Modeling Diagnostic Label Correlation for Automatic ICD Coding

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4043–4052 June 6–11, 2021. ©2021 Association for Computational Linguistics 4043 Modeling Diagnostic Label Correlation for Automatic ICD Coding Shang-Chi Tsai∗ Chao-Wei Huang∗ Yun-Nung Chen National Taiwan University, Taipei, Taiwan {d08922014,f07922069}@csie.ntu.edu.tw y.v.chen@ieee.org Abstract Given the clinical notes written in electronic health records (EHRs), it is challenging to pre- dict the diagnostic codes which is formulated as a multi-label classification task. The large set of labels, the hierarchical dependency, and the imbalanced data make this prediction task extremely hard. Most existing work built a bi- nary prediction for each label independently, ignoring the dependencies between labels. To address this problem, we propose a two-stage framework to improve automatic ICD coding by capturing the label correlation. Specifically, we train a label set distribution estimator to rescore the probability of each label set can- didate generated by a base predictor. This pa- per is the first attempt at learning the label set distribution as a reranking module for medi- cal code prediction. In the experiments, our proposed framework is able to improve upon best-performing predictors on the benchmark MIMIC datasets. 1 1 Introduction Clinical notes from electronic health records (EHRs) are free-from text generated by clinicians during patient visits. The associated diagnostic codes from the International Classification of Dis- eases (ICD) represent diagnostic and procedural information of the visit. The ICD codes provide an standardized and systematic way to encode infor- mation and has several potential use cases (Choi et al., 2016). Considering that manual ICD coding has been shown to be labor-intensive (O’malley et al., 2005), several approaches for automatic ICD cod- ing has been proposed and investigated by the re- search community (Perotte et al., 2013; Kavuluru et al., 2015). With recent introduction of deep neural networks, the performance of automatic ∗Equal contribution. 1The source code of this project is available at https://github.com/MiuLab/ICD-Correlation. …250.61, 357.2, 564.0, 564.5, 401.9… a set of ICD prediction Constipation Functional diarrhea 564.0 : 564.5 : Figure 1: An example of conflicting predictions. While these two codes share the same root in the hierarchical ICD struc- ture and are semantically similar, they are unlikely to appear together. ICD coding has been improved significantly (Choi et al., 2016; Shi et al., 2017; Mullenbach et al., 2018; Baumel et al., 2018; Xie and Xing, 2018; Li and Yu, 2020; Vu et al., 2020; Cao et al., 2020). Prior work on neural models mostly treated the task of automatic ICD coding as a multi-label clas- sification problem. These models mostly employ a shared text encoder, and build one binary clas- sifier for each label on top of the encoder. This architecture along side with binary cross-entropy loss make the prediction of each label independent of each other, which might lead to incomplete or conflicting predictions. An example of such er- ror is shown in Figure 1. This issue is especially problematic in ICD code prediction, since the ICD codes share a …

Content truncated to 500 words