Multi-disease association propagation situation prediction method and system based on graph network

Abstract

The application provides a multi-disease correlation propagation situation prediction method and system based on a graph network, relates to the technical field of disease propagation situation prediction, and comprises the following steps: collecting multi-source monitoring data and constructing a ternary heterogeneous graph; a graph neural network prediction model is constructed; the graph neural network prediction model is trained by using training data, so that a trained graph neural network prediction model is obtained; the ternary heterogeneous graph to be predicted is input into the trained graph neural network prediction model, so that a situation prediction result of each disease node at a future time step is obtained. The application realizes accurate prediction of the multi-disease collaborative propagation effect by constructing a disease-population-environment ternary heterogeneous graph network and adopting dynamic weight modulation based on a popular situation imbalance for the comorbidity edge.

Description

Technical Field

[0001] This invention relates to the field of disease transmission trend prediction technology, and in particular to a method and system for predicting the transmission trend of multiple diseases based on graph networks. Background Technology

[0002] Predicting disease transmission trends is a core task of public health management. Accurately predicting disease development trends is crucial for formulating effective prevention and control strategies and optimizing the allocation of medical resources. Especially in complex scenarios where multiple infectious and chronic diseases coexist, the epidemiological links between different diseases pose even greater challenges to transmission trend prediction.

[0003] With the development of graph neural network technology, some researchers have attempted to introduce graph structures into the field of disease prediction. Existing methods mainly focus on the static relationships between diseases, using graph networks with fixed weights to model comorbidity among multiple diseases. However, these methods have significant limitations in dealing with the problem of collaborative transmission of multiple diseases, lacking consideration of the impact of dynamic changes in the epidemic situation on the comorbidity effect, and making it difficult to accurately capture the interaction patterns of different diseases at different stages of the epidemic.

[0004] Chinese patent CN113593716B discloses a method for predicting infectious disease trends. This method addresses the issue of the classic SEIR model not considering the infectious characteristics of asymptomatic incubation periods by adding the isolation of asymptomatic infected individuals to the SEIR model. It constructs a dynamic model encompassing five population groups: susceptible individuals, exposed individuals, isolated exposed individuals, infected individuals, and recovered individuals. The model uses a system of differential equations to describe population state transitions and employs the Monte Carlo method to optimize and fit the model parameters, thereby predicting the number of confirmed infections and susceptible individuals in the future. However, it primarily focuses on predicting chronic disease comorbidities in individual patients, neglecting the dynamic characteristics of disease transmission at the population level and failing to dynamically modulate the weights of comorbidity edges to reflect real-time changes in the prevalence intensity of different diseases. Therefore, it is difficult to apply to scenarios involving the prediction of multi-disease-related transmission trends. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for predicting the spread of multiple diseases based on graph networks. It solves the technical problem that the prediction of the spread of multiple diseases cannot accurately depict the impact of the dynamic changes in the epidemic situation on the comorbidity effect in the existing technology. By constructing a disease-population-environment ternary heterogeneous graph network, dynamic weight modulation based on the imbalance of the epidemic situation is applied to the comorbidity edge, and a cross-disease interaction modeling mechanism driven by the change in the spread situation is introduced to achieve accurate prediction of the synergistic spread effect of multiple diseases.

[0006] The technical solution of this invention is implemented as follows: On the one hand, this invention provides a method for predicting the spread of multiple diseases based on graph networks, including: Collect multi-source monitoring data and construct a ternary heterogeneous graph. The ternary heterogeneous graph includes a set of nodes and a set of edges connecting the nodes. The set of nodes includes a set of disease nodes, a set of population nodes, and a set of environment nodes. The set of edges includes transmission chain edges, comorbidity edges, susceptibility edges, social edges, and environmental impact edges. A graph neural network prediction model is constructed, which includes a graph convolutional layer, a temporal coding layer, a cross-disease interaction layer, a spatial propagation layer, and a prediction output layer. The graph neural network prediction model is trained using training data to obtain a trained graph neural network prediction model; The ternary heterogeneous graph to be predicted is input into a trained graph neural network prediction model to obtain the situation prediction results of each disease node in the future time step.

[0007] Based on the above technical solutions, preferably, the training of the graph neural network prediction model specifically includes: A1. The co-infected edges are weighted using a dynamic weight modulation mechanism to obtain time-varying co-infected edge weights; A2. Use graph convolutional layers to extract features from the ternary heterogeneous graph, and perform aggregation operations on the features of adjacent nodes of each node according to edge type to obtain the embedding representation of each node. A3. A temporal coding layer is used to encode the embedding representation of disease nodes at multiple historical time steps to obtain the temporal-aware feature vector of disease nodes. A4. A cross-disease interaction layer is used to model the cross-disease interaction of disease nodes based on the temporal perception feature vectors of disease nodes. The interaction weights between disease node pairs are modulated according to the propagation momentum of comorbid disease nodes to obtain the cross-disease interaction enhancement feature vectors of disease nodes. A5. Spatial propagation layer is used to model the spatial propagation of the cross-disease interaction enhancement feature vector of the disease node, and the spatial propagation enhancement feature vector of the disease node is obtained. A6. The prediction output layer is used to map the spatial propagation enhancement feature vector of the disease node to obtain the disease situation prediction value. The loss function is constructed based on the error between the prediction value and the true label and then optimized and trained.

