Infectious disease transmission prediction method and system based on dynamic space-time diagram neural network

Abstract

The invention relates to the technical field of epidemic situation transmission prediction, in particular to an infectious disease transmission prediction method and system based on a dynamic space-time diagram neural network, and the method comprises the following steps: extracting space-time features from historical case data through a dynamic space-time diagram learning module, and generating an adjacent matrix representing an infectious disease transmission network; predicting a future epidemic situation development trend based on the adjacent matrix through an infectious disease dynamic modeling module; wherein the dynamic space-time diagram learning module comprises a time sequence convolutional network, a multi-head self-attention mechanism and a long short-term memory network; the infectious disease dynamic modeling module comprises a time sequence convolutional network, a graph neural network and a gating circulation unit.According to the infectious disease transmission prediction method and system based on the dynamic space-time graph neural network, by introducing virtual nodes and a dynamic graph structure, the response ability to external factors is improved, and the prediction accuracy is improved; data missing and external influence factors can be effectively processed, and a real propagation mode can be embodied from the generated propagation network.

Description

Technical Field

[0001] The present invention relates to the technical field of epidemic transmission prediction, and specifically to an infectious disease transmission prediction method and system based on a dynamic spatio-temporal graph neural network. Background Art

[0002] Accurately predicting the dynamics of an epidemic has become an important task in public health decision-making. Traditional epidemic prediction models, such as the SIR (Susceptible-Infected-Removed) and SEIR (Susceptible-Exposed-Infected-Removed) models, although providing a basic framework for describing disease transmission, usually focus on the analysis of the state transitions of each group, but are inadequate when dealing with complex spatial and temporal dependencies. These models often assume that the epidemic spreads in a closed environment, ignoring the dynamic interactions between regions and the impact of external factors such as population mobility and policy changes on disease transmission. Previous studies have shown that this limitation leads to a significant reduction in the accuracy of traditional models in the application of epidemic prediction, especially when the disease network exhibits spatial and temporal variations.

[0003] In recent years, the development of complex network analysis and graph neural network (GNN) technology has provided new ideas for epidemic modeling. Through dynamic graph neural networks, researchers can dynamically capture the disease transmission network, adapting to the dynamic changes within a region and the influences from neighboring regions or the outside. However, there are still many challenges in current research, especially in the simulation of the impact of an epidemic on population mobility. Many existing models rely on traffic flow networks or human mobility data. The acquisition of such data is often difficult and cannot fully reflect the dynamic transmission of the disease. For example, the mobility network may overlook some important transmission paths, resulting in an incomplete or inaccurate representation of the epidemic network.

[0004] Although existing epidemic models provide important theoretical support for basic disease transmission analysis, their scope of application and accuracy have obvious deficiencies. Traditional SIR and SEIR models usually assume that regions are closed and perform transmission predictions based on static data, failing to fully consider factors such as population mobility and policy interventions. Therefore, once these models are applied to a dynamic epidemic environment, their prediction results often deviate far from the real situation. In addition, the limitations of many traditional models are reflected in their dependence on the mobility network. Traffic flow and human-to-human mobility data are often difficult to obtain, resulting in the constructed epidemic transmission network lacking authenticity.

[0005] In addition, to handle dynamic changes, existing models usually need to continuously update historical data. However, the acquisition of data is often limited by the observed regions and time, resulting in the neglect of important transmission paths and interaction relationships. This shortcoming significantly reduces the prediction accuracy of traditional models when facing complex and uncertain epidemic transmissions, thus affecting the formulation and implementation of public health policies. Therefore, in view of the above situation, there is an urgent need to develop an infectious disease transmission prediction method and system based on dynamic spatio-temporal graph neural networks to overcome the deficiencies in current practical applications. Summary of the Invention

[0006] An object of the present invention is to provide an infectious disease transmission prediction method and system based on dynamic spatio-temporal graph neural networks to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] An infectious disease transmission prediction method based on dynamic spatio-temporal graph neural networks includes the following steps:

