Research through Evaluation for Large Language Model in Patient-Clinician Communications.

Area of Science:

• Healthcare technology evaluation
• Artificial intelligence in medicine
• Clinical informatics

Background:

• Large Language Models (LLMs) show promise in healthcare, but evaluating their outputs at scale is difficult.
• Current metrics are often insufficient for assessing advanced LLM-based healthcare tools.
• Human evaluator recruitment for large-scale LLM output assessment presents significant challenges.

Purpose of the Study:

• To propose and validate a Research through Evaluation (RtE) approach for refining metrics and improving assessments of LLM-generated healthcare content.
• To enhance the transparency and robustness of evaluation methodologies for AI in clinical settings.

Main Methods:

• A retrospective comparative study was conducted, comparing human clinical care team responses with GPT-4-generated responses for prostate cancer patient inquiries.
• The Research through Evaluation (RtE) method was deployed, involving three rounds of co-evaluations with clinical professionals.
• An LLM-as-graders study was performed using four LLMs to assess response quality.

Main Results:

• Iterative evaluations using the RtE method enabled human graders to better adapt metrics to their clinical practices.
• LLMs, in the LLM-as-graders study, demonstrated a preference for the GPT-4 in-basket bot's responses over those from the human clinical care team.
• The RtE approach proved effective in refining evaluation metrics for LLM-generated content.

Conclusions:

• The Research through Evaluation (RtE) method enhances the transparency and robustness of evaluation metrics for LLM-generated content in healthcare.
• The RtE methodology shows potential for generalization to the evaluation of AI-generated content in various clinical practices.
• This study highlights the utility of RtE in addressing challenges in evaluating large-scale LLM applications in medicine.