(PI-029 (RPT-003)) Secure and Accessible Agentic AI for Chronic Wound Care Education: A Comparative Evaluation of Open-Source and Proprietary Language Models

Necip Gürler, PhD – eKare, Inc.; Samet Hüsnü Bülbül, MSc – eKare, Inc.; Dilek Yilmaz Akyaz, RN, Msc, PhD(c) – Graduate School of Health Sciences, Koç University; Özgür Güler, PhD – eKare, Inc.; Asım Evren Yantaç, PhD, Assoc. Prof. – Koç University; Emir Bedirhan Aydın, Undergraduate Student – Koç University; Holly Kirkland-Kyhn, PhD, FNP-c, GNP-c, CWCN, FAANP – Uc Davis, Betty Irene School Of Nursıng

Introduction:

Artificial intelligence (AI)–supported educational tools are increasingly used in nursing education to promote self-directed learning, clinical reasoning, and guideline-based decision-making. However, widespread adoption remains limited due to concerns related to cost, data privacy, institutional autonomy, and the reliability of AI-generated content. These challenges are particularly relevant in chronic wound care education, a clinically critical yet underrepresented domain within undergraduate nursing curricula. Retrieval-augmented generation (RAG)–based agentic AI systems offer a potential solution by grounding responses in verified clinical guidelines while allowing flexible, locally controlled deployment. This study aimed to develop and evaluate a modular agentic AI chatbot framework for chronic wound care education and to compare the educational performance of open-source and proprietary large language model configurations.

Methods:

A comparative experimental study design was employed. A modular, RAG-based agentic AI chatbot framework was developed using curated international chronic wound care guidelines. A validated dataset of 100 wound care–specific questions, spanning multiple cognitive levels, was independently submitted to two chatbot configurations: an open-source, on-premises model and a proprietary, cloud-based model. Six wound care experts evaluated the generated responses under blinded conditions using a structured rubric assessing accuracy, relevance, clarity, and content coverage on a 5-point Likert scale. Descriptive and comparative statistical analyses were conducted to examine differences in response quality between configurations.

Results:

Both chatbot configurations generated clinically accurate and guideline-aligned responses across all evaluation domains. The proprietary configuration demonstrated slightly higher overall performance, with the largest difference observed in content coverage. Nevertheless, the open-source configuration achieved strong guideline adherence, with more than half of responses demonstrating complete alignment with reference standards. Differences between configurations were consistent across expert raters and question types, indicating stable comparative performance.

Discussion:

Findings suggest that a modular, agentic AI chatbot framework can effectively support chronic wound care education. Although proprietary models demonstrated marginally higher overall scores, the open-source configuration achieved near-comparable educational performance while offering key advantages in cost, data security, and institutional control. These results highlight the feasibility of secure, locally deployable AI systems as scalable and equitable educational tools for nursing programs, particularly in resource-constrained settings.