[0009] Extract spatio-temporal features from historical case data through a dynamic spatio-temporal graph learning module to generate an adjacency matrix representing the infectious disease transmission network;

[0010] Predict the future epidemic development trend based on the adjacency matrix through an epidemiology dynamic modeling module;

[0011] Among them, the dynamic spatio-temporal graph learning module includes a temporal convolutional network, a multi-head self-attention mechanism, and a long short-term memory network, which are used to capture the spatio-temporal dependence relationships between regions;

[0012] The epidemiology dynamic modeling module includes a temporal convolutional network, a graph neural network, and a gated recurrent unit, which are used to predict the epidemic trend by combining virtual nodes.

[0013] As a further aspect of the present invention: The specific steps for the dynamic spatio-temporal graph learning module to generate the adjacency matrix include:

[0014] Use a temporal convolutional network to extract temporal features from the input data to obtain a temporal feature matrix;

[0015] Calculate the attention weights between regions through a multi-head self-attention mechanism to generate a weighted feature representation;

[0016] Process the weighted feature representation through a long short-term memory network to generate an adjacency matrix representing the transmission relationship between regions.

[0017] As a further aspect of the present invention: The calculation of the multi-head self-attention mechanism includes:

[0018] Calculate the query matrix Q = ZW Q, key matrix K = ZW K and value matrix V = ZW V , where Q represents the query matrix; K represents the key matrix; V represents the value matrix; Z represents the input feature; W Q represents the weight of the query matrix; W K represents the weight of the key matrix; W V represents the weight of the value matrix;

[0019] Calculate the attention weight through the attention mechanism: where att represents the calculated attention weight; K T represents the transpose of the key matrix K; d k represents the dimension of the key vector, used to scale the dot product to prevent gradient disappearance; softmax represents the Softmax function, which converts the calculated attention scores into a probability distribution;

[0020] Apply the attention weight to the value matrix to obtain the weighted feature representation: Z att = att · V.

[0021] As a further solution of the present invention: The epidemiology dynamic modeling module introduces virtual nodes, and the generation method thereof includes:

[0022] Perform a pooling operation on the time feature matrix to aggregate the information of all nodes;

[0023] Extract virtual node features from the pooled features through a multi-layer perceptron;

[0024] Add the pooled features to the output of the multi-layer perceptron to generate the virtual node feature representation x Λ , and the specific formula is as follows:

[0025] x Λ = f init (Pooling(X tcn )) + Pooling(X tcn );

[0026] where Pooling(·) represents the average pooling operation, used to aggregate the information of all nodes; f init is a function that extracts useful information from the pooled features through an MLP to generate the feature representation of the virtual node.

[0027] As a further solution of the present invention: The graph neural network updates the node features through multi-layer graph convolution operations, and the specific formula is as follows:

[0028]

[0029] where, Denote the updated feature of node i at time t, i.e., the hidden state; W represents the learnable weight matrix; Denote the neighbor set of node i at the k-th layer; Denote the normalized attention weight from node i to its neighbor node j at the k-th layer at time t; Denote the temporal feature of node j extracted from the input data at time t; Denote the normalized attention weight from node i to the virtual node Λ at time t; Denote the feature of the virtual node at time t;

[0030] Subsequently, the gated recurrent unit module is used to capture the long-term dependencies in the time series. The hidden state update formula of the gated recurrent unit is:

[0031] H gru = GRU(H, H prev );

[0032] where, H gru represents the output of the GRU module, H is the hidden state of the GRU module at the current moment, and H prev is the hidden state of the GRU module at the previous moment;

[0033] Finally, by concatenating the output of the GRU module with the temporal feature X tcn and then inputting it into a multi-layer perceptron to generate the prediction result That is:

[0034]

[0035] where, represents the predicted value of the future time step; is the concatenation symbol.

