Enhancing Clinical Note Generation with ICD-10, Clinical Ontology Knowledge Graphs, and Chain-of-Thought Prompting...

Area of Science:

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

Background:

• The Health Information Technology for Economic and Clinical Health (HITECH) Act and 21st Century Cures Act have increased electronic health record (EHR) data.
• Physicians spend significant time manually writing clinical notes, impacting patient wait times and diagnostic speed.
• Large language models (LLMs) show potential for generating human-like text, including clinical documentation.

Purpose of the Study:

• To investigate the use of Chain-of-Thought (CoT) prompt engineering for improving LLM-based clinical note generation.
• To enhance LLM clinical note generation by integrating semantic search results and clinical knowledge graphs.
• To evaluate the effectiveness of advanced prompting techniques against standard methods for EHR data.

Main Methods:

• Utilized International Classification of Diseases (ICD) codes and basic patient information as input for LLM prompts.
• Implemented a hybrid approach combining traditional CoT with semantic search.
• Integrated a knowledge graph derived from clinical ontology to enrich domain-specific information.
• Tested the prompting technique on six clinical cases from the CodiEsp dataset using GPT-4.

Main Results:

• The developed CoT prompting strategy, enhanced with semantic search and knowledge graphs, outperformed standard one-shot prompts.
• Generated clinical notes demonstrated improved quality and domain-specific accuracy.
• GPT-4, when guided by the advanced prompting technique, produced superior clinical notes compared to baseline methods.

Conclusions:

• Advanced CoT prompt engineering, augmented with semantic search and knowledge graphs, is a promising method for improving LLM-based clinical note generation.
• This approach can help reduce physician documentation time and potentially improve patient care efficiency.
• Further research can explore scaling this technique across diverse clinical settings and EHR systems.