Human-Machine Agreement in Medical Ethics: Patient Autonomy Case-Based Evaluation of Large Language Models.

Area of Science:

• Medical Ethics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Medical ethics, grounded in beneficence, nonmaleficence, patient autonomy, and justice, guides clinical practice.
• Patient autonomy is paramount in ethical dilemmas, emphasizing patient needs.
• Large language models (LLMs) are increasingly used in medicine, raising ethical questions about reasoning and context.

Purpose of the Study:

• To evaluate foundational large language models (LLMs) in medical ethics, focusing on patient autonomy.
• To explore methods for improving LLM performance using patient autonomy-based hypothetical cases.
• To address the lack of research on LLM applications in rapidly evolving medical ethics.

Main Methods:

• Evaluated three foundational LLMs (ChatGPT, LLaMA, Gemini) using hypothetical patient autonomy cases.
• Compared LLM responses to physician consensus using Cohen κ statistic.
• Employed McNemar test for iterative improvement analysis and significance testing (P<.05).

Main Results:

• Initial agreement between foundational LLMs and physician consensus was slight to fair (Cohen κ).
• Iterative improvement techniques significantly enhanced LLM agreement to substantial or higher (κ=0.73-0.82).
• Improvement was statistically significant across all tested LLMs (P=.006 for ChatGPT, P<.001 for Gemini and LLaMA).

Conclusions:

• Foundational LLMs demonstrate potential in medical ethics but require cautious application.
• Human oversight in testing and established techniques can align LLM responses with human ethical reasoning.
• Further research and iterative refinement are crucial for safe and effective LLM integration into medical ethics.