[0008] Based on the above technical solutions, preferably, step A1 specifically includes: The initial weights of the comorbidity edges are calculated based on the proportion of comorbidity in the comorbidity disease node pairs statistically analyzed in the epidemiological database and the multiplier of the severity of the comorbidity disease node pairs recorded in the clinical data. The prevalence data of comorbid disease node pairs at each time point are extracted from multi-source monitoring data. The higher prevalence value of the comorbid disease node pair is calculated as the dominant trend strength. The ratio of the absolute difference in prevalence of two comorbid disease nodes to the sum of prevalence is calculated as the trend imbalance. Dynamic modulation coefficients are constructed based on the strength of the dominant situation and the degree of situational imbalance. The initial weights are then scaled and modulated using these dynamic modulation coefficients to obtain the co-problem edge weights that change dynamically over time.

[0009] Based on the above technical solutions, preferably, step A2 specifically includes: The ternary heterogeneous graph is decomposed into multiple homogeneous subgraphs according to the edge type, and each homogeneous subgraph corresponds to one edge type; Perform graph convolution operation on each homogeneous subgraph, and aggregate the features of its neighboring nodes for each node to obtain a node representation based on the edge type; By fusing node representations based on different edge types, a node embedding representation is obtained.

[0010] Based on the above technical solutions, preferably, step A3 specifically includes: The embedding representations of disease nodes in continuous historical time steps are arranged in chronological order to form a time-series vector sequence; A bidirectional long short-term memory network is used to encode the temporal vector sequence to obtain the forward hidden state and the backward hidden state. The forward hidden state and the backward hidden state are concatenated to obtain the bidirectional temporally dependent disease node representation. A temporal attention mechanism is used to calculate attention weights for the bidirectional temporal dependent disease node representations at multiple historical time steps. Based on the attention weights, the disease node representations at multiple historical time steps are weighted and aggregated to obtain the temporal-aware feature vector of the disease node at the current time.

[0011] Based on the above technical solutions, preferably, step A4 specifically includes: Extract all disease nodes and all co-disease edges from the ternary heterogeneous graph to form a co-disease subgraph; For disease node pairs connected by co-disease edges in the co-disease subgraph, a graph attention mechanism is used to calculate the interaction impact score between disease nodes. The difference between the temporal sensing feature vector of each disease node at the current moment and the temporal sensing feature vector at the previous moment is calculated as the change in the transmission trend. For the changes in the propagation status of two disease nodes connected by a co-disease edge, a propagation status modulation coefficient is constructed. The interaction influence score is modulated using the propagation status modulation coefficient and then normalized to obtain the interaction influence weight between disease nodes. Based on the interaction influence weight, the temporal-aware feature vectors of disease nodes connected by co-disease edges are weighted and aggregated to update, resulting in cross-disease interaction enhancement feature vectors of disease nodes.

[0012] Based on the above technical solutions, the preferred method for calculating the interaction influence score is as follows: ; in, Represents disease nodes and disease nodes The interaction between them affects the score. This represents a vector of learnable attention parameters. This indicates a query for the transformation matrix. Represents the key transformation matrix. Represents disease nodes The time-aware feature vector at the current moment, Represents disease nodes The time-aware feature vector at the current moment, This represents the dynamic weight of the comorbidity edge. This represents a vector concatenation operation. This represents the activation function. Indicates the sensitivity coefficient to the spread of the virus. Represents disease nodes At the present moment Compared to the previous moment The characteristic represents the change in the propagation trend. Represents disease nodes The amount of change in the spread trend, It represents the square of the Euclidean distance between the two changes in propagation status.

[0013] Based on the above technical solutions, preferably, step A5 specifically includes: The cross-disease interaction enhancement feature vector of the disease node is fused with the embedded representation of the corresponding population node in different geographical regions to construct a disease-region feature matrix. Calculate the spatial proximity between geographic regions, perform graph convolution operation on spatially adjacent region nodes, and obtain a region feature representation that integrates spatial proximity information; Based on population flow data, flow connection edges between regions are constructed. An attention mechanism is used to calculate the weight of the flow connection. The flow-related regional node features are weighted and aggregated to obtain a regional feature representation that integrates spatial proximity information and population flow information. The regional feature representations that integrate spatial proximity information and population flow information are aggregated back to the corresponding disease nodes and then fused with the cross-disease interaction enhancement feature vectors of the disease nodes to obtain the spatial propagation enhancement feature vectors of the disease nodes.

