基于检索增强的大型语言模型用于指导老年人使用大麻二酚的证据支持
Retrieval-Augmented Large Language Models for Evidence-Informed Guidance on Cannabidiol Use in Older Adults
- Lawrence Bloomberg Faculty of Nursing, University of Toronto（多伦多大学劳伦斯·布隆伯格护理学院）
- KITE Research Institute, Toronto Rehabilitation Institute, University Health Network（大学健康网络多伦多康复研究所KITE研究所）
- College of Engineering and Technology, American University of the Middle East（中东美国大学工程与技术学院）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
本文提出一种结合结构化提示工程与大麻二酚证据的检索增强型大型语言模型框架，用于为老年人提供安全指导，同时开发了无标注评估框架以评估AI在敏感健康场景中的可靠性。
AI中文摘要：
老年人常患有慢性疾病，如疼痛和睡眠障碍，可能考虑使用大麻二醇进行症状管理。安全使用需要适当剂量、谨慎滴定和药物相互作用意识，但污名化和有限的健康素养常限制理解。基于大型语言模型和检索增强生成的对话式人工智能系统可能支持大麻二醇教育，但其安全性和可靠性尚不充分。本研究开发了一种结合结构化提示工程与精选大麻二醇证据的检索增强型大型语言模型框架，生成针对老年人（包括认知障碍者）的上下文感知指导。我们还提出了一个自动化、无标注的评估框架，用于在无标准化基准的情况下评估领先的独立模型和检索增强模型。通过变化症状、偏好、认知状态、人口统计学、共病、药物、大麻史和照顾者支持，生成了64种多样化的用户场景。评估了多种最先进的模型，包括一种新型的集成检索架构，整合了多个检索系统。在三种自动化评估策略中，检索增强模型始终比独立模型产生更谨慎且符合指南的推荐，其中集成方法表现最佳。这些发现表明，结构化检索提高了AI驱动的大麻二醇教育的可靠性和安全性，并提供了一种可重复使用的框架，用于评估用于敏感健康情境的AI工具。
英文摘要：
Older adults commonly experience chronic conditions such as pain and sleep disturbances and may consider cannabidiol for symptom management. Safe use requires appropriate dosing, careful titration, and awareness of drug interactions, yet stigma and limited health literacy often limit understanding. Conversational artificial intelligence systems based on large language models and retrieval-augmented generation may support cannabidiol education, but their safety and reliability remain insufficiently evaluated. This study developed a retrieval-augmented large language model framework that combines structured prompt engineering with curated cannabidiol evidence to generate context-aware guidance for older adults, including those with cognitive impairment. We also proposed an automated, annotation-free evaluation framework to benchmark leading standalone and retrieval-augmented models in the absence of standardized benchmarks. Sixty-four diverse user scenarios were generated by varying symptoms, preferences, cognitive status, demographics, comorbidities, medications, cannabis history, and caregiver support. Multiple state-of-the-art models were evaluated, including a novel ensemble retrieval architecture that integrates multiple retrieval systems. Across three automated evaluation strategies, retrieval-augmented models consistently produced more cautious and guideline-aligned recommendations than standalone models, with the ensemble approach performing best. These findings demonstrate that structured retrieval improves the reliability and safety of AI-driven cannabidiol education and provide a reproducible framework for evaluating AI tools used in sensitive health contexts.