[0036] As a further solution of the present invention: The method further includes optimizing the model through a loss function. The loss function includes the case number prediction error and the propagation network alignment error. The specific formula is as follows:

[0037]

[0038] where, is the mean square error of case number prediction; is the mean square error of propagation network alignment; λ is a hyperparameter used to balance the importance of the two parts of the loss; M is the content that can be provided by datasets such as population flow information and geographical connectivity information, which can assist in generating the infectious disease propagation network to a certain extent; Y represents the true value corresponding to the prediction result in the dataset; The adjacency matrix generated for the above dynamic spatio-temporal graph learning module reflects the real infectious disease transmission network.

[0039] An infectious disease transmission prediction system based on a dynamic spatio-temporal graph neural network, comprising:

[0040] A dynamic spatio-temporal graph learning module, configured to extract spatio-temporal features from historical case data and generate an adjacency matrix;

[0041] An infectious disease dynamic modeling module, configured to predict the future epidemic development trend based on the adjacency matrix;

[0042] Wherein, the structures of the dynamic spatio-temporal graph learning module and the infectious disease dynamic modeling module are the same as those of the corresponding modules in the above method.

[0043] As a further solution of the present invention: the dynamic spatio-temporal graph learning module includes a temporal convolutional network, a multi-head self-attention mechanism, and a long short-term memory network;

[0044] The infectious disease dynamic modeling module includes a temporal convolutional network, a graph neural network, and a gated recurrent unit.

[0045] As a further solution of the present invention: the system further includes a virtual node generation unit, configured to generate a virtual node feature representation x Λ , and the specific formula is as follows:

[0046] x Λ = f init (Pooling(X tcn )) + Pooling(X tcn );

[0047] Wherein, Pooling(·) represents an average pooling operation, used to aggregate the information of all nodes; f init is a function for extracting useful information from the pooled features through an MLP, and generates the feature representation of the virtual node.

[0048] As a further solution of the present invention: the system further includes a loss function optimization unit, configured to optimize the model through the case number prediction error and the propagation network alignment error, and the specific formula is as follows:

[0049]

[0050] Wherein, is the mean square error of case number prediction; is the mean square error of propagation network alignment; λ is a hyperparameter, used to balance the importance of the two parts of the loss; M is the content that can be provided by datasets such as population flow information and geographical connectivity information, and can assist in the generation of the infectious disease transmission network to a certain extent; Y represents the prediction result in the dataset The corresponding true value; It is the adjacency matrix generated by the above dynamic spatio-temporal graph learning module, reflecting the real infectious disease transmission network.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] By combining the dynamic spatio-temporal graph learning module and the dynamic modeling module of epidemiology, the present invention significantly improves the accuracy and robustness of infectious disease transmission prediction. At the same time, in the prediction tasks of COVID-19 and influenza in multiple countries and regions, it outperforms traditional models (such as linear regression and LSTM, etc.), especially maintaining high accuracy in the case of sparse data (50% regional observations).

[0053] By introducing virtual nodes and dynamic graph structures to improve the response ability to external factors, the model can effectively handle data missing and external influencing factors. At the same time, it can reflect the real transmission pattern from the generated transmission network (the Hits@10 index exceeds 0.5), providing an intuitive analysis of the transmission relationship, which not only improves the prediction accuracy but also effectively guides the epidemic control measures.

[0054] In addition, through the dynamic learning and prediction of the graph neural network, the present invention will significantly enhance the understanding and control ability of the epidemic transmission, more flexibly respond to public health emergencies, provide an effective support model for public health decision-making, and also demonstrate strong generalization ability, being applicable to the prediction tasks of various infectious diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the overall architecture in the embodiment of the present invention;

[0056] Among them, a is the overall framework of the present invention, b is an example of the regional case number of historical data, c is the dynamic spatio-temporal graph learning module in the framework, and d is the epidemic dynamic modeling module in the framework. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than 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 making creative efforts belong to the scope of protection of the present invention.

[0058] The following describes the specific implementation of the present invention in detail in conjunction with specific embodiments.

