How to leverage large language models for automatic ICD coding

Anonymous

Abstract: ICD coding, which indicates assigning appropriate ICD codes to clinical notes, is imperative for various healthcare circumstances such as health expense claims, insurance claims, and disease research. However, clinical notes contain numerous non-grammatical expressions, abbreviations, professional terms, and synonyms, rendering them notably noisy compared to general documents. Additionally, ICD coding also presents challenges such as a broad label space and a long-tail problem. While Large Language Models (LLMs) possess exceptional ability for natural language comprehension and thus hold potential for high-quality ICD coding, fine-tuning considering the unique properties of clinical notes and ICD codes is requisite. In this research, we propose a novel fine-tuning framework for LLMs toward automatic ICD coding. Our framework includes additional structures of label attention mechanism, note-relevant knowledge injection based on medical expressions, and knowledge-driven sampling for input clinical notes to navigate the input token limitations of LLMs. Our experiments on the MIMIC-III-50 dataset demonstrate that our framework achieves higher scores across both micro and macro measurements compared to the vanilla fine-tuning framework, with notably enhanced performance improvements observed in encoder-decoder models.

Paper Type: long

Research Area: Information Extraction

Contribution Types: NLP engineering experiment

Languages Studied: English

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.