Turning Dialogues Into Event Data: Lessons From GPT-Based Recognition of Nursing Actions.

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Nursing Process Analysis

Background:

• Home-based nursing care involves extensive documentation, contributing to significant administrative burden.
• Structured event logs are crucial for process analysis and quality improvement in healthcare.

Purpose of the Study:

• To evaluate the feasibility of using large language models (LLMs) for generating structured event logs from home care nursing conversations.
• To reduce the documentation burden on healthcare professionals and enable data-driven process insights.

Main Methods:

• An exploratory study utilized 27 transcribed audio recordings of home care visits.
• A Generative Pre-Trained Transformer (GPT) model was employed to identify nursing interventions using the Nursing Interventions Classification (NIC) system.
• Iterative prompt engineering and evaluation involved computer scientists and nurse researchers.

Main Results:

• GPT showed capability in extracting explicitly stated and temporally aligned nursing interventions.
• Implicit or ambiguous information presented challenges for accurate extraction.
• Five guidelines were developed for optimizing LLM application in this domain, addressing data, labeling, confidence, and output customization.

Conclusions:

• LLMs demonstrate potential for converting unstructured clinical dialogue into structured care data.
• Expert supervision and prompt refinement are essential for reliable LLM application in healthcare.
• Careful implementation is necessary to ensure accuracy, transparency, and trust in real-world healthcare settings.