Research on Intelligent Diagnosis Model of Electronic Medical Record Based on Graph Transformer

Intelligent diagnosis based on the Electronic Medical Record (EMR) is an important research content of biomedical informatics. Recent studies have showed that graph representation learning methods can improve the performance of diagnosis prediction tasks by utilizing the graph structural information of EMR. However, the original EMR data often do not contain such complete and explicit structural information. It is challenging to learn EMR structural information from EMR data. In this paper, we propose a Statistics and Knowledge based Graph Transformer (S_K_GT) model to jointly learn EMR graph and perform diagnosis prediction tasks. Specifically, we firstly use conditional probability to initially construct the adjacency matrix of EMR graph and send it to Transformer encoder to further learn the connection relationship between EMR nodes. Then, we construct a knowledge attention network based on medical knowledge graph as expert knowledge to optimize the EMR graph deep learning process. Finally, we learn precise EMR graph representation for intelligent diagnosis. Experiments on EMR data in both Chinese and English show that the proposed model outperforms the state-of-the-art methods in all diagnosis prediction tasks.