Prompt Engineering Paradigms for Medical Applications: Scoping Review

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Prompt engineering enhances large language model (LLM) capabilities, particularly vital in medicine due to specialized language.
• Clinical natural language processing (NLP) faces challenges with complex medical texts and privacy.
• Prompt engineering offers a novel method for extracting clinical information from medical texts.

Purpose of the Study:

• To review prompt engineering research and technical approaches in medical applications.
• To provide an overview of opportunities and challenges for clinical practice.
• To establish reporting guidelines for future medical prompt engineering studies.

Main Methods:

• Searched medical, computer science, and informatics databases, including preprints.
• Included studies from 2022-2024 using prompt learning (PL), prompt tuning (PT), and prompt design (PD).
• Extracted data on prompt paradigms, LLMs, languages, topics, baselines, and strategies.

Main Results:

• 114 prompt engineering studies were included; prompt design (PD) was most prevalent (78 papers).
• ChatGPT was the most used LLM; 7 studies used it with sensitive clinical data.
• Chain-of-thought was the most frequent PD technique; many PD studies lacked non-prompt baselines.

Conclusions:

• Prompt engineering shows promise in medicine, but reporting inconsistencies exist.
• Guidelines are provided to improve future research and reporting standards.
• Summarized data aims to facilitate further advancements in medical prompt engineering.