Think FAST: a novel framework to evaluate fidelity, accuracy, safety, and tone in conversational AI health coach dialogues.

Developments in Machine Learning based Conversational and Generative Artificial Intelligence (GenAI) have created opportunities for sophisticated Conversational Agents to augment elements of healthcare. While not a replacement for professional care, AI offers opportunities for scalability, cost effectiveness, and automation of many aspects of patient care. However, to realize these opportunities and deliver AI-enabled support safely, interactions between patients and AI must be continuously monitored and evaluated against an agreed upon set of performance criteria. This paper presents one such set of criteria which was developed to evaluate interactions with an AI Health Coach designed to support patients receiving obesity treatment and deployed with an active patient user base. The evaluation framework evolved through an iterative process of development, testing, refining, training, reviewing and supervision. The framework evaluates at both individual message and overall conversation level, rating interactions as Acceptable or Unacceptable in four domains: Fidelity, Accuracy, Safety, and Tone (FAST), with a series of questions to be considered with respect to each domain. Processes to ensure consistent evaluation quality were established and additional patient safety procedures were defined for escalations to healthcare providers based on clinical risk. The framework can be implemented by trained evaluators and offers a method by which healthcare settings deploying AI to support patients can review quality and safety, thus ensuring safe adoption.