Large Language Model-Driven Knowledge Graph Construction in Sepsis Care Using Multicenter Clinical Databases:...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Data Science

Background:

• Sepsis presents significant heterogeneity and data complexity, challenging traditional knowledge graph construction.
• Large Language Models (LLMs) offer a novel approach to integrate and analyze unstructured clinical data for improved sepsis management.

Purpose of the Study:

• To develop a comprehensive sepsis knowledge graph using GPT-4.0 and multicenter clinical databases.
• To enhance the understanding of sepsis and provide actionable insights for clinical decision-making.
• To establish a multicenter sepsis database (MSD) to support knowledge graph development.

Main Methods:

• Collected clinical guidelines, public databases, and real-world data from three hospitals (10,544 sepsis patients).
• Employed GPT-4.0 with advanced prompt engineering for entity recognition and relationship extraction.
• Constructed a nuanced sepsis knowledge graph integrating diverse data sources.

Main Results:

• Established a sepsis database with 10,544 patient records.
• Developed a sepsis knowledge graph with 1894 nodes and 2021 relationships across nine entity concepts.
• GPT-4.0 achieved superior F1-scores (76.76% on sepsis data, 65.42% on CMeEE dataset) for entity recognition and relationship extraction compared to other models.

Conclusions:

• Pioneering use of LLMs (GPT-4.0) for comprehensive sepsis knowledge graph construction.
• Advanced prompt engineering and multicenter data integration enhanced efficiency and accuracy.
• The sepsis knowledge graph offers a robust framework for clinical decision-making and future research.