How to leverage large language models for automatic ICD coding.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Accurate International Classification of Diseases (ICD) coding is crucial for healthcare administration, research, and claims processing.
• Manual ICD coding is inefficient, error-prone, and struggles with the complexity and noise inherent in clinical notes.
• Existing automated methods face challenges due to the broad label space and long-tail distribution of ICD codes.

Purpose of the Study:

• To develop and evaluate a novel fine-tuning framework for Large Language Models (LLMs) to automate ICD coding.
• To address the limitations of LLMs in processing noisy clinical text and handling the complexities of ICD code assignment.
• To improve the accuracy and efficiency of automatic ICD coding in healthcare settings.

Main Methods:

• Proposed a fine-tuning framework for LLMs specifically designed for automatic ICD coding.
• Incorporated a label attention mechanism to focus on relevant diagnostic labels.
• Integrated note-relevant medical knowledge and employed knowledge-driven sampling to manage LLM input token constraints.

Main Results:

• The proposed framework demonstrated superior performance compared to vanilla fine-tuning on the MIMIC-III-50 dataset.
• Significant improvements in micro and macro accuracy and F1 scores were observed.
• Encoder-decoder LLM architectures showed particularly notable gains with the enhanced fine-tuning approach.

Conclusions:

• The novel fine-tuning framework effectively enhances LLM capabilities for automatic ICD coding.
• The framework successfully addresses challenges posed by clinical note characteristics and ICD code distribution.
• This approach offers a promising solution for accurate and efficient automated ICD coding in clinical practice.