Secure and accessible AI in nursing education: a modular agentic chatbot framework comparing ChatGPT-4o with an open-source LLM for chronic wound care

Abstract

To develop and evaluate a modular agentic AI chatbot framework for chronic wound care education in nursing and to compare open-source and commercial configurations. Background: Integrating artificial intelligence (AI) into nursing education can enhance learning experiences, but widespread adoption is limited by costs, data governance and concerns about model reliability. Modular, retrieval-augmented generation (RAG)-based agentic AI frameworks might address these limitations by improving accuracy, flexibility and local control. Design: A comparative experimental pre-implementation study using a validated 100-item chronic wound care question dataset aligned with international clinical guidelines. Methods: A modular, agentic chatbot framework employing RAG technology was developed and evaluated using open-source and commercial AI models. Six experts independently rated responses on accuracy, relevance, clarity and coverage using a 5-point Likert scale. Analyses included linear mixed-effects modeling and Wilcoxon signed-rank tests. Dataset validity was confirmed using the Davis technique. Results: Both chatbot configurations produced clinically accurate, guideline-aligned responses. The commercial model achieved higher overall scores (mean difference = +1.01 on a 20-point scale, p < 0.001), with the largest domain-specific difference observed in coverage (+0.39 on a 5-point scale, p < 0.001). The open-source configuration demonstrated strong adherence to guidelines, with 56% of responses fully aligned with clinical recommendations. Conclusions: The open-source, on-premises AI agent demonstrated near-comparable performance to ChatGPT-4o while offering clear advantages in cost, security and institutional autonomy. Its capacity to support guideline-based instruction and promote equitable access makes it a promising tool for nursing education.