EMPOWER: Evolutionary Medical Prompt Optimization With Reinforcement Learning.

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Clinical Informatics

Background:

• Large Language Models (LLMs) show promise in healthcare but require specialized prompt engineering for reliability.
• Existing prompt optimization methods often overlook crucial medical domain knowledge and safety standards.
• Ensuring clinical utility necessitates advanced techniques for prompt quality enhancement.

Purpose of the Study:

• To introduce EMPOWER, a novel evolutionary framework designed to elevate the quality of medical prompts for LLMs.
• To address the limitations of current prompt engineering approaches in the medical domain.
• To facilitate the responsible and effective integration of LLMs into clinical settings.

Main Methods:

• Developed EMPOWER, incorporating a medical terminology attention mechanism for specialized representation learning.
• Implemented a multi-dimensional evaluation architecture assessing clarity, specificity, clinical relevance, and factual accuracy.
• Utilized structure-preserving algorithms, including a component-level evolutionary algorithm and a semantic verification module, to maintain clinical reasoning integrity and medical knowledge adherence.

Main Results:

• EMPOWER achieved a 24.7% reduction in factually incorrect content within medical prompts.
• Demonstrated a 19.6% improvement in domain specificity for LLM outputs.
• Achieved 15.3% higher clinician preference in blinded evaluations across diagnostic, therapeutic, and educational tasks.

Conclusions:

• The EMPOWER framework significantly enhances medical prompt quality, addressing critical challenges in LLM application.
• This approach facilitates more reliable and clinically useful LLM outputs in healthcare.
• EMPOWER promotes responsible integration of advanced AI tools into medical practice.