Toward integrating Large Language Models with Multi-armed bandit for personalized interventions in digital mental health

• Jacob Gordon, Computer Science, University of Delaware

• Matthew Mauriello, Computer Science, University of Delaware

Computer Science students suffer from elevated mental health issues compared to other undergraduate students. On the other hand, there is limited research on developing solutions to improve the mental health of university CS students. Digital mental health is promising for scaling psychological well-being but is often insufficiently personalized. On the other hand, psychology-based interventions offer highly effective support but can be prohibitively ineffective if not provided at the right moment. Our research draws inspiration from cognitive and psychology based interventions to explore whether and how large language models (LLMs) might address personalization challenges in digital mental health. I have developed an application InterventionGPT, which delivers interventions incorporating device interactions data from a user’s personal computer utilizing three different techniques. The first one uses just a contextual Multi-Armed Bandit (cMAB) to select the intervention and the second only uses the LLM based AI agent. The last one uses a cMAB to select the intervention type and an LLM based AI agent to personalize the intervention message. In a small qualitative-based feasibility study among three participants, we compare the three methods along with the perceived effectiveness of each. This research will inform the development of a larger usability study to fully ascertain the acceptability of each method in real-time.