. The method according to claim 1, characterized in that, In step S2, the adaptive lag window mechanism is implemented as follows: For the m-th type of meteorological data in region i at time t, let the maximum lag length be L, and assign trainable weight vectors to historical observations over the past L time steps. Normalized weights are obtained after softmax normalization. , ; By applying normalized weights to the historical observations and then summing the results, the lagged aggregate value of this type of meteorological data is obtained: ； in For region i in The m-th type of meteorological data observation value at time m, This is the corresponding lagged aggregation value.

3. The method according to claim 2, characterized in that, In step S2, the meteorological lag feature vector It is composed of the lagged aggregated values ​​of five types of meteorological data: temperature, relative humidity, ultraviolet radiation, precipitation, and wind speed. ； Subscript These correspond to temperature, relative humidity, ultraviolet radiation, precipitation, and wind speed, respectively.

4. The method according to claim 1, characterized in that, In step S3, the probability distribution distance metric uses the Wasserstein-2 distance: Meteorological data of region i at time t are modeled as multivariate Gaussian distributions. ,in Let be the mean vector of the multivariate Gaussian distribution of the meteorological state of region i at time t. Let be the covariance matrix of the multivariate Gaussian distribution of the meteorological state of region i at time t; Environmental differences between region i and region j for: ； in For vectors Norm, for The positive definite square root is calculated using eigenvalue decomposition. This is the trace operation of a matrix.

5. The method according to claim 4, characterized in that, In step S3, the edge weights are generated based on the environmental differences between region i and region j at the current time. and its change relative to the previous moment As input, trained parameters After weighting both factors, the edge weights of the dynamic environment similarity graph are obtained by mapping using the sigmoid function: ； in It is the sigmoid activation function. Let be the edge weights between region i and region j at time t.

6. The method according to claim 3, characterized in that, In step S4, the inter-regional transmission intensity is determined by the pathogen export capacity of the starting region i. Directed population flow from region i to region j and the acceptance capacity of susceptible populations in region j. Definition of product: ； in The population flow from region i to region j per unit time is taken from population flow data in multi-source data. by Construct a directed propagation graph for the weights of the directed edges.

7. The method according to claim 6, characterized in that, Pathogen output capacity of origin region i Defined as: ； in Let i be the infection rate in region i at time t. Let i be the number of infected cases in region i at time t. Let be the total resident population of region i; The meteorological-pathogen activity function is constructed by multiplying the Arrhenius temperature term, the absolute humidity segment term, and the ultraviolet inactivation term by the hysteresis aggregated values ​​of temperature, relative humidity, and ultraviolet radiation, with the input of temperature, relative humidity, and ultraviolet radiation. The acceptance level of susceptible populations in region j Defined as: ； in Let J be the baseline reproduction number for region j. Let be the density of susceptible populations in region j at time t.

8. The method according to claim 1, characterized in that, In step S5, the spatial aggregation process of the multi-relationship graph neural network is as follows: Using node features as input, graph convolution is performed on dynamic environment similarity graph, directed propagation graph and static geographic adjacency graph as adjacency structures respectively to obtain a three-way node spatial aggregation representation; With trainable attention weights , , The weighted summation of the three-way node spatial aggregation representations yields the fused node spatial representation. ; in and Updated dynamically over time. The weights are time-independent static weights, and the three satisfy the normalization constraint.

9. The method according to claim 8, characterized in that, In step S5, the gated loop unit continuously... The fusion node spatial representation sequence at each time step Temporal encoding is performed on the input, and the spatiotemporal hidden layer representation of each region is output. ; After mapping by a multilayer perceptron, the output region i is predicted to have a sequence of infected cases for the next H time steps. ; An end-to-end spatiotemporal prediction model composed of multi-relation graph neural networks, gated recurrent units, and multilayer sensing mechanisms is jointly trained using the mean square error over the entire region and the entire prediction step size as the loss function. ； in Let i be the predicted number of infected cases in region i at time t+h. Where N is the actual number of infected cases, H is the total number of regions, and H is the prediction step size. After training, the predicted sequence The risk thresholds are determined by summing and aggregating data at the county, city, and provincial administrative levels, and combining this with the average historical infection cases at each level, thus outputting the infection risk level for each region.

10. An infectious disease infection prediction system based on environmental analysis, characterized in that, The system is used to implement the method according to any one of claims 1-9, and the system comprises: The data acquisition and alignment module is used to collect multi-source data for each prediction area. The multi-source data includes meteorological data, population flow data, historical infection cases and population statistics. The above multi-source data is aligned with regional codes and timestamps to form a multivariate time series tensor. The meteorological lag feature extraction module is used to perform weighted aggregation of historical time series meteorological data in multivariate time series tensors using an adaptive lag window mechanism to obtain meteorological lag feature vectors for each region. The dynamic environment similarity graph construction module is used to model the meteorological data of each region as a probability distribution, use the probability distribution distance to measure the environmental differences between regions, and combine the dynamic change rate of environmental differences between regions to generate edge weights and construct a dynamic environment similarity graph. The directed propagation graph construction module is used to decompose the cross-regional propagation process into three aspects: the pathogen output capacity of the originating region, the directed population flow between regions, and the acceptance capacity of the susceptible population in the destination region. The product of these three factors is used as the inter-regional propagation intensity to construct a directed propagation graph. At the same time, a static geographic adjacency graph is constructed based on the spatial adjacency relationships between regions. The spatiotemporal prediction module is used to concatenate meteorological lag feature vectors with historical infection case numbers and population statistics to form node features. It uses dynamic environmental similarity graphs, directed propagation graphs and static geographic adjacency graphs as graph structure inputs, and employs a multi-relation graph neural network to spatially aggregate the node features. Then, it encodes the aggregated temporal features through a gated recurrent unit and outputs the predicted number of future infection cases for each region.

Images