Leveraging LLMs for Efficient Data Structure Standardization in Chinese Medical Examination Reports.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Chinese medical examination reports lack standardized terminology, hindering health assessment and data analysis.
• Variations in reporting across hospitals create significant challenges for consistent data interpretation.

Purpose of the Study:

• To evaluate the efficacy of large language models (LLMs) in standardizing Chinese medical examination reports.
• To assess prompt engineering and LoRA fine-tuning for improving LLM performance on medical text data.

Main Methods:

• Utilized a large dataset of over 900,000 Chinese medical reports.
• Employed prompt engineering and LoRA fine-tuning techniques on various LLMs, including Qwen2.5-14B.
• Conducted cross-validation experiments to evaluate model generalization on unseen hospital data.

Main Results:

• Fine-tuned LLMs, especially Qwen2.5-14B, achieved high accuracy in standardizing department and item detail names.
• LLMs significantly outperformed traditional methods like BERT in medical report standardization.
• Achieved an average accuracy of 98.34% on unseen hospital data, demonstrating strong generalization capabilities.

Conclusions:

• LLMs show significant potential for standardizing Chinese medical examination reports.
• This research provides a foundation for automated processing and analysis of large-scale medical data.
• Fine-tuned LLMs offer a robust solution to the challenge of non-standardized medical terminologies.