AI Chatbots in Health: Implementing an LLM-Based Solution to Promote Physical Activity

Abstract

With the emergence of powerful generative AI models comes the possibility of creating knowledgeable and engaging chatbots, which have the potential to significantly enhance several areas of the user’s life. This thesis focuses on the design and implementation of FysBot, an application with an integrated chatbot that aims to increase the user’s physical activity levels. In collaboration with a PhD project, FysBot’s design is guided by scientific evidence on user preferences in physical activity chatbots, with a strong emphasis on the security and privacy of the users’ data.

To facilitate personalized conversations on different topics, the use of chatbot personas is proposed. A chatbot persona is specialized for a single conversation topic and is responsible for generating an appropriate response to the user. FysBot implements personas for the following conversation topics: exercise recommendations, hiking trip suggestions and step progress review. In addition, a default persona is used to handle queries unrelated to the listed conversation topics. In order to mediate a query to the correct persona, queries are first presented to a routing persona, which then routes the query to the appropriate destination.

FysBot’s design has undergone testing via a usability study conducted using the think-aloud method. From this study, the overall impression of FysBot has been positive. However, several areas require improvement to enhance the user experience and intuitiveness. Addressing these areas will, in turn, increase engagement and promote healthier lifestyles among users.