[0059] Please refer to Figure 1, the method for predicting the spread of infectious diseases based on a dynamic spatio-temporal graph neural network provided by the embodiments of the present invention includes the following steps:

[0060] Extract spatio-temporal features from historical case data through a dynamic spatio-temporal graph learning module to generate an adjacency matrix representing the infectious disease transmission network;

[0061] Predict the future epidemic development trend based on the adjacency matrix through an epidemiological dynamic modeling module;

[0062] Among them, the dynamic spatio-temporal graph learning module includes a temporal convolutional network, a multi-head self-attention mechanism, and a long short-term memory network, which are used to capture the spatio-temporal dependence relationships between regions;

[0063] The epidemiological dynamic modeling module includes a temporal convolutional network, a graph neural network, and a gated recurrent unit, which are used to predict the epidemic trend in combination with virtual nodes.

[0064] In an embodiment of the present invention, the specific steps for the dynamic spatio-temporal graph learning module to generate an adjacency matrix include:

[0065] Use a temporal convolutional network to extract temporal features from the input data to obtain a temporal feature matrix;

[0066] Calculate the attention weights between regions through a multi-head self-attention mechanism to generate a weighted feature representation;

[0067] Process the weighted feature representation through a long short-term memory network to generate an adjacency matrix representing the propagation relationship between regions.

[0068] In an embodiment of the present invention, the calculation of the multi-head self-attention mechanism includes:

[0069] Calculate the query matrix Q = ZW Q 、the key matrix K = ZW K and the value matrix V = ZW V . Among them, Q represents the query matrix; K represents the key matrix; V represents the value matrix; Z represents the input feature; W Q represents the weight of the query matrix; W K represents the weight of the key matrix; W V represents the weight of the value matrix;

[0070] Calculate the attention weights through the attention mechanism: where att represents the calculated attention weights; K T represents the transpose of the key matrix K; d k represents the dimension of the key vector, which is used to scale the dot product to prevent gradient disappearance; softmax represents the Softmax function, which converts the calculated attention scores into a probability distribution;

[0071] Apply the attention weights to the value matrix to obtain a weighted feature representation: Z att = att·V.

[0072] In one embodiment of the present invention, the epidemiology dynamic modeling module introduces virtual nodes, and the generation method thereof includes:

[0073] Perform a pooling operation on the time feature matrix to aggregate the information of all nodes;

[0074] Extract virtual node features from the pooled features through a multi-layer perceptron;

[0075] Add the pooled features to the output of the multi-layer perceptron to generate a virtual node feature representation x Λ , and the specific formula is as follows:

[0076] x Λ = f init (Pooling(X tcn )) + Pooling(X tcn );

[0077] Among them, Pooling(·) represents an average pooling operation for aggregating the information of all nodes; f init is a function that extracts useful information from the pooled features through an MLP to generate a feature representation of the virtual node.

[0078] In one embodiment of the present invention, the graph neural network updates the node features through multi-layer graph convolution operations, and the specific formula is as follows:

[0079]

[0080] Among them, represents the updated feature of node i at time t, that is, the hidden state; W represents a learnable weight matrix; represents the neighbor set of node i in the k-th layer; represents the normalized attention weight from node i to its neighbor node j in the k-th layer at time t; represents the time feature of node j extracted from the input data at time t; represents the normalized attention weight from node i to the virtual node Λ at time t; represents the feature of the virtual node at time t;

[0081] Subsequently, the gated recurrent unit module is used to capture the long-term dependencies in the time series, and the hidden state update formula of the gated recurrent unit is:

[0082] H gru = GRU(H, H prev );

[0083] Among them, H gru represents the output of the GRU module, H is the hidden state of the GRU module at the current moment, and H prev is the hidden state of the GRU module at the previous moment;

[0084] Finally, by concatenating the output of the GRU module with the time feature X tcn and then inputting it into a multi-layer perceptron (MLP) as a fully connected output layer to generate a prediction result That is:

[0085]

[0086] Among them, represents the predicted value of the future time step; is the concatenation symbol.

