Coexistence of AI Nurse and AI-Simulated Patient in Nursing Education Using a Bidirectional Fine-Tuning Pipeline: Development and Content Analysis Study
JMIR Biomedical Engineering · 2026-11-06 · 11 · e100778-e100778
DOI: 10.2196/100778 ↗Abstract
Abstract Background Simulated patient systems are essential in nursing education but face actor scarcity, instructor burden, cost, and disrupted continuity where in-person simulation cannot be sustained. Existing large language model (LLM)–based systems address these issues but treat the dialogic partner—the AI nurse—as a static prompt or unspecialized model, compromising data collection quality and pedagogical applicability. Objective We report (1) the construction of 2 specialized fine-tuned models, NURSE-ONE (AI nurse) and PATIENT-ONE (AI-simulated patient), via a 5-step bidirectional fine-tuning pipeline; (2) their evaluation using Berelson content analysis validated by independent human coders; (3) an exploratory third-party evaluation by nursing faculty; and (4) the design of a coexistence simulation ecosystem supporting 3 operating modes, including offline disaster preparedness applicability. Methods On a single consumer-grade graphics processing unit (NVIDIA RTX 4090; 24 GB video random access memory), the multimodal base model gemma4-E4B-it was fine-tuned via low-rank adaptation in five steps: (1) a 251-scenario base from 29 institutionally authored nursing care plans and 20 items from the 115th Japanese National Nursing Examination; (2) a provisional dialogue system (AutoGen v18, maxmsg=25) to collect training dialogues; (3) NURSE-ONE fine-tuning across 3 training volume conditions (245, 735, and 1225 cases); (4) 4-condition Berelson comparison; and (5) PATIENT-ONE fine-tuning on NURSE-ONE–generated dialogues with a 3-condition comparison. Interrater reliability was confirmed by 2 independent nursing experts (196 blinded utterances, κ=0.770) and corroborated by 2 LLM raters (κ=0.742). Ten nursing faculty evaluated 3 coexistence modes in a working web user interface (21-item rubric; item-level content validity index=1.00). Results Four-condition Berelson analysis ( χ 2 15 =42.0, P
EUVIMED