Research On The Prediction Model Of Infectious Disease Epidemic Trend Based On Graph Neural Network

The prediction and analysis of the epidemic trend of infectious diseases,as well as the prediction of the trend of population flow,are important contents of the prevention and control of infectious diseases.Although there have been research efforts regarding the trends of infectious diseases,they still have some deficiencies,such as only being able to support static predictions of the number of infections,merely considering symptoms of infections and their transmission trajectories,lacking analysis of infectious disease trends from a graph-based perspective to provide effective strategies to alter these trends,and suffering large errors in predicting population flow trends due to data scarcity.To solve the problems,the thesis studies the infectious disease epidemic trend prediction model based on graph neural network,including: proposing a multi-factor collaborative prediction model based on graph neural network that takes into account infection features and transmission trajectories,enabling dynamic predictions of infection numbers;introducing a key node search model based on graph neural network and deep reinforcement learning,training an intelligent agent to efficiently search for individuals with strong possibility of infection in the transmission network;presenting a population flow trend prediction model based on graph attention network that makes up for the lack of data through reinforcement learning method and predicts of population flow trend through graph attention network and long and short memory neural network.The specific details are as follows:(1)To address the challenge of dynamically predicting trends in infectious disease outbreaks,a multi-factor collaborative prediction model based on graph neural network,referred to as the trace-aware Graph Convolutional Network(TGCN),is proposed.TGCN first considers the track,physiological,and pathological features of infected individuals and creates a multi-factor co-propagation network;then TGCN uses graph neural network to complete the node classification task on the multi-factor co-propagation network.Furthermore,the model is integrated with the Susceptible-Exposed-Infected-Removed(SEIR)model to achieve dynamic prediction of the number of infections.Experimental results conducted on COVID-19 infection datasets from four different countries demonstrate that TGCN’s predictions of infection numbers align closely with real statistical data,exhibiting the smallest prediction errors among the tested methods.(2)To address the analysis of infectious disease trends and the problem of efficiently searching for individuals with strong infection possibility,a key node search model based on graph neural networks and deep reinforcement learning,named Vital Node Searcher(VNS),is proposed.Key nodes are typically treated as nodes in a network that exert a decisive impact on network performance.By adding or removing key nodes,one can achieve changes in network performance with minimal cost.Individuals with strong infection possibilities can be considered key nodes.Removing individuals with high infection possibilities from the transmission network maximizes the reduction in infection spread risk.VNS first employs the graph embedding method to acquire the embedding vector representations of the target network and then generates graphs as training data.A deep reinforcement learning algorithm is employed to enhance the agent’s ability to search for key nodes.Experimental results conducted on both generated graph datasets and real-world datasets demonstrate that VNS achieves faster and more efficient searching for individuals with a strong infection possibility.(3)To predict population flow,a population trend prediction model based on graph attention networks,called the Dynamic Attention Network(DAT),is proposed.Mastering the flow of populations during a specific time period helps determine the flow direction of patients and infected people.DAT first employs a reinforcement learning method to generate dynamic graphs,solving the problem of data scarcity.DAT then extracts spatiotemporal features of the target network by combining graph attention networks with long-shortterm memory neural networks,thus achieving the prediction of population mobility trends.Experimental results indicate that DAT can overcome data scarcity and achieve more accurate predictions of population flow trends.In conclusion,the thesis proposes an approach that predicts the epidemic trend of infectious diseases,then studies how to search for individuals with a high possibility of transmission,and finally designs an prediction approach to population flow trends.Experiments verified the effectiveness of the proposed models.