From aligned models to trusted interfaces : explainable health intervention and transparent health information seeeking

This thesis presents research on two interrelated themes in applying large language models (LLMs) to health contexts: (1) the model theme, aligning LLMs with domain expertise for digital counseling. (2) the interface theme, exploring people’s trust in LLM-powered health information. Specifically, the first theme focuses on improving the explainability and controllability of LLMs for therapeutic dialogue generation. Initial evaluation showed that LLMs struggle to produce emotionally nuanced and contextually appropriate reflections in Motivational Interviewing (MI). To address this, we created expert-annotated bilingual MI dataset that captures therapeutic communication strategies. Building on this, a Script-Strategy Aligned Generation (SSAG) was proposed, where the LLM first predicts a therapeutic strategy before generating dialogue. This structured approach enhanced both explainability and adherence to evidence-based principles, offering a more controllable way to integrate LLMs into digital psychotherapy. Research on the second theme explored trust in LLM-powered health information seeking. The first study compared LLM-powered conversational search with use of a traditional web-based search engine. The results showed that participants express more trust in LLM’s conversational responses. Research also revealed a preference for text-based interfaces over speech-based and embodied interfaces. Crucially, a mixed-methods study on source attribution showed that while LLM-generated information was highly trusted, content explicitly labeled as human-sourced was perceived as more trustworthy than content labeled as AI-generated, regardless of its actual origin. These findings underscore the nuanced relationship between source transparency and trust. By bridging model alignment with human-centric assessment, this work provides new knowledge for designing explainable, transparent, and trustworthy LLM-powered applications in digital health.

• View at Worldcat

Free Full Text ( Final Version , 72mb )