Extraction of Transitional Relations in Healthcare Processes from Chinese Medical Text based on Deep Learning

Abstract

Knowledge extraction in healthcare has become a recent research hotspot due to the rapid development of AI technology as well as the urgent demand in the medicine research and practice. Traditional methods focus on extracting terms and relations using pattern matching and machine learning, but they depend on the construction of rules and the accuracy of feature extraction. In contrast, deep learning methods have a stronger ability to process data with high dimensions. Moreover, current study neglects the fact that the key knowledge in medicine is the process of diagnosing and treating patients, i.e., the relations that causes the status change in a healthcare workflow. This paper first identifies the relations relevant for medical processes, and adopts the method of relationship extraction based on Bidirection Gated Recurrent Unit (BiGRU) model and attention mechanism (BiGRU-ATT), in order to retrieve these relations from Chinese medical text. The experimental results show that regarding Chinese medical entity relationship extraction, we can achieve better accuracy rate and recall rate than using convolutional neural network(CNN).