[0014] More preferably, the loss function includes a prediction loss term, a comorbidity consistency loss term, and a temporal smoothing constraint term, calculated as follows: ; ; ; in, Denotes the joint loss function. This represents the predicted loss term. This indicates the loss of comorbidity. Represents the time-series smoothing constraint term. This represents the weighting coefficient of the comorbidity consistency loss term. This represents the loss weight coefficient of the time series smoothing constraint term. Let this be the set of all edges sharing the same disease. Represents disease nodes and disease nodes Dynamic weights of the comorbidity sides between them Represents disease nodes The situation prediction sequence, Represents disease nodes The situation prediction sequence, Represents Euclidean distance. Represents disease nodes At any moment The time-aware feature vector, Represents a set of disease nodes. Indicates the number of historical time steps.

[0015] In addition, this invention also provides a multi-disease association transmission trend prediction system based on graph networks to realize the above-mentioned multi-disease association transmission trend prediction method based on graph networks, including: The heterogeneous graph construction module is used to collect multi-source monitoring data and construct a ternary heterogeneous graph. The ternary heterogeneous graph includes a set of nodes and a set of edges connecting the nodes. The set of nodes includes a set of disease nodes, a set of population nodes, and a set of environment nodes. The set of edges includes transmission chain edges, comorbidity edges, susceptibility edges, social edges, and environmental influence edges. The model building module is used to build a graph neural network prediction model, which includes a graph convolutional layer, a temporal coding layer, a cross-disease interaction layer, a spatial propagation layer, and a prediction output layer. The model training module is used to train the graph neural network prediction model using training data to obtain a trained graph neural network prediction model. The situation prediction module is used to input the ternary heterogeneous graph to be predicted into the trained graph neural network prediction model to obtain the situation prediction results of each disease node in the future time step.

[0016] The present invention has the following advantages over the prior art: (1) By constructing a disease-population-environment ternary heterogeneous graph, disease nodes, population nodes and environment nodes are incorporated into a unified graph network framework. Five types of heterogeneous edges are defined: transmission chain edge, comorbidity edge, susceptibility edge, social edge and environmental influence edge to characterize multi-dimensional relationships. This can comprehensively model the complex coupling relationship between disease transmission chain, population susceptibility differences, social network structure and environmental factors in the multi-disease transmission system, thereby improving the ability of the prediction model to represent the real transmission network. (2) By using dynamic weight modulation based on the imbalance of the epidemic situation on the comorbidity edge, the weight of the comorbidity edge can be automatically amplified to reflect the nonlinear effect of high-incidence diseases on patients with low-incidence diseases, thereby improving the modeling accuracy of complex scenarios such as the increased risk of comorbidity in patients with chronic respiratory diseases during the influenza outbreak. (3) By calculating the change in the characteristic representation of disease nodes at adjacent times as the change in the propagation trend, and constructing modulation coefficients based on comorbid disease nodes, the interaction influence scores between diseases are dynamically modulated, accurately capturing the synergistic resonance effect and propagation decoupling effect between diseases, and improving the prediction accuracy of multi-disease synergistic propagation patterns. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart of the multi-disease association transmission trend prediction method based on graph networks of the present invention; Figure 2 This is a schematic diagram of the ternary heterogeneous graph of the multi-disease association transmission trend prediction method based on graph networks of the present invention; Figure 3 This is a diagram of the graph neural network prediction model for the multi-disease association and transmission trend prediction method based on graph networks of the present invention. Detailed Implementation

[0019] 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.

[0020] like Figure 1 and Figure 2 As shown, this invention provides a method for predicting the spread of multiple diseases based on graph networks, including: Collect multi-source monitoring data and construct a ternary heterogeneous diagram The ternary heterogeneous diagram Includes node set The set of edges connecting nodes , where the set of nodes Includes disease node set Crowd Node Set and environment node set ,Right now edge set Including the edge of the transmission chain comorbidity Susceptible edges Social side and environmental impact ; A graph neural network prediction model is constructed, which includes a graph convolutional layer, a temporal coding layer, a cross-disease interaction layer, a spatial propagation layer, and a prediction output layer. The graph neural network prediction model is trained using training data to obtain a trained graph neural network prediction model; The ternary heterogeneous graph to be predicted is input into a trained graph neural network prediction model to obtain the situation prediction results of each disease node in the future time step.

[0021] Understandable, disease node set Include Each disease node represents a specific infectious or chronic disease requiring monitoring, such as seasonal influenza, pneumonia caused by COVID-19, chronic obstructive pulmonary disease, or asthma; the population node set... Include Individual population nodes, each node representing a population cluster divided by geographical region or demographic characteristics, such as a street community population divided by administrative district, or an elderly population or a children population divided by age group; environmental node set. Include There are several environmental nodes, each representing a category of environmental factors that influence disease transmission, such as meteorological conditions nodes and medical resource nodes.

