Language Models for Value-Informed Proxy Decision Support

AbstractWhen patients lose decision-making capacity, large language models (LLMs) may help proxies choose treatments that honor incapacitated patients' values and preferences, especially when advance directives or formally identified proxies are ambiguous or unavailable. In a primary analysis, we prospectively enrolled 30 adults representing 15 dyads (15 patients alongside their legal, patient-designated proxy decision-maker) and compared patient-proxy and patient-LLM binary decisional agreement on multiple-choice hypothetical clinical scenarios, with and without providing patients' value profiles to the LLM; patients' value profiles were not provided to the human proxy. We found that patient-proxy decisional agreement (i.e., choosing the same treatment) on hypothetical clinical scenarios was 44.7% (95% confidence interval [CI], 28.6 to 61.9), similar to previous reports. Patient-LLM agreement was greater at 72.6% (95% CI, 61.9 to 85.7). When the LLM was stripped of information describing patients' values, its ability to choose the same treatment as the patient decreased substantially and matched the ability of human proxies. In secondary analyses, we explored the extraction of value profiles from electronic health record (EHR) data for patients who lose decisional capacity before formally documenting their values, and we critically evaluated value profiles written at lower reading levels. We found that the LLM inferred patients' values from real-world EHR clinical notes with low accuracy, underscoring the importance of having patients formally document their values before decisional capacity is lost. LLM utility degraded when value profiles were written at lower reading levels. Value-informed LLMs show promise as a proof of concept in the potential to assist human proxies in making treatment decisions that reflect the values of patients who have lost decisional capacity. Our findings caution against inferring patients' values from EHR data and using value profiles that are written at lower reading levels.