ai
Researchers Develop RESPECT Conversational AI System for Informed Consent
A team at Stanford University created RESPECT, an LLM-based assistant that uses retrieval-augmented generation to answer questions about clinical trial informed consent documents. The system was evaluated for accuracy, safety through appropriate refusal rates, and stakeholder feedback from research staff.
May 8, 8:00 PM·1m read1 source
swissinfo.chAudio version
Tap play to generate a narrated version.
Informed consent is a cornerstone of clinical research. It typically includes written materials and an oral discussion between the investigator and participant. In practice both components tend to be templated and standardized, limiting opportunities for meaningful individualized dialog.
Researchers developed RESPECT (RESearch Participant Engagement and Consent Tool), an LLM consent assistant that utilizes retrieval-augmented generation to ground responses in informed consent source documents. The system aims to enhance accessibility of informed consent while maintaining accuracy, safety and appropriateness before research deployment.
The team evaluated accuracy through leave-one-out cross-validation and question rephrasing analysis. These tests demonstrated high accuracy in information retrieval for the RAG system.
A novel safety evaluation framework was introduced that measures two dimensions: appropriate refusal and utility. Appropriate refusal tracks how often the system refuses questions it should not answer. Utility tracks how often it answers questions it should answer.
This approach generalizes simple refusal rates by plotting a Refusal-Utility Curve analogous to ROC-AUC curves. RESPECT demonstrated significantly higher appropriate refusal rates compared to GPT-4. The improvement came at the cost of reduced utility in answering legitimate questions.
Stakeholder evaluations were conducted with research staff to assess accuracy, comprehensiveness and satisfaction. RESPECT represents the first RAG-based LLM consent assistance tool developed specifically for research contexts. It demonstrated improved safety through higher appropriate refusal rates.
The novel Refusal-Utility Curve evaluation framework provides researchers with a tool for assessing safety-utility tradeoffs in LLM systems. This enables informed decisions about deploying such tools in healthcare research settings. The study is funded by the Stanford University School of Medicine Department of Psychiatry & Behavioral Sciences 2024 Innovator Grants Program.
