RESPECT: a conversational AI system for informed consent with accuracy, safety, and stakeholder-centered evaluation

Area of Science:

• Clinical Research Ethics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Informed consent (IC) is crucial in clinical research, but current practices often lack individualized dialogue.
• Large Language Models (LLMs) present opportunities to enhance IC accessibility and engagement.
• Ensuring accuracy and safety is paramount before deploying LLMs in research settings.

Purpose of the Study:

• To develop and evaluate RESPECT (RESearch Participant Engagement and Consent Tool), an LLM-based assistant for clinical research informed consent.
• To assess the accuracy, safety, and utility of the RESPECT tool using a novel evaluation framework.
• To compare RESPECT's performance against existing LLMs like GPT-4.

Main Methods:

• Developed RESPECT, an LLM consent assistant employing Retrieval-Augmented Generation (RAG) grounded in IC documents.
• Evaluated accuracy via leave-one-out cross-validation and question rephrasing.
• Introduced a Refusal-Utility Curve (RUC) framework to measure appropriate refusal and utility.
• Conducted stakeholder evaluations with research staff.

Main Results:

• The RAG system demonstrated high accuracy in information retrieval from IC documents.
• RESPECT achieved significantly higher appropriate refusal rates compared to GPT-4.
• RESPECT's utility in answering legitimate questions was reduced at the cost of increased safety.
• Stakeholder evaluations indicated positive assessments of accuracy, comprehensiveness, and satisfaction.

Conclusions:

• RESPECT is the first RAG-based LLM consent assistant for research, enhancing safety through improved appropriate refusal rates.
• The RUC framework offers a valuable method for assessing safety-utility tradeoffs in LLM systems.
• Findings support informed decisions regarding the deployment of LLM tools in healthcare research.