[0087] In an embodiment of the present invention, the method further includes optimizing the model through a loss function, and the loss function includes a case number prediction error and a propagation network alignment error. The specific formula is as follows:

[0088]

[0089] Among them, is the mean square error of case number prediction; is the mean square error of propagation network alignment; λ is a hyperparameter used to balance the importance of the two parts of the loss; M is the content that can be provided by datasets such as population flow information and geographical connectivity information, which can assist in the generation of the infectious disease propagation network to a certain extent; Y represents the true value corresponding to the prediction result in the dataset ; is the adjacency matrix generated by the above dynamic spatio-temporal graph learning module, reflecting the real infectious disease propagation network.

[0090] In an embodiment of the present invention, an infectious disease transmission prediction system based on a dynamic spatio-temporal graph neural network includes:

[0091] A dynamic spatio-temporal graph learning module for extracting spatio-temporal features from historical case data and generating an adjacency matrix;

[0092] An infectious disease dynamic modeling module for predicting the future epidemic development trend based on the adjacency matrix;

[0093] Among them, the structures of the dynamic spatio-temporal graph learning module and the infectious disease dynamic modeling module are the same as those of the corresponding modules in the above method.

[0094] In one embodiment of the present invention, the dynamic spatio-temporal graph learning module includes a temporal convolutional network, a multi-head self-attention mechanism, and a long short-term memory network;

[0095] The dynamic epidemiology modeling module includes a temporal convolutional network, a graph neural network, and a gated recurrent unit.

[0096] The system further includes a virtual node generation unit for generating a virtual node feature representation x Λ , and the specific formula is as follows:

[0097] x Λ = f init (Pooling(X tcn )) + Pooling(X tcn );

[0098] where Pooling(·) represents an average pooling operation for aggregating information of all nodes; f init is a function for extracting useful information from the pooled features through an MLP to generate the feature representation of the virtual node.

[0099] The system further includes a loss function optimization unit for optimizing the model through the case number prediction error and the propagation network alignment error. The specific formula is as follows:

[0100]

[0101] where is the mean squared error of case number prediction; is the mean squared error of propagation network alignment; λ is a hyperparameter for balancing the importance of the two parts of the loss; M is the content that can be provided by datasets such as population flow information and geographical connectivity information, which can assist in generating the infectious disease propagation network to a certain extent; Y represents the true value corresponding to the prediction result in the dataset; is the adjacency matrix generated by the above dynamic spatio-temporal graph learning module, reflecting the real infectious disease propagation network.

[0102] An overall overview of the dynamic epidemiology modeling framework is as Figure 1As shown, this framework is mainly composed of two core parts: the dynamic spatio-temporal graph learning module (sub-graph c) and the epidemic dynamics modeling module (sub-graph d). The dynamic spatio-temporal graph learning module (sub-graph c) uses a Temporal Convolutional Network (TCN), a Multi-Head Attention (MHA) mechanism, and a Long Short-Term Memory (LSTM) network to reconstruct a dynamic interaction network (as shown in the middle of sub-graph a) from historical case data (as shown in the historical cases on the left side of sub-graph a), aiming to capture the spatio-temporal dependencies evolving over time between different regions and map the regional case features shown in sub-graph b to the corresponding interaction network. Subsequently, the reconstructed dynamic interaction network is input into the epidemic dynamics modeling module (sub-graph d), which adopts a framework based on Graph Neural Networks (GNN) and node long short-term memory networks to predict the future epidemic spread trend (as shown in the predicted cases on the right side of sub-graph a). Among them, the circles represent different geographical regions, where the solid circles indicate the regions where cases have been observed, and the dashed circles represent the regions where cases have not been observed. Diamond nodes are introduced as virtual nodes to simulate the influence of external factors or systems. The edges connecting these nodes represent the interaction weights between regions, and the color of the edges fades from yellow to red, indicating that the interaction intensity changes from weak to strong. At the same time, the depth of the color of the circles reflects the number of cases in each region, and the darker the blue, the more cases there are in that region. By adaptively modeling spatio-temporal dependencies, this framework provides a data-driven method for understanding and predicting the spread dynamics of the epidemic.

