Multi-label Diagnosis Classification of Swedish Discharge Summaries – ICD-10 Code Assignment Using KB-BERT

Proceedings of Recent Advances in Natural Language Processing, pages 1158–1166 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_130 1158 Multi-label Diagnosis Classification of Swedish Discharge Summaries – ICD-10 Code Assignment Using KB-BERT Sonja Remmera,b remmer@dsv.su.se Anastasios Lamproudisa anastasios@dsv.su.se aDepartment of Computer and Systems Sciences (DSV) Stockholm University, Kista, Sweden bNorwegian Centre for E-health Research, Tromsø, Norway Hercules Dalianisa,b hercules@dsv.su.se Abstract The International Classification of Diseases (ICD) is a system for systematically recording patients’ diagnoses. Clinicians or professional coders assign ICD codes to patients’ medical records to facilitate funding, research, and ad- ministration. In most health facilities, clinical coding is a manual, time-demanding task that is prone to errors. A tool that automatically assigns ICD codes to free-text clinical notes could save time and reduce erroneous coding. While many previous studies have focused on ICD coding, research on Swedish patient records is scarce. This study explored different approaches to pairing Swedish clinical notes with ICD codes. KB-BERT, a BERT model pre-trained on Swedish text, was compared to the traditional supervised learning models Support Vector Machines, Decision Trees, and K-nearest Neighbours used as the baseline. When considering ICD codes grouped into ten blocks, the KB-BERT was superior to the base- line models, obtaining an F1-micro of 0.80 and an F1-macro of 0.58. When considering the 263 full ICD codes, the KB-BERT was outper- formed by all baseline models at an F1-micro and F1-macro of zero. Wilcoxon signed-rank tests showed that the performance differences between the KB-BERT and the baseline mod- els were statistically significant. 1 Introduction There are both administrative and statistical pur- poses of ICD coding. Administrative to reimburse the clinical unit or hospital, but also to plan health- care. The codes are assigned by both treating physi- cians and designated coders. The current version of the ICD system, ICD-10, contains tens of thou- sands of codes divided into 22 chapters (WHO, 2016). ICD coding is time-consuming and error-prone, either missing the main diagnosis or displaying er- rors in the coding in up to 20 per cent of the patient records (Jacobsson and Serd´en, 2013). Therefore, it would be valuable to have a supporting tool to assist the physician or coder in choosing among the codes. In this article, Swedish patient records in the medical speciality of gastrointestinal surgery and their already assigned ICD-10 codes are used to per- form supervised learning to predict ICD-10 codes. More specifically, the part of the patient records that summarises the patient’s care period at the time of the discharge, the discharge summaries, and their associated ICD-10 codes are used. The assigned codes belong to the Swedish version of the ICD- 10 system known as ICD-10-SE (Socialstyrelsen, 2018). The codes considered are both full ICD codes at the highest level of granularity and the full codes grouped into ten blocks. The research question is how the deep learning language model KB-BERT, compared to the traditional supervised learning models Support Vector Machines, Deci- sion Trees, and K-Nearest Neighbours performs in pairing discharge summaries with the correct ICD codes. 2 Related …

Content truncated to 500 words