https://www.mdu.se/

Fall injuries are the leading cause of hospital admissions and result in both health-related consequences and economic repercussions. This thesis contributes to the PRE-fall research project by focusing on the implementation and evaluation of utilizing different prompting techniques to personalize communication with a Large Language Model (LLM), integrated as a virtual coach in an e-health mobile application for fall-preventive physical activity(PA). Previous work within the PRE-fall research project built a solid foundation of the application, focusing on the knowledge- and practice components of The Knowledge, Attitude and Practice (KAP) theory. Building on this foundation, this thesis explores how the user’s attitude towards PA can help personalize the tonal output of a virtual coach, while still maintaining relevance and substance.

An onboarding module was implemented using Flutter and Python to collect information regarding the user’s attitude towards PA. This information is then utilized to increase personalization, aiming to enhance both motivation and adherence towards completing fall-preventive PA programs. Exploring two different prompting techniques, the LLM’s responses to exercise-related questions and statements were evaluated by ten respondents through a survey with respect to tone, relevance and substance in relation to the fictional users’ attitude.

The results show that the tonal output can be well adjusted with lightweight prompting techniques while maintaining relevance and substance but also highlights the importance of how the instructions and example prompts are formulated and structured. This work aims to guide in situations where prompting techniques are deemed suitable to adjust the tonal output of an LLM, in contexts where improving users’ motivation and adherence are important.