JMIR Publications

ABSTRACT

Background:

Sepsis is a complex, life-threatening condition that presents significant challenges due to its heterogeneity and the vast, unstructured data associated with it. Traditional methods of knowledge graph construction struggle to manage this complexity effectively.

Objective:

This study aims to harness the capabilities of large language models (LLMs) in conjunction with extensive real-world data, to develop a detailed and methodical knowledge graph focused on sepsis. The goal is to enhance our comprehension of sepsis and to furnish actionable insights for its clinical management. Additionally, we established a multicenter sepsis database (MSD) to enrich our analysis and findings.

Methods:

Our methodology involved the collection of clinical guidelines, public databases, and substantial real-world data pertinent to sepsis. Utilizing GPT-4.0, we executed we carried out tasks of relationship extraction and entity recognition tasks through innovative prompt engineering techniques. This approach facilitated the construction of a nuanced sepsis knowledge graph.

Results:

We established a sepsis database that includes three centers and encompasses over 10,000 individuals. Importantly, we identified nine entity concepts and types and defined eight semantic relationships, successfully integrating the gathered knowledge through the Unified Medical Language System (UMLS) entity linker. As a result, we compiled a comprehensive sepsis knowledge graph, comprising 1,894 nodes and 2,021 distinct relationships.

Conclusions:

This study represents a groundbreaking effort in employing prompt engineering with GPT4.0 to establish a database and knowledge graph, thereby facilitating a systematic and in-depth understanding of sepsis. The inventive application of prompt engineering opens new avenues for the advancement of knowledge graphs, providing significant technical support to enhance both the efficiency and quality of knowledge graph construction.