Correlation-driven framework based on graph convolutional network for clinical disease classification

With the increasing popularity of computer-aided technology applied in medicine, great achievements have been made in certain diseases. However, due to the similarity of clinical and histological features, the problem of disease classification has not been well resolved, especially the Crohn's disease (CD) and intestinal tuberculosis (ITB). In this paper, a novel sample connection driven framework named RFG-GCN is presented to overcome this. Firstly, the RFG-GCN employs a random forest based graph generation algorithm (RFG) to convert structured data into graph data, which considers the correlation between samples. Under the training of two layer-wise graph convolution neural network, an efficient model is established. Our findings show that the classification performance has been significantly improved compared withother methods. Furthermore, to show the extensibility and consistency, we apply our framework to other four public medical datasets. The results indicate that the RFG-GCN framework can achieve competitive and better classification performance.