JMIR Publications

ABSTRACT

Digital behavioral interventions are increasingly used to support chronic disease self-management, yet many systems rely on predetermined content that limits personalization and sustained engagement. Large language models (LLMs) offer new opportunities to deliver conversational behavioral support. However, integrating LLMs into behavioral interventions requires careful architectural, methodological, and ethical planning, which may be challenging for researchers without formal training in artificial intelligence. This viewpoint provides a structured introduction to LLMs tailored to behavioral science. We describe foundational concepts in natural language processing and transformer-based architectures, outline the core components of LLM-based systems, including prompting strategies, context management, retrieval-augmented generation, and guardrails, and illustrate these principles through our experience integrating a proprietary LLM into a mobile self-management intervention for individuals with systemic sclerosis. Building on this case example, we propose a phased design workflow to guide early-stage development and responsible implementation, along with a decision framework to help researchers navigate scientific and logistical trade-offs between proprietary models and other alternatives. The considerations presented here are informed by formative implementation efforts and are intended to support early-stage design decisions for LLM-based behavioral interventions. As these interventions continue to evolve, rigorous evaluation and interdisciplinary collaboration will be important to ensure that these systems improve personalization and scalability while maintaining safety and scientific rigor.