[0103] This model is mainly composed of two core modules: the dynamic spatio-temporal graph learning module and the epidemic dynamics modeling module.

[0104] The core task of the dynamic spatio-temporal graph learning module is to extract spatio-temporal features from the input time series data and generate an adjacency matrix representing the infectious disease transmission network. This adjacency matrix can reflect the transmission relationships between different regions and provide a basis for subsequent epidemic prediction.

[0105] The input data is represented as where represents the set of real numbers, L is the time step, N is the number of regions, and D is the feature dimension (such as the number of cases). First, the present invention extracts time features through a temporal convolutional network. The temporal convolutional network captures the dependencies in the time series through one-dimensional convolutional operations, and the output is where F is the feature dimension after convolution.

[0106] Next, the present invention introduces a multi-head self-attention mechanism to enhance information interaction between regions. The specific steps are as follows:

[0107] 1. Calculate the query matrix Q = ZW Q and the key matrix K = ZW K and the value matrix V = ZW V . Where Q represents the query matrix; K represents the key matrix; V represents the value matrix; Z represents the input feature; W Q represents the weight of the query matrix; W K represents the weight of the key matrix; W V represents the weight of the value matrix;

[0108] 2. Calculate the attention weight through the attention mechanism: where att represents the calculated attention weight; K T represents the transpose of the key matrix K; d k represents the dimension of the key vector, which is used to scale the dot product to prevent gradient disappearance; softmax represents the Softmax function, which converts the calculated attention scores into a probability distribution;

[0109] 3. Apply the attention weight to the value matrix to obtain the weighted feature representation: Z att = att·V.

[0110] To further capture the interaction between regions, the present invention reshapes the attention features and calculates the interaction weight matrix E between regions through matrix multiplication. Subsequently, E is processed by LSTM to capture the dependencies in the time dimension, and the output is E LSTM . Finally, an adjacency matrix is generated through a Multilayer Perceptron (MLP) and a Sigmoid activation function This matrix represents the propagation relationship between regions.

[0111] The dynamic modeling module of epidemiology further predicts the future epidemic development trend based on the adjacency matrix generated by the dynamic spatio-temporal graph learning module . This module consists of three sub-modules: a Temporal Convolutional Network module (TCN module), a Graph Neural Network module (GNN module), and a Gated Recurrent Unit (GRU) module (GRU module).

[0112] First, the TCN module extracts time features from the input data, and the output is X tcn . Then, the GNN module enhances the generalization ability of the model by introducing virtual nodes. The features of the virtual nodes are generated through pooling operations and MLP, denoted as x Λ . The specific formula is as follows:

[0113] x Λ = f init (Pooling(X tcn )) + Pooling(X tcn );

[0114] Where Pooling(·) represents the average pooling operation for aggregating information of all nodes; f init is a function that extracts useful information from the pooled features through an MLP, generating the feature representation of the virtual node. The final virtual node feature x Λ is obtained by adding the original pooled features and the MLP output through a residual connection, retaining the original information and enhancing the stability and performance of the model. The GNN module updates the node features through multiple layers of graph convolution operations, and the formula is as follows:

[0115]

[0116] Where represents the updated feature of node i at time t, that is, the hidden state; W represents the learnable weight matrix; represents the set of neighbors of node i in the k-th layer; represents the normalized attention weight from node i to its neighbor node j in the k-th layer at time t; represents the temporal feature of node j extracted from the input data at time t; represents the normalized attention weight from node i to the virtual node Λ at time t; represents the feature of the virtual node at time t;

[0117] Subsequently, the gated recurrent unit module is used to capture the long-term dependencies in the time series. The formula for updating the hidden state of the gated recurrent unit is:

[0118] H gru = GRU(H, H prev );

[0119] Where H gru represents the output of the GRU module, H is the hidden state of the GRU module at the current moment, and H prev is the hidden state of the GRU module at the previous moment;

[0120] Finally, by concatenating the output of the GRU module with the temporal feature X tcn and then inputting it into a fully connected output layer (MLP) to generate the prediction result That is:

[0121]

[0122] Where The predicted value representing the future time step; is the concatenation symbol.

