Authors: Ning Liu, Wei Zhang, Xiuxing Li, Haitao Yuan and Jianyong Wang
Abstract: Text-oriented clinical diagnosis inference is to predict a set of diagnoses for a specific patient given its medical notes. Due to the great potential of automatic diagnosis inference, machine learning methods have began to be applied to this domain. However, existing approaches focus on performing either labor-intensive feature engineering or sequential modeling of each medical note separately, without considering the information sharing among similar patients, which is essential for evidencebased medicine, an emerging new diagnosis process. Motivated by this issue and the recently proposed graph convolutional network (GCN) for text classification, we propose to apply GCN for the text-oriented clinical diagnosis inference task. To encode the comorbidity of diagnoses into the GCN model and allow information sharing between patients, we devise a coupled graph convolutional neural networks (CGCN), where a note-dependent graph and a label-dependent graph are learned collaboratively with hyperplane projection to ensure they are in the same semantic space. The comprehensive results on two real datasets show that our method outperforms the state-of-art methods in text-oriented diagnosis inference.