Automatic ICD Coding Exploiting Discourse Structure and Reconciled Code Embeddings

Proceedings of the 29th International Conference on Computational Linguistics, pages 2883–2891 October 12–17, 2022. 2883 Automatic ICD coding exploiting discourse structure and reconciled code embeddings Shurui Zhang1, Bozheng Zhang1, Fuxin Zhang1, Bo Sang1 and Wanchun Yang ∗21 1MsunHealth Co., LTD 2Shandong Jiaotong University {zhang65821, oscarzbz, zfx050621, kangtasz}@gmail.com yangwanchun82@gmail.com Abstract The International Classification of Diseases (ICD) is the foundation of global health statis- tics and epidemiology. The ICD is designed to translate health conditions into alphanumeric codes. A number of approaches have been pro- posed for automatic ICD coding, since manual coding is labor-intensive and there is a global shortage of healthcare workers. However, exist- ing studies did not exploit the discourse struc- ture of clinical notes, which provides rich con- textual information for code assignment. In this paper, we exploit the discourse structure by leveraging section type classification and section type embeddings. We also focus on the class-imbalanced problem and the hetero- geneous writing style between clinical notes and ICD code definitions. The proposed rec- onciled embedding approach is able to tackle them simultaneously. Experimental results on the MIMIC dataset show that our model outper- forms all previous state-of-the-art models by a large margin. The source code is available at https://github.com/discnet2022/discnet 1 Introduction The International Classification of Diseases (ICD) is a classification system maintained by the World Health Organization. The system is designed to map health conditions to pre-defined ICD codes, allowing the world to share healthcare data in a con- sistent way between different regions. The founda- tion of global health statistics and epidemiology is based on the ICD. The ICD coding task (as shown in table 1) is usually performed by professional coders. Coders review the whole clinical documents and manu- ally assign the most appropriate codes. However, manual coding is labor-intensive, expensive, and error-prone. The approximate cost of ICD coding is estimated to be about $25 billion per year in the US (Lang, 2007). ∗Corresponding author. Clinical Document History of Present Illness: A 62-year-old male with Type II diabetes mellitus, coronary artery disease, hypertension, chronic kidney disease... Past Medical History: Hypertension Type II Diabetes Mellitus s/p cervical laminoplasty... Brief Hospital Course: ... Discharge Diagnosis: Anasarca Heart failure with restrictive physiology... Assigned ICD codes 428.31 Diastolic heart failure 584.9 Acute renal failure, unspecified 427.32 Atrial flutter ... Table 1: An illustration of ICD coding task In recent years, deep learning approaches have demonstrated promising results on ICD coding. Some of these studies improved clinical document representation by leveraging Convolutional Neu- ral Networks (CNN) (Mullenbach et al., 2018; Xie et al., 2019). The others improved ICD code repre- sentation by exploiting the dependencies between codes (Xie and Xing, 2018; Vu et al., 2020; Cao et al., 2020). However, these approaches entail lim- itations. Firstly, they ignore the discourse structure of clinical documents. Secondly, most of these ap- proaches did not consider the writing style discrep- ancies between ICD code descriptions and relevant clinical documents related to the codes. Thirdly, most of these approaches did not consider the class imbalanced …

Content truncated to 500 words