JMIR Publications

ABSTRACT

Background:

LLMs are increasingly being explored as new tools to support health information seeking. However, stroke survivors may face unique challenges when using digital health technologies due to stroke-related functional impairments. Understanding how stroke survivors interact with LLM systems and adapt AI-generated information to their own health contexts is important for developing more inclusive and patient-centered AI health tools.

Objective:

To explore how stroke survivors adaptively use large language models (LLMs) for health information seeking and how they understand and apply AI-generated information within their functional, health, and social contexts.

Methods:

In this qualitative study, 11 stroke survivors were recruited. Participants performed standardized and personalized health information-seeking tasks using three LLMs (DouBao, DeepSeek, and AQ), followed by evaluation of AI-generated responses via the CLEAR scale. Data collection integrated a think-aloud approach with semi-structured interviews, and inductive thematic analysis was applied for data analysis.

Results:

CLEAR scores indicated generally high evaluations of LLM-generated health information, with researcher ratings slightly higher than participant ratings. Think-aloud findings revealed different processing patterns across tasks: participants primarily verified AI information in the standardized task, whereas the personalized task involved greater contextual adaptation and application, with participants progressively supplementing personal health information. Overall, LLM-based health information seeking among stroke survivors was a process of continuous adaptation. Participants adjusted their interaction approaches according to their functional abilities and progressively supplemented health contexts and clarified their needs during interactions. They also evaluated AI outputs based on credibility cues, uncertainty, personal health conditions, and potential risks. Whether AI-generated recommendations could be translated into actual actions depended on their alignment with patients’ functional abilities and real-life circumstances. Furthermore, this process involved dynamic collaboration among patients, family members, and AI systems.

Conclusions:

Stroke survivors’ use of large language models for health information seeking is a process of continuous adaptation across interaction, evaluation, and action. Future generative AI systems should tailor interaction modalities and the difficulty of recommendations to patients’ functional abilities, clearly communicate uncertainty and applicability boundaries for high-risk health information, and support family collaboration with patient authorization, thereby helping transform AI-generated health information from being merely accessible to being safely usable and practically actionable.