JMIR Publications

ABSTRACT

Background:

The rapid integration of artificial intelligence (AI) into healthcare has outpaced formal training in health professions education. Generative AI chatbots — including simulated patients, clinical coaches, and formative assessment tools — offer scalable, interactive learning experiences that do not require programming expertise. Despite growing interest, many health educators lack a structured framework for designing and deploying these tools effectively.

Objective:

This tutorial presents a five-section, sequentially ordered workflow to guide health professions educators through the design, build, testing, and iterative refinement of no-code AI chatbots using generative AI platforms.

Methods:

The workflow was developed through iterative faculty development activities at the Yong Loo Lin School of Medicine, National University of Singapore, and talks given at USA, China and Malaysia, drawing on two chatbot initiatives: the Virtual Integrated Patient (VIP), a free-text conversational platform for clinical history-taking practice; and the Depression Avatars project, in which ChatGPT-4 was used to generate clinically accurate patient scripts and D-ID software was used to produce photorealistic AI avatar videos of patients with psychiatric presentations including low mood and chest tightness. The workflow synthesises design principles from educational theory — including the TPACK framework and feedback science — with practical experience configuring no-code LLM platforms.

Results:

Preliminary evaluation of the Depression Avatars project found that medical students responded positively to AI avatar-based psychiatric patient simulations, with the majority rating them as more engaging and educationally useful than traditional formats. Qualitative evaluation of the VIP platform (n=8) identified flexibility and self-directed practice as key strengths, while emotional realism and interactivity were noted as areas for improvement. These findings directly informed the workflow sections addressing persona design, graduated disclosure, multimodal avatar integration, and pilot testing.

Conclusions:

No-code AI chatbots are feasible, acceptable, and potentially effective tools for health professions education when designed with clear pedagogical intent. The five-section workflow — from defining purpose and persona to embedding ethics — provides a practical, platform-agnostic framework for health educators. While the primary audience is health professions education, the workflow is transferable across higher education disciplines. The tools described will evolve; the design principles will not.