Automated ICD Coding Based on Neural Machine Translation

With the rapid development of the medical field, the electronic medical record (EMR) is a rich source of clinical information for medical study. However, electronic medical records are full of redundant information, which increases the difficulty of statistical analysis. The International Classification of Diseases (ICD) is widely used to describe the diagnosis of patients. A reliable automated ICD coding system can improve the quality of clinical decision support. Manual coding is time-consuming, expensive, and error-prone. To reduce coding errors and cost, we aim at developing an ICD code assignment system that automatically and accurately assigns the diagnostic description(DD) to ICD codes. Automated ICD coding is the multi-label classification task, which is to assign the ICD codes to the diagnostic description. In this paper, we apply the Neural Machine Translation(NMT) to the multi-label classification task of the code assignment problem. We proposed RAANMT(Recurrent-Based Encoder and Average Attention-Based Decoder for Neural Machine Translation), which can extract the relationship between the text and the labels to improve the automated ICD coding. Moreover, we implement experiments on Chinese dataset CDTD-TP and English dataset MIMIC III. Extensive experiments show that our RAANMT model can improve the performance of the automated code assignment.