From Patient Consultations to Graphs: Leveraging LLMs for Patient Journey Knowledge Graph Construction.

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Knowledge Representation

Background:

• The healthcare industry is shifting towards patient-centric models, necessitating a comprehensive understanding of patient journeys.
• Existing healthcare data systems lack holistic representations, impeding effective care coordination.
• Patient Journey Knowledge Graphs (PJKGs) offer a solution by unifying diverse patient information into structured formats.

Purpose of the Study:

• To present a methodology for constructing PJKGs using Large Language Models (LLMs).
• To demonstrate how PJKGs can capture temporal and causal relationships within patient data.
• To enable advanced reasoning and personalized insights for improved healthcare.

Main Methods:

• Utilized LLMs to process clinical documentation and patient-provider conversations.
• Developed a methodology for building PJKGs from unstructured and semi-structured health data.
• Evaluated four LLMs (Claude 3.5, Mistral, Llama 3.1, ChatGPT4o) for PJKG construction.

Main Results:

• All evaluated LLMs demonstrated perfect structural compliance in PJKG construction.
• Significant variations were observed among LLMs in medical entity processing, computational efficiency, and semantic accuracy.
• The constructed PJKGs effectively capture temporal and causal relationships between clinical events.

Conclusions:

• PJKGs provide a robust framework for integrating disparate patient information.
• LLM-driven PJKG construction advances patient-centric healthcare by enabling enhanced care coordination and outcome prediction.
• Further research is needed to optimize LLM performance for specific medical entity recognition and semantic accuracy in PJKGs.