Retrieval-Augmented Generation for ICD-10 Coding in German Clinical Texts - A Technical Case Report.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Documentation

Background:

• Manual International Classification of Diseases (ICD) coding of German clinical texts is inefficient and prone to errors.
• Existing automated methods struggle with low-resource languages and rare codes.

Purpose of the Study:

• To develop a semi-automated pipeline for efficient ICD coding of unstructured German medical documentation.
• To address limitations of current fine-tuned language models in German clinical text analysis.

Main Methods:

• Implemented a Retrieval-Augmented Generation (RAG) framework with Named Entity Recognition, semantic retrieval, and abbreviation resolution.
• Utilized Sentence-BERT embeddings, FAISS indexing, and a Mistral-Small-Instruct generative model for ICD code assignment.
• Combined semantic similarity and generative refinement for accurate code selection.

Main Results:

• The developed pipeline integrates multiple NLP techniques for semi-automated ICD coding.
• Identified semantic retrieval and diagnosis normalization as key areas for error reduction.
• Demonstrated a novel approach for handling low-resource clinical text coding challenges.

Conclusions:

• The RAG pipeline offers a promising approach for semi-automated ICD coding in German.
• Future work should enhance domain-specific embeddings and context-aware prompting for improved accuracy.
• The system aims to increase the accuracy and usability of clinical coding in real-world environments.