An Empirical Study of LLMs for Engineering Medical Consent.

Abstract

The process of obtaining patient informed consent for personal health information management and agreement on treatments has in recent years become a significant burden for healthcare providers, who often struggle to integrate complex consent requirements into existing clinical workflows. The difficulty increases as consent requirements must be continuously updated and maintained to accommodate new templates and regulatory specifications. Recent advances in Large Language Models (LLMs) offer potential on automating the engineering of medical consent. This paper examines the capabilities of popular LLMs (LLaMa, Gemma, DeepSeek and Phi-3) to be used as collaborative tools to help clinicians generate high-quality, compliant, and patient-centric consent forms from natural language inputs. We identified key clinical challenges and evaluated each LLM's ability to address them through tasks in formalization, comprehensibility, and compliance. Qualitative expert feedback shows that LLMs are effective companions, but a "one-size-fits-all" approach is insufficient. The paper provides useful implications to practitioners that helps them select the appropriate LLM for consent requirements engineering based on the context of the consent requirement.