Integrating Human Feedback in a Virtual Smoking Cessation Coach Optimizing Behavioral and Identity Outcomes with Reinforcement Learning

Abstract

The use of chatbots in eHealth applications has grown significantly, providing an interactive platform for users to improve their health and lifestyle. To enhance these applications, human feedback has been incorporated, with a coach offering personalized guidance to help users achieve their goals. However, deciding whether to provide human feedback or rely solely on automated systems remains a challenge, especially given the cost constraints associated with human involvement. This study presents a reinforcement learning (RL) model designed to optimize this decision by analyzing users’ states and predicting the potential benefit of such interventions. The model was trained on data from a longitudinal study involving over 500 daily smokers and vapers who interacted with a virtual coach across multiple sessions. Our findings indicate that the RL model effectively assists in determining whether human feedback is valuable, increasing the likelihood of behavior change, enhancing users’ quitter identities, and further reducing smoking/vaping frequency. This research contributes to the development of more effective, resource-efficient eHealth interventions.