[0022] propagation chain edge Connecting disease nodes and population nodes represents the infection pressure and transmission intensity of a specific disease on a specific population; comorbidity edges Connecting different disease nodes represents the comorbidity and mutual influence among multiple diseases; susceptibility edges Connecting population nodes and disease nodes represents the susceptibility and infection risk of a specific population to a specific disease; social edge Connecting different population nodes represents the contact network and population flow relationships between populations; environmental impact edges Connecting environmental nodes and disease nodes represents the regulatory role and influence mechanism of environmental factors on disease transmission.

[0023] This invention constructs a disease-population-environment ternary heterogeneous graph, incorporating disease nodes, population nodes, and environment nodes into a unified graph network framework. It defines five types of heterogeneous edges—transmission chain edges, comorbidity edges, susceptibility edges, social edges, and environmental influence edges—to characterize multi-dimensional relationships. This overcomes the limitations of existing technologies that only consider disease-patient bipartite graphs, and can comprehensively model the complex coupling relationships of disease transmission chains, population susceptibility differences, social network structures, and environmental factors in multi-disease transmission systems, thereby improving the predictive model's ability to represent real transmission networks.

[0024] In one embodiment of the present invention, for disease nodes and crowd nodes The propagation chain between The formula for calculating its weight is: ; in, Indicates the edge of the propagation chain The weight, Indicates disease The basic reproduction number, Represents a crowd The proportion of susceptible population in the population, Indicates disease With the crowd The contact rate coefficient between them.

[0025] In one embodiment of the present invention, for the crowd node and disease nodes The susceptible edge between The formula for calculating its weight is: ; in, Indicates susceptible edge The weight, This represents the comprehensive susceptibility assessment function. Represents a crowd The age structure feature vector, This represents the vector of prevalence of chronic underlying diseases in this population. This indicates the average level of immunity in this population. This indicates that the population is targeting the disease. Vaccination coverage rate.

[0026] In one embodiment of the present invention, for the crowd node and Social borders between The formula for calculating its weight is: ; in, Indicates social edge The weight, Indicates from the crowd Flowing to the crowd Number of people Indicates the number of people moving in the opposite direction. and These represent the total population of the two groups, This represents the contact propagation efficiency coefficient.

[0027] In one embodiment of the present invention, for environmental nodes and disease nodes Environmental impact between The formula for calculating its weight is: ; in, Indicates the environmental impact edge The weight, Indicates environmental factors For disease The baseline coefficient of influence The sensitivity coefficient represents the impact of changes in environmental factors on a disease. Represents environment nodes The current eigenvalue.

[0028] like Figure 3 As shown, in one embodiment of the present invention, the training of the graph neural network prediction model specifically includes: A1. The co-infected edges are weighted using a dynamic weight modulation mechanism to obtain time-varying co-infected edge weights; A2. Use graph convolutional layers to extract features from the ternary heterogeneous graph, and perform aggregation operations on the features of adjacent nodes of each node according to edge type to obtain the embedding representation of each node. A3. A temporal coding layer is used to encode the embedding representation of disease nodes at multiple historical time steps to obtain the temporal-aware feature vector of disease nodes. A4. A cross-disease interaction layer is used to model the cross-disease interaction of disease nodes based on the temporal perception feature vectors of disease nodes. The interaction weights between disease node pairs are modulated according to the propagation momentum of comorbid disease nodes to obtain the cross-disease interaction enhancement feature vectors of disease nodes. A5. Spatial propagation layer is used to model the spatial propagation of the cross-disease interaction enhancement feature vector of the disease node, and the spatial propagation enhancement feature vector of the disease node is obtained. A6. The prediction output layer is used to map the spatial propagation enhancement feature vector of the disease node to obtain the disease situation prediction value. The loss function is constructed based on the error between the prediction value and the true label and then optimized and trained.

[0029] In one embodiment of the present invention, step A1 specifically includes: The initial weights of the comorbidity edges are calculated based on the proportion of comorbidity in the comorbidity disease node pairs statistically analyzed in the epidemiological database and the multiplier of the severity of the comorbidity disease node pairs recorded in the clinical data. The prevalence data of comorbid disease node pairs at each time point are extracted from multi-source monitoring data. The higher prevalence value of the comorbid disease node pair is calculated as the dominant trend strength. The ratio of the absolute difference in prevalence of two comorbid disease nodes to the sum of prevalence is calculated as the trend imbalance. Based on the dominant situation strength and situation imbalance, dynamic modulation coefficients are constructed. The initial weights are then scaled and modulated using these dynamic modulation coefficients to obtain the co-problem edge weights that dynamically change over time. in, Represents disease nodes and disease nodes The comorbidity between them is always present Crowd nodes Dynamic weights at the location, This represents the proportion of the comorbid population for disease nodes obtained from epidemiological databases. This indicates the factor by which comorbidity disease nodes recorded from clinical data affect the severity of the disease. This indicates disease nodes obtained from multi-source surveillance data. At any moment Crowd nodes The prevalence of the disease at the site This indicates disease nodes obtained from multi-source surveillance data. At any moment Crowd nodes The prevalence of the disease at the site The sensitivity coefficient to imbalance. It is a numerically stable term.

