Research On Clinical Data Classification And Diagnosis Based On Graph Convolution Neural Network Algorithm

With the increasing popularity of artificial intelligence technology in auxiliary medicine,many clinical problems have been better solved.Among them,clinical disease diagnosis is a very important task in clinical medicine.The early diagnosis method based on doctors’ clinical experience has a high misdiagnosis rate and missed diagnosis rate,which leads to patients bearing high treatment costs and unnecessary pain.Therefore,it is a good solution to use artificial intelligence technology to assist medical disease diagnosis.Artificial intelligence has made great breakthroughs in the classification of some diseases.However,the classification of diseases with high similarity in clinical and histological characteristics has not been well resolved,such as the differentiation of Crohn’s disease and intestinal tuberculosis.Therefore,there are still many challenges in the application of artificial intelligence in clinical disease classification and diagnosis.Aiming at the challenges described above,this thesis first proposes a graph convolution network framework(RF-GCB)based on random forest graph generation algorithm to solve this problem.RF-GCN uses a graph generation algorithm based on random forest to convert structured data into graph data.The algorithm considers the potential correlation between samples,and establishes an effective classification model through the feature representation of graph convolution learning nodes.Finally,comparing the performance of six benchmark models on four medical data sets,the experimental results show that the proposed algorithm model has better classification performance and can improve the performance of the model by 5%-20%.Finally,this thesis further improves the RF-GCN framework and proposes a graph attention network(HR-GCN)with multi-level graph generation algorithm.HR-GCN supports the generation of multi-level graph data,and adopts graph attention network to consider the contribution degree of neighbor nodes to represent learning.Finally,through comparative experiments on four base map neural network model frameworks and four medical data sets,the experimental results show that the algorithm model of HR-GCN proposed in this thesis has an improvement of 2%-5% under the RF-GCN algorithm model.