Enhancing medical coding efficiency through domain-specific fine-tuned large language models.

Area of Science:

• Artificial Intelligence
• Health Informatics
• Medical Coding Automation

Background:

• Medical coding is a critical healthcare operation, but current manual processes are inefficient, prone to errors (up to 20%), and costly ($18.2 billion annually).
• Existing large language models (LLMs) show limited accuracy when applied to medical coding tasks.
• There is a need for automated solutions to improve the accuracy and efficiency of medical code generation.

Purpose of the Study:

• To evaluate the effectiveness of fine-tuning LLMs with specialized ICD-10 knowledge for automating medical code generation.
• To assess the impact of domain-specific training on LLM performance in medical coding.
• To compare the performance of different LLM models and training approaches.

Main Methods:

• A two-phase fine-tuning approach was employed for LLMs.
• Phase 1 involved initial fine-tuning using 74,260 ICD-10 code-description pairs.
• Phase 2 focused on enhanced training to address linguistic and lexical variations in clinical documentation.

Main Results:

• Initial fine-tuning dramatically increased exact matching from less than 1% to 97%.
• Enhanced fine-tuning further improved performance, achieving 69.20% exact match and 87.16% category match on real-world clinical notes.
• Evaluations were conducted on both proprietary (GPT-4o mini) and open-source (Llama) LLM models.

Conclusions:

• Domain-specific fine-tuned LLMs can effectively automate medical code generation.
• This approach significantly reduces manual coding burdens and improves the reliability of medical coding.
• The findings support the integration of specialized LLMs into healthcare operations for enhanced efficiency and accuracy.