Personalizing Mental Health Chatbots for Young People

Depression is a prevalent mental disorder among young people, yet many affected individuals do not receive adequate support because of stigma, attitudinal barriers, and long waiting times for treatment. Mental health chatbots offer anonymous, easily accessible, and conversational support, but existing systems often fail to address young users' specific needs, the effects of different content personalization strategies remain poorly understood, and the rise of large language models (LLMs) introduces new design and evaluation challenges. This dissertation investigates how personalized mental health chatbots for young people can be designed, instantiated, and evaluated to improve user engagement and effectiveness across three complementary levels: youth-sensitive design, personalized content selection, and the personalization of therapeutic content during ongoing conversations.
The first part examines youth-specific needs through a qualitative study of young people with depression and derives design recommendations for chatbots to treat depression, highlighting the importance of age-sensitive content and personalization. The second part develops and evaluates TheraBot in two field experiments that compare standardized content with chatbot-led, user-led, and hybrid content personalization. ... mehrThe results show that content personalization can improve therapeutic outcomes, but only hybrid personalization consistently outperformed standardized content; moreover, the effects of personalization depended on individual characteristics such as depression severity, stated preferences, and personality traits, indicating that the personalization strategy itself should be adapted to the user.
The third part investigates how therapeutic content can be personalized during ongoing conversations through an LLM-based behavioral activation chatbot and how such a system can be iteratively refined. Using a structured prompt architecture, artificial users, and clinical expert fidelity assessment, the study shows that the chatbot can deliver behavioral activation with robust protocol adherence in most sessions while also revealing shortcomings in clinical judgment, rapport building, and handling skeptical or information-seeking user steering.
Overall, the dissertation provides evidence-based design recommendations, empirically validated knowledge on content personalization, and a reusable evaluation workflow for LLM-based systems, thereby advancing research and practice on personalized mental health chatbots for young people.