[0030] Understandably, the dynamic modulation term in the formula is composed of the product of two factors, the first factor being... This represents the prevalence intensity of the dominant disease, taking the one with the higher prevalence among two diseases. When the prevalence of both diseases is very low, this factor is close to zero, resulting in a weak dynamic modulation effect. When at least one disease is in a high-incidence period, this factor is large, significantly activating dynamic modulation. The second factor... The imbalance degree is represented by the degree of epidemic imbalance. When the prevalence of two diseases is equal, the imbalance degree is zero, indicating a perfectly balanced state where comorbidity weights are not amplified. When the prevalence of one disease is much higher than that of the other, the imbalance degree approaches one, indicating a highly imbalanced state that triggers significant weight amplification. Dynamic weights, through the coupled design of dominant strength and imbalance degree, capture the asymmetry of disease epidemic trends on the amplification effect of comorbidity. This accurately reflects the nonlinear enhancement of comorbidity risk in patients with chronic respiratory diseases during influenza outbreaks, thereby improving the accuracy of multi-disease transmission trend prediction.

[0031] This invention employs dynamic weight modulation of comorbidity edges based on the imbalance of epidemic trends. By using the prevalence data of comorbidity disease nodes at various times and coupling the dominant disease epidemic intensity with the degree of epidemic imbalance, the weight of comorbidity edges is dynamically adjusted according to the epidemic trend. This overcomes the problem of existing technologies using static comorbidity edge weights, which cannot reflect the amplification effect of the imbalance of disease prevalence on comorbidity. When one disease is in a high-incidence period and another disease is in a low-incidence period, the weight of comorbidity edges can be automatically amplified to reflect the nonlinear impact of the high-incidence disease on patients with the low-incidence disease, thereby improving the modeling accuracy of complex scenarios such as the increased risk of comorbidity in patients with chronic respiratory diseases during influenza outbreaks.

[0032] In one embodiment of the present invention, step A2 specifically includes: The ternary heterogeneous graph is decomposed into multiple homogeneous subgraphs according to the edge type, and each homogeneous subgraph corresponds to one edge type; Perform graph convolution operation on each homogeneous subgraph, and aggregate the features of its neighboring nodes for each node to obtain a node representation based on the edge type; By fusing node representations based on different edge types, a node embedding representation is obtained.

[0033] In one embodiment of the present invention, step A3 specifically includes: The embedding representations of disease nodes in continuous historical time steps are arranged in chronological order to form a time-series vector sequence; A bidirectional long short-term memory network is used to encode the temporal vector sequence to obtain the forward hidden state and the backward hidden state. The forward hidden state and the backward hidden state are concatenated to obtain the bidirectional temporally dependent disease node representation. A temporal attention mechanism is used to calculate attention weights for the bidirectional temporal dependent disease node representations at multiple historical time steps. Based on the attention weights, the disease node representations at multiple historical time steps are weighted and aggregated to obtain the temporal-aware feature vector of the disease node at the current time.

[0034] In one embodiment of the present invention, step A4 specifically includes: Extract all disease nodes and all co-disease edges from the ternary heterogeneous graph to form a co-disease subgraph; For disease node pairs connected by co-disease edges in the co-disease subgraph, a graph attention mechanism is used to calculate the interaction impact score between disease nodes. The difference between the temporal sensing feature vector of each disease node at the current moment and the temporal sensing feature vector at the previous moment is calculated as the change in the transmission trend. For the changes in the propagation status of two disease nodes connected by a co-disease edge, a propagation status modulation coefficient is constructed. The interaction influence score is modulated using the propagation status modulation coefficient and then normalized to obtain the interaction influence weight between disease nodes. Based on the interaction influence weight, the temporal-aware feature vectors of disease nodes connected by co-disease edges are weighted and aggregated to update, resulting in cross-disease interaction enhancement feature vectors of disease nodes.

[0035] Specifically, the method for calculating the interaction influence score is as follows: in, Represents disease nodes and disease nodes The interaction between them affects the score. This represents a vector of learnable attention parameters. This indicates a query for the transformation matrix. Represents the key transformation matrix. Represents disease nodes The time-aware feature vector at the current moment, Represents disease nodes The time-series-aware feature vector at the current moment, This represents the dynamic weight of the comorbidity edge. This represents a vector concatenation operation. This represents the activation function. Indicates the sensitivity coefficient to the spread of the virus. Represents disease nodes At the present moment Compared to the previous moment The characteristic represents the change in the propagation trend. Represents disease nodes The amount of change in the spread trend, This represents the dot product between two changes in propagation status. It represents the square of the Euclidean distance between the two changes in propagation status.

