Protocol For Human Evaluation of Artificial Intelligence Chatbots in Clinical Consultations

Abstract

AbstractBackgroundGenerative artificial intelligence (AI) technology has the revolutionary potentials to augment clinical practice and telemedicine. The nuances of real-life patient scenarios and complex clinical environments demand a rigorous, evidence-based approach to ensure safe and effective application.MethodsWe present a protocol for the systematic evaluation of generative AI large language models (LLMs) as chatbots within the context of clinical microbiology and infectious disease consultations. We aim to critically assess the clinical accuracy, comprehensiveness, coherence, and safety of recommendations produced by leading generative AI models, including Claude 2, Gemini Pro, GPT-4.0, and a GPT-4.0-based custom AI chatbot.DiscussionA standardised healthcare-specific prompt template is employed to elicit clinically impactful AI responses. Generated responses will be graded by a panel of human evaluators, encompassing a wide spectrum of domain expertise in clinical microbiology and virology and clinical infectious diseases. Evaluations are performed using a 5-point Likert scale across four clinical domains: factual consistency, comprehensiveness, coherence, and medical harmfulness. Our study will offer insights into the feasibility, limitations, and boundaries of generative AI in healthcare, providing guidance for future research and clinical implementation. Ethical guidelines and safety guardrails should be developed to uphold patient safety and clinical standards.