TreeMAN: Tree-enhanced Multimodal Attention Network for ICD Coding

Proceedings of the 29th International Conference on Computational Linguistics, pages 3054–3063 October 12–17, 2022. 3054 TreeMAN: Tree-enhanced Multimodal Attention Network for ICD Coding Zichen Liu†, Xuyuan Liu†, Yanlong Wen†∗, Guoqing Zhao‡, Fen Xia‡, Xiaojie Yuan† †College of Computer Science, Nankai University ‡Mashang Consumer Finance Co.,Ltd. {liuzichen,wenyl,yuanxj}@dbis.nankai.edu.cn hsuyuanliu0204@mail.nankai.edu.cn {guoqing.zhao02,fen.xia}@msxf.com Abstract ICD coding is designed to assign the disease codes to electronic health records (EHRs) upon discharge, which is crucial for billing and clin- ical statistics. In an attempt to improve the effectiveness and efficiency of manual coding, many methods have been proposed to automat- ically predict ICD codes from clinical notes. However, most previous works ignore the deci- sive information contained in structured med- ical data in EHRs, which is hard to be cap- tured from the noisy clinical notes. In this pa- per, we propose a Tree-enhanced Multimodal Attention Network (TreeMAN) to fuse tab- ular features and textual features into multi- modal representations by enhancing the text representations with tree-based features via the attention mechanism. Tree-based features are constructed according to decision trees learned from structured multimodal medical data, which capture the decisive information about ICD coding. We can apply the same multi-label classifier from previous text mod- els to the multimodal representations to pre- dict ICD codes. Experiments on two MIMIC datasets show that our method outperforms prior state-of-the-art ICD coding approaches. The code is available at https://github. com/liu-zichen/TreeMAN. 1 Introduction The International Classification of Diseases (ICD), maintained by the World Health Organization, is a hierarchical classification of codes representing diseases, injuries, and so on. ICD codes have been used in diverse areas, including insurance re- imbursement, epidemiology, and clinical research (Park et al., 2000). In the hospital, when patients discharge, their electronic health records (EHRs) and all associated data are transferred to the information management department, where clinical coders manually assign ∗Corresponding author Automatic Multimodal ICD Coding Model Clinical Text ITEM VALUE Medicine Insulin … … Hematocrit abnormal ICD Code Disease Name 250.13 type I diabetes mellitus 403.90� Hypertensive renal disease 584.9 Acute renal failure 585.3 Chronic kidney disease … … ITEM VALUE …Heart Rate 120 bpm … O2 Flow 15 L/min Figure 1: An example of automatic multimodal ICD coding. Model inputs include physiological data and medical records in addition to clinical text. the appropriate ICD codes using rigid ICD coding guidelines after reviewing records (O’malley et al., 2005). The manual code assignment is expensive, labor-intensive, and error-prone due to the large volume of medical record information and high professional requirements (Nguyen et al., 2018). Since deep learning has achieved great success in lots of healthcare applications (Cai et al., 2019), many neural methods have been proposed to auto- mate the ICD coding process by researchers (Teng et al., 2022). Recent works formulate automated ICD coding as a multi-label document classifica- tion task, using clinical notes as model input, pre- dicting coding with a multi-label classifier, and learning text features through word embedding techniques and neural networks such as RNNs and CNNs (Mullenbach et al., 2018; Vu …

Content truncated to 500 words