[0036] The first part of the formula for calculating the interaction influence score employs an attention mechanism, mapping the representations of two disease nodes to the query and key spaces. It combines the weights of co-disease edges as prior information on association strength to learn the dynamic interaction patterns between diseases. The second part of the formula is the propagation momentum shift modulation term, where the first factor in the modulation term... The dot product of two propagation momentum vectors reflects the consistency in direction and intensity of disease propagation trends. A positive and large dot product indicates a high degree of consistency in the propagation momentum directions of the two diseases, with both showing significant fluctuations. This indicates a strong synergistic effect, amplifying the interaction effect score. Conversely, a negative dot product indicates opposite propagation momentum directions, weakening the interaction effect score. The second factor in the modulation term... Let be the decay function of the propagation momentum difference, where The squared Euclidean distance, representing the difference between two momentum vectors, measures the degree of difference in the disease transmission patterns along their evolutionary trajectories. When the transmission momentum vectors of the two diseases are very close, the squared distance approaches zero, and the attenuation function value approaches one, thus not suppressing the modulation effect. When the transmission momentum vectors differ greatly, the squared distance increases, the attenuation function value decreases, and the modulation effect is suppressed, avoiding spurious resonance amplification that occurs when the momentum dot product is positive but the actual evolutionary trajectories of the two diseases differ too much.

[0037] Comorbidity subgraph is defined as This subgraph contains only the set of disease nodes. and the set of common disease edges connecting them This clearly characterizes which diseases have a comorbidity relationship and the strength of that relationship. For any two disease nodes connected by a comorbidity edge in the comorbidity subgraph... and The formula for calculating the interaction influence weight is: ; in Indicates disease For disease The interaction influence weight, Indicates the relationship between disease and comorbidity in the comorbidity subgraph. There exists a set of all neighboring disease nodes connected by a co-disease edge. Indicates disease and The interaction between them affects the score.

[0038] This invention calculates the change in the feature representation of disease nodes at adjacent time points as the change in the propagation trend, and constructs modulation coefficients based on the change in the propagation trend of comorbid disease node pairs. This dynamically modulates the interaction influence scores between diseases, enabling the model to identify the directional consistency and intensity synergy of the propagation trends of multiple diseases. When two diseases are simultaneously in a rapid rise or rapid fall phase, the interaction influence is enhanced; when the propagation trends are opposite, the interaction influence is weakened. This accurately captures the synergistic resonance effect and propagation decoupling effect between diseases, solving the problem of existing technologies that only calculate interaction weights based on node representation similarity while ignoring the temporal characteristics of propagation dynamics, and improving the prediction accuracy of multi-disease synergistic propagation patterns.

[0039] In one embodiment of the present invention, after calculating the interaction influence weights, the disease node The representation vector is updated by aggregating the weighted representations of all co-disease neighboring diseases, and the update formula is: ; in This indicates a disease node after cross-disease interaction. The new representation vector, Represents the original representation vector. The value transformation matrix is ​​represented by the summation term, which represents the weighted contribution of all comorbid neighbor diseases.

[0040] This update mechanism ensures that the representation of each disease node not only includes its own transmission characteristics, but also incorporates information on the prevalence of other diseases that have a comorbid relationship with it.

[0041] In one embodiment of the present invention, step A5 specifically includes: The cross-disease interaction enhancement feature vector of the disease node is fused with the embedded representation of the corresponding population node in different geographical regions to construct a disease-region feature matrix. Calculate the spatial proximity between geographic regions, perform graph convolution operation on spatially adjacent region nodes, and obtain a region feature representation that integrates spatial proximity information; Based on population flow data, flow connection edges between regions are constructed. An attention mechanism is used to calculate the weight of the flow connection. The flow-related regional node features are weighted and aggregated to obtain a regional feature representation that integrates spatial proximity information and population flow information. The regional feature representations that integrate spatial proximity information and population flow information are aggregated back to the corresponding disease nodes and then fused with the cross-disease interaction enhancement feature vectors of the disease nodes to obtain the spatial propagation enhancement feature vectors of the disease nodes.

[0042] In one embodiment of the present invention, the loss function includes a prediction loss term, a comorbidity consistency loss term, and a temporal smoothing constraint term, and the calculation formula is as follows: ; ; ; in, Denotes the joint loss function. This represents the predicted loss term. This indicates the loss of comorbidity. Represents the time-series smoothing constraint term. This represents the weighting coefficient of the comorbidity consistency loss term. This represents the loss weight coefficient of the time series smoothing constraint term. Let this be the set of all edges sharing the same disease. Represents disease nodes and disease nodes Dynamic weights of the comorbidity sides between them Represents disease nodes The situation prediction sequence, Represents disease nodes The situation prediction sequence, Represents Euclidean distance. Represents disease nodes At any moment The time-aware feature vector, Represents a set of disease nodes. Indicates the number of historical time steps.

[0043] Understandably, the total loss function integrates the three loss terms through a weighted sum, with weighting coefficients... and The model was optimized using the validation set. The Adam optimizer was used for parameter updates, with a learning rate set to 0.001. A learning rate decay strategy was employed to gradually reduce the learning rate during training.

