Incorporating Patient Values in Large Language Model Recommendations for Surrogate and Proxy Decisions

Area of Science:

• Medical Ethics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Shared decision-making for incapacitated patients often fails to honor patient values due to clinician stress and biases.
• Advance directives have limitations in completion rates and applicability.
• Investigating large language models (LLMs) to support clinical decisions for incapacitated patients.

Purpose of the Study:

• To evaluate the potential of LLMs in incorporating patient values for critical care decisions.
• To assess LLM performance in generating treatment recommendations aligned with patient values.

Main Methods:

• Simulated 50 decisionally incapacitated patient scenarios with diverse value profiles (numeric, text, free-text).
• Utilized pre-trained LLMs for treatment extraction and value-based recommendation generation.
• Compared LLM outputs against domain expert adjudications.

Main Results:

• LLMs accurately extracted treatments in 88% of scenarios.
• LLM recommendations scored 3.92/5.00 for medical plausibility and 3.58/5.00 for reflecting patient values.
• Free-text, unstructured patient value narratives yielded the highest scores.

Conclusions:

• LLMs demonstrate potential as support tools for honoring incapacitated patients' wishes.
• LLMs can aid surrogates, proxies, and clinicians in value-concordant critical care decisions.
• Further research can refine LLM application in clinical decision support.