Optimizing Automated KCD Coding: A Retrieval-Verification Approach

Abstract

This study proposes a two-step Retrieval-Verification system for automating the assignment of Korean Standard Classification of Diseases (KCD) codes to free-text diagnoses. The system uses SapBERT-XLMR for initial retrieval, followed by Llama 3.1 for final verification and code selection. Combining the two models improved accuracy to 82.3%. Future work aims to improve the system's performance on abbreviations and conduct experiment with a larger dataset.

Keywords: Clinical coding; Embedding; KCD; Language models.