[0044] This invention maintains a statistical correlation between the prediction curves of comorbid disease pairs and the strength of the comorbidity relationship by constraining the prediction curves. It directly introduces the dynamic weights of the comorbidity edges into the loss function optimization process, so that the graph neural network prediction model focuses not only on the accuracy of individual disease predictions during training, but also on the coordination of prediction trends among multiple diseases. This enhances the model's ability to learn the overall pattern of multi-disease association and transmission, and improves the internal consistency and reliability of prediction results.

[0045] This invention also provides a multi-disease association transmission trend prediction system based on graph networks, used to implement the multi-disease association transmission trend prediction method based on graph networks as described above, including: The heterogeneous graph construction module is used to collect multi-source monitoring data and construct a ternary heterogeneous graph. The ternary heterogeneous graph includes a set of nodes and a set of edges connecting the nodes. The set of nodes includes a set of disease nodes, a set of population nodes, and a set of environment nodes. The set of edges includes transmission chain edges, comorbidity edges, susceptibility edges, social edges, and environmental influence edges. The model building module is used to build a graph neural network prediction model, which includes a graph convolutional layer, a temporal coding layer, a cross-disease interaction layer, a spatial propagation layer, and a prediction output layer. The model training module is used to train the graph neural network prediction model using training data to obtain a trained graph neural network prediction model. The situation prediction module is used to input the ternary heterogeneous graph to be predicted into the trained graph neural network prediction model to obtain the situation prediction results of each disease node in the future time step.

[0046] 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 the spread of multiple diseases based on graph networks, characterized in that, include: Collect multi-source monitoring data and construct a ternary heterogeneous graph. The ternary heterogeneous graph includes a set of nodes and a set of edges connecting the nodes. The set of nodes includes a set of disease nodes, a set of population nodes, and a set of environment nodes. The set of edges includes transmission chain edges, comorbidity edges, susceptibility edges, social edges, and environmental impact edges. A graph neural network prediction model is constructed, which includes a graph convolutional layer, a temporal coding layer, a cross-disease interaction layer, a spatial propagation layer, and a prediction output layer. The graph neural network prediction model is trained using training data to obtain a trained graph neural network prediction model; The ternary heterogeneous graph to be predicted is input into a trained graph neural network prediction model to obtain the situation prediction results of each disease node in the future time step.

2. The method for predicting the spread of multiple diseases based on graph networks as described in claim 1, characterized in that: The training of the graph neural network prediction model specifically includes: A1. The co-infected edges are weighted using a dynamic weight modulation mechanism to obtain time-varying co-infected edge weights; A2. Use graph convolutional layers to extract features from the ternary heterogeneous graph, and perform aggregation operations on the features of adjacent nodes of each node according to edge type to obtain the embedding representation of each node. A3. A temporal coding layer is used to encode the embedding representation of disease nodes at multiple historical time steps to obtain the temporal-aware feature vector of disease nodes. A4. A cross-disease interaction layer is used to model the cross-disease interaction of disease nodes based on the temporal perception feature vectors of disease nodes. The interaction weights between disease node pairs are modulated according to the propagation momentum of comorbid disease nodes to obtain the cross-disease interaction enhancement feature vectors of disease nodes. A5. Spatial propagation layer is used to model the spatial propagation of the cross-disease interaction enhancement feature vector of the disease node, and the spatial propagation enhancement feature vector of the disease node is obtained. A6. The prediction output layer is used to map the spatial propagation enhancement feature vector of the disease node to obtain the disease situation prediction value. The loss function is constructed based on the error between the prediction value and the true label and then optimized and trained.

3. The method for predicting the spread of multiple diseases based on graph networks as described in claim 2, characterized in that: Step A1 specifically includes: The initial weights of the comorbidity edges are calculated based on the proportion of comorbidity in the population of the comorbidity disease node pairs in the statistics and the multiplier of the severity of the comorbidity disease node pairs recorded in the clinical data. The prevalence data of comorbid disease node pairs at each time point are extracted from multi-source monitoring data. The prevalence value of the comorbid disease node pair is calculated as the dominant trend strength. The ratio of the absolute difference in prevalence of two comorbid disease nodes to the sum of prevalence is calculated as the trend imbalance degree. Dynamic modulation coefficients are constructed based on the strength of the dominant situation and the degree of situational imbalance. The initial weights are then scaled and modulated using these dynamic modulation coefficients to obtain the co-problem edge weights that change dynamically over time.

4. The method for predicting the spread of multiple diseases based on graph networks as described in claim 3, characterized in that: Step A2 specifically includes: The ternary heterogeneous graph is decomposed into multiple homogeneous subgraphs according to the edge type, and each homogeneous subgraph corresponds to one edge type; Perform graph convolution operation on each homogeneous subgraph, and aggregate the features of its neighboring nodes for each node to obtain a node representation based on the edge type; By fusing node representations based on different edge types, a node embedding representation is obtained.