[0123] In addition, the training objective of the present invention is to optimize the model by minimizing the loss function. The loss function includes two parts: the case number prediction error and the propagation network alignment error. The specific formula is as follows:

[0124]

[0125] Among them, is the mean squared error of the case number prediction; is the mean squared error of the propagation network alignment; λ is a hyperparameter used to balance the importance of the two parts of the loss; M is the content that can be provided by datasets such as population flow information and geographical connectivity information, which can assist in the generation of the infectious disease propagation network to a certain extent; Y represents the true value corresponding to the prediction result in the dataset; is the adjacency matrix generated by the above dynamic spatio-temporal graph learning module, reflecting the real infectious disease propagation network.

[0126] Through the above technical solutions, the present invention can effectively capture the spatio-temporal dynamic characteristics of infectious disease transmission and still achieve accurate epidemic prediction in the case of lack of complete mobility data.

[0127] It should be noted that in the present invention, it should be understood that although this specification is described according to the embodiments, not each embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An infectious disease transmission prediction method based on a dynamic spatio-temporal graph neural network, characterized in that It includes the following steps: Extract spatio-temporal features from historical case data through a dynamic spatio-temporal graph learning module to generate an adjacency matrix representing the infectious disease transmission network; Predict the future epidemic development trend based on the adjacency matrix through an epidemiology dynamic modeling module; Among them, the dynamic spatio-temporal graph learning module includes a temporal convolutional network, a multi-head self-attention mechanism, and a long short-term memory network, which are used to capture the spatio-temporal dependence relationships between regions; The epidemiology dynamic modeling module includes a temporal convolutional network, a graph neural network, and a gated recurrent unit, which are used to predict the epidemic trend by combining virtual nodes.

2. The infectious disease transmission prediction method based on a dynamic spatio-temporal graph neural network according to claim 1, wherein The specific steps for the dynamic spatio-temporal graph learning module to generate an adjacency matrix include: Use a temporal convolutional network to extract temporal features from the input data to obtain a temporal feature matrix; Calculate the attention weights between regions through a multi-head self-attention mechanism to generate a weighted feature representation; Process the weighted feature representation through a long short-term memory network to generate an adjacency matrix representing the propagation relationships between regions.

3. The infectious disease transmission prediction method based on a dynamic spatio-temporal graph neural network according to claim 2, wherein The calculation of the multi-head self-attention mechanism includes: Calculate the query matrix Q = ZW Q and the key matrix K = ZW K and the value matrix V = ZW V , where Q represents the query matrix; K represents the key matrix; V represents the value matrix; Z represents the input feature; W Q represents the weight of the query matrix; W K represents the weight of the key matrix; W V represents the weight of the value matrix; Calculate the attention weights through the attention mechanism: where att represents the calculated attention weights; K T represents the transpose of the key matrix K; d k represents the dimension of the key vector, which is used to scale the dot product to prevent the vanishing gradient; softmax represents the Softmax function, which converts the calculated attention scores into a probability distribution; Apply the attention weights to the value matrix to obtain a weighted feature representation: Z att = att · V.

4. The infectious disease transmission prediction method based on a dynamic spatio-temporal graph neural network according to claim 1, wherein The epidemiology dynamic modeling module introduces virtual nodes, and its generation method includes: Perform a pooling operation on the temporal feature matrix to aggregate the information of all nodes; Extract virtual node features from the pooled features through a multi-layer perceptron; Add the pooled features to the output of the multi-layer perceptron to generate the virtual node feature representation x Λ , and the specific formula is as follows: x Λ = f init (Pooling(X tcn )) + Pooling(X tcn ); Among them, Pooling(·) represents the average pooling operation, which is used to aggregate the information of all nodes; f init is a function that extracts useful information from the pooled features through an MLP, generating the feature representation of the virtual node.

