A Novel Framework Based on Medical Concept Driven Attention for Explainable Medical Code Prediction via External Knowledge

Findings of the Association for Computational Linguistics: ACL 2022, pages 1407 - 1416 May 22-27, 2022 c⃝2022 Association for Computational Linguistics A Novel Framework Based on Medical Concept Driven Attention for Explainable Medical Code Prediction via External Knowledge Tao Wang Linhai Zhang Chenchen Ye Junxi Liu Deyu Zhou∗ School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China {wangtao98, lzhang472, chenchenye, liujunxi, d.zhou}@seu.edu.cn Abstract Medical code prediction from clinical notes aims at automatically associating medical codes with the clinical notes. Rare code prob- lem, the medical codes with low occurrences, is prominent in medical code prediction. Recent studies employ deep neural networks and the external knowledge to tackle it. However, such approaches lack interpretability which is a vital issue in medical application. Moreover, due to the lengthy and noisy clinical notes, such approaches fail to achieve satisfactory results. Therefore, in this paper, we propose a novel framework based on medical concept driven attention to incorporate external knowledge for explainable medical code prediction. In spe- cific, both the clinical notes and Wikipedia doc- uments are aligned into topic space to extract medical concepts using topic modeling. Then, the medical concept-driven attention mecha- nism is applied to uncover the medical code related concepts which provide explanations for medical code prediction. Experimental re- sults on the benchmark dataset show the supe- riority of the proposed framework over several state-of-the-art baselines. 1 Introduction Medical codes, also known as ICD codes, are or- ganized by International Classification of Diseases (ICD, recent versions are ICD-9 and ICD-10) tax- onomies. Each medical code corresponds to a dis- ease, procedure or sign, and so on. Medical codes can abstract away fine details of free-text clinical notes, which provide great convenience for ana- lyzing clinical data directly (Shull, 2019; Bai and Vucetic, 2019). It is time consuming, costly and error-prone for manual medical coding due to the large menu of options (over 15,000 codes in ICD- 9) and the complex lengthy clinical notes (Adams et al., 2002; Lang, 2007). Medical code predic- tion aims at automatically associating the relevant ∗corresponding author Figure 1: An example of a clinical note annotated with 3-digit ICD-9 code “250” and the corresponding Wikipedia document, where words in red are medical concept-indicative words which can be employed as ev- idences to infer medical codes. medical codes with the clinical notes. Treating medical code prediction as a multi-label text classification problem, many machine learning based approaches have been proposed including Bayesian-based (Larkey and Croft, 1995) and Sup- port Vector Machine based (Lita et al., 2008; Per- otte et al., 2014). With the success of deep learning, many researchers propose neural networks with at- tention mechanism (Mullenbach et al., 2018; Li and Yu, 2020; Vu et al., 2020) to identify represen- tative words in clinical notes and those with large weights serve as evidence for prediction. Li and Yu (2020) utilize convolutional neural networks with several fixed window sizes to capture various medical patterns, then identify …

Content truncated to 500 words