5. The method for predicting the spread of multiple diseases based on graph networks as described in claim 4, characterized in that: Step A3 specifically includes: The embedding representations of disease nodes in continuous historical time steps are arranged in chronological order to form a time-series vector sequence; A bidirectional long short-term memory network is used to encode the temporal vector sequence to obtain the forward hidden state and the backward hidden state. The forward hidden state and the backward hidden state are concatenated to obtain the bidirectional temporally dependent disease node representation. A temporal attention mechanism is used to calculate attention weights for the bidirectional temporal dependent disease node representations at multiple historical time steps. Based on the attention weights, the disease node representations at multiple historical time steps are weighted and aggregated to obtain the temporal-aware feature vector of the disease node at the current time.

6. The method for predicting the spread of multiple diseases based on graph networks as described in claim 5, characterized in that: Step A4 specifically includes: Extract all disease nodes and all co-disease edges from the ternary heterogeneous graph to form a co-disease subgraph; For disease node pairs connected by co-disease edges in the co-disease subgraph, a graph attention mechanism is used to calculate the interaction impact score between disease nodes. The difference between the temporal sensing feature vector of each disease node at the current moment and the temporal sensing feature vector at the previous moment is calculated as the change in the transmission trend. For the changes in the propagation status of two disease nodes connected by a co-disease edge, a propagation status modulation coefficient is constructed. The interaction influence score is modulated using the propagation status modulation coefficient and then normalized to obtain the interaction influence weight between disease nodes. Based on the interaction influence weight, the temporal-aware feature vectors of disease nodes connected by co-disease edges are weighted and aggregated to update, resulting in cross-disease interaction enhancement feature vectors of disease nodes.

7. The method for predicting the spread of multiple diseases based on graph networks as described in claim 6, characterized in that: The method for calculating the interaction impact score is as follows: ； in, Represents disease nodes and disease nodes The interaction between them affects the score. This represents a vector of learnable attention parameters. This indicates a query transformation matrix. Represents the key transformation matrix. Represents disease nodes The time-aware feature vector at the current moment, Represents disease nodes The time-aware feature vector at the current moment, This represents the dynamic weight of the comorbidity edge. This represents a vector concatenation operation. This represents the activation function. Indicates the sensitivity coefficient to the propagation situation. Represents disease nodes At the present moment Compared to the previous moment The characteristic represents the change in the propagation trend. Representing disease nodes The amount of change in the spread trend, It represents the square of the Euclidean distance between the two changes in propagation status.

8. The method for predicting the spread of multiple diseases based on graph networks as described in claim 7, characterized in that: Step A5 specifically includes: The cross-disease interaction enhancement feature vector of the disease node is fused with the embedded representation of the corresponding population node in different geographical regions to construct a disease-region feature matrix. Calculate the spatial proximity between geographic regions, perform graph convolution operation on spatially adjacent region nodes, and obtain a region feature representation that integrates spatial proximity information; Based on population flow data, flow connection edges between regions are constructed. An attention mechanism is used to calculate the weight of the flow connection. The flow-related regional node features are weighted and aggregated to obtain a regional feature representation that integrates spatial proximity information and population flow information. The regional feature representations that integrate spatial proximity information and population flow information are aggregated back to the corresponding disease nodes and then fused with the cross-disease interaction enhancement feature vectors of the disease nodes to obtain the spatial propagation enhancement feature vectors of the disease nodes.

9. The method for predicting the spread of multiple diseases based on graph networks as described in claim 2, characterized in that: The loss function includes a prediction loss term, a comorbidity consistency loss term, and a temporal smoothing constraint term, and its calculation formula is as follows: ； ； ； in, Denotes the joint loss function. This represents the predicted loss term. This indicates the loss of comorbidity. Represents the time-series smoothing constraint term. This represents the weighting coefficient of the comorbidity consistency loss term. This represents the loss weight coefficient of the time series smoothing constraint term. Let this be the set of all edges sharing the same disease. Representing disease nodes and disease nodes Dynamic weights of the comorbidity sides between them Representing disease nodes The situation prediction sequence, Representing disease nodes The situation prediction sequence, Represents Euclidean distance. Representing disease nodes At any moment The time-aware feature vector, Represents a set of disease nodes. Indicates the number of historical time steps.

10. A multi-disease association transmission trend prediction system based on graph networks, characterized in that, The method for predicting the spread of multiple diseases based on graph networks as described in any one of claims 1-9 includes: The heterogeneous graph construction module is used to collect multi-source monitoring data and construct a ternary heterogeneous graph. The ternary heterogeneous graph includes a set of nodes and a set of edges connecting the nodes. The set of nodes includes a set of disease nodes, a set of population nodes, and a set of environment nodes. The set of edges includes transmission chain edges, comorbidity edges, susceptibility edges, social edges, and environmental influence edges. The model building module is used to build a graph neural network prediction model, which includes a graph convolutional layer, a temporal coding layer, a cross-disease interaction layer, a spatial propagation layer, and a prediction output layer. The model training module is used to train the graph neural network prediction model using training data to obtain a trained graph neural network prediction model. The situation prediction module is used to input the ternary heterogeneous graph to be predicted into the trained graph neural network prediction model to obtain the situation prediction results of each disease node in the future time step.

Images