5. The method for predicting the spread of infectious diseases based on the dynamic spatio-temporal graph neural network according to claim 1, characterized in that, The graph neural network updates the node features through multi-layer graph convolution operations, and the specific formula is as follows: Among them, represents the updated feature of node i at time t, that is, the hidden state; W represents the learnable weight matrix; represents the neighbor set of node i in the k-th layer; represents the normalized attention weight from node i to its neighbor node j in the k-th layer at time t; represents the temporal feature of node j extracted from the input data at time t; represents the normalized attention weight from node i to the virtual node Λ at time t; represents the feature of the virtual node at time t; Subsequently, the gated recurrent unit module is used to capture the long-term dependence relationships in the time series, and the hidden state update formula of the gated recurrent unit is: H gru = GRU(H, H prev ); Among them, H gru represents the output of the GRU module. H is the hidden state of the GRU module at the current moment, and H prev is the hidden state of the GRU module at the previous moment; Finally, by concatenating the output of the GRU module with the temporal feature X tcn and then inputting it into a multi-layer perceptron to generate the prediction result That is: Among them, represents the predicted value at a future time step; is a concatenation symbol.

6. The method for predicting the spread of infectious diseases based on the dynamic spatio-temporal graph neural network according to claim 1, wherein The method further includes optimizing the model through a loss function, and the loss function includes the case number prediction error and the propagation network alignment error, and the specific formula is as follows: Among them, is the mean squared error of the case number prediction; is the mean squared error of the propagation network alignment; λ is a hyperparameter used to balance the importance of the two parts of the loss; M is the content that can be provided by datasets such as population mobility information and geographical connectivity information, which can assist in the generation of the infectious disease transmission network to a certain extent; Y represents the true value corresponding to the prediction result in the dataset; is the adjacency matrix generated by the above dynamic spatio-temporal graph learning module, reflecting the real infectious disease transmission network.

7. An infectious disease transmission prediction system based on a dynamic spatio-temporal graph neural network, characterized in that, It includes: A dynamic spatio-temporal graph learning module, which is used to extract spatio-temporal features from historical case data and generate an adjacency matrix; An epidemiology dynamic modeling module, which is used to predict the future epidemic development trend based on the adjacency matrix; Among them, the structures of the dynamic spatio-temporal graph learning module and the epidemiology dynamic modeling module are the same as those of the corresponding modules in any one of claims 1-6.

8. The infectious disease transmission prediction system based on the dynamic spatio-temporal graph neural network according to claim 7, wherein The dynamic spatio-temporal graph learning module includes a temporal convolutional network, a multi-head self-attention mechanism, and a long short-term memory network; The epidemiology dynamic modeling module includes a temporal convolutional network, a graph neural network, and a gated recurrent unit.

9. The infectious disease transmission prediction system based on the dynamic spatio-temporal graph neural network according to claim 7, wherein The system further includes a virtual node generation unit for generating a virtual node feature representation x Λ , and the specific formula is as follows: x Λ = f init (Pooling(X tcn )) + Pooling(X tcn ); Among them, Pooling(·) represents the average pooling operation, which is used to aggregate the information of all nodes; f init is a function that extracts useful information from the pooled features through an MLP to generate the feature representation of the virtual node.

10. The infectious disease transmission prediction system based on a dynamic spatio-temporal graph neural network according to claim 7, wherein The system further includes a loss function optimization unit, which is used to optimize the model through the case number prediction error and the propagation network alignment error, and the specific formula is as follows: Among them, is the mean squared error of case number prediction; is the mean squared error of propagation network alignment; λ is a hyperparameter used to balance the importance of the two parts of the loss; M is the content that can be provided by datasets such as population mobility information and geographical connectivity information, which can assist in the generation of the infectious disease propagation network to a certain extent; Y represents the true value corresponding to the prediction result in the dataset; is the adjacency matrix generated by the above dynamic spatio-temporal graph learning module, reflecting the real infectious disease propagation network.

Images