Abstract

The increasing use and functionality of Large Language Model (LLM)-powered chatbots creates an opportunity to adapt these for other applications than information retrieval. One potential application is within psychiatry where research has shown that chatbots can complement psychiatrists to further aid the patient by being more readily available to either simply chat, answer questions or provide support. One psychiatric diagnosis where LLM-powered chatbots may be helpful is selective mutism in children. However, hosting a chatbot in the cloud raises concerns regarding the integrity and privacy of the patient’s personal data. The aim of this thesis is therefore to investigate both the potential need for a companion-like chatbot during the treatment of selective mutism in children and the feasibility of running this chatbot locally on a small portable computer, like a Jetson Nano, to ensure data integrity and privacy. The potential role of a chatbot during treatment is investigated through a literature review and interviews with two psychologists who work with children diagnosed with selective mutism. The feasibility of running a sufficiently capable LLM is evaluated by testing whether an LLM can run on a Jetson Nano and produce responses quickly and coherently enough for conversational use. A prototype chatbot is also developed as a proof-of-concept to demonstrate how LLMs could be
implemented. The results show that a chatbot companion can be helpful during the treatment of selective mutism in children if the chatbot is adapted to the child’s age, as the treatment differs for different age groups. The developed prototype shows how a locally hosted chatbot could be implemented on capable hardware. However, the Jetson Nano is neither modern nor powerful enough to run LLMs capable of generating sufficient responses. Newer Jetson models may however, provide sufficient computational capacity for such applications.