Research On The Diagnosis Method Of Mild And Severe COVID-19 Based On Graph Representation Learnin

The Corona Virus Disease 2019(COVID-19)is a global pandemic,which has spread rapidly across the world and brought an enormous burden on the global healthcare system.The symptoms of COVID-19 patients develop fast and the severe symptoms quickly lead to a variety of other diseases,even to death.Hence,finding an antidote against the COVID-19 to reduce the clinicians’ workloads and designing an effective treatment plan is an urgent problem that needs to be solved.In this thesis our study focus on COVID-19 developing severe symptoms based on graph representation learning.First,we propose a novel structural attention graph neural network,named SAGNN,which can identify the severe COVID-19 cases from the mild cases.The SAGNN first aggregates the features in a given sample graph with hierarchical information inherent to the physical structure of lungs,and then uses the structural attention mechanism to effectively fuse the different features to obtain the final graph representation for classification tasks.The Focal-Loss is used to solve the issue of imbalanced group distribution,which further improve the overall network performance.Furthermore,we propose multi-task learning via structural attention graph neural networks,named MSAGNN,a joint classification and regression method to determine whether the patient would develop severe symptoms in the later time,and if yes,predict the possible conversion time that the patient would spend to convert to the severe stage.Based on SAGNN,MSAGNN can combine the lung structure,features extracted from the chest CT image and the patient non-imaging information to conduct a diagnosis of COVID-19 and predict the conversion time from a mild case to a severe case.To verify the effectiveness of the proposed methods,the experiments are on a COVID-19 dataset,including 1687 cases,confirmed by the Chinese Center for Disease Control and Prevention.In the single classification tasks,the specificity and the sensitivity of SAGNN is 91.3% and 72.7%,respectively,and the sensitivity is higher than all comparison methods in this thesis.Moreover,joint classification and regression,MSAGNN improves the performance of above tasks(classification)and reduces the RMSE by 0.16 ～ 1.11 compared with the comparison method in the time estimation task.In addition,the ablation experiments show the significance of the introduced parts,i.e.,graph structure,structural attention mechanism and demographic information.In summary,both SAGNN and MSAGNN have the best performance in the experiments.