JMIR Publications

ABSTRACT

Background:

Simulated patients (SPs) are essential in nursing education but face persistent challenges: trained-actor scarcity, instructor burden, cost, and disrupted continuity during pandemics or armed-conflict environments where in-person simulation cannot be sustained. Recent LLM-based simulated patient systems address some of these issues, yet existing studies focus exclusively on the patient role, leaving the dialogic partner—the AI nurse—as a static prompt or unspecialized model. This asymmetry compromises both data-collection quality and pedagogical applicability.

Objective:

We report (a) the construction of two specialized fine-tuned models, NURSE-ONE (AI nurse) and PATIENT-ONE (AI simulated patient), through a five-step bidirectional fine-tuning pipeline; (b) an evaluation of the resulting models using Berelson content analysis; and (c) the design and analysis of a coexistence simulation ecosystem in which the two models operate jointly to support three pedagogical operating modes. We additionally examine the system's applicability to disaster-preparedness nursing education.

Methods:

On a single consumer-grade GPU (NVIDIA RTX 4090, 24 GB VRAM), the multimodal base model gemma4-E4B-it was fine-tuned via Low-Rank Adaptation (LoRA) under a five-step pipeline: (1) construction of a 251-scenario base from 29 institutionally-authored standard nursing care plans and 20 situational items from the 115th Japanese National Nursing Examination; (2) provisional dialogue system using AutoGen v18; (3) NURSE-ONE fine-tuning with three training-volume conditions; (4) four-condition Berelson comparison; (5) PATIENT-ONE fine-tuning using NURSE-ONE-generated dialogue data. All adapters were merged using PEFT, converted to GGUF q4_k_m (5.07 GB per role), and served via Ollama. Three coexistence operating modes were operationalized in a working web UI.

Results:

Four-condition Berelson analysis (κ=0.742, χ²=41.991, df=15, p=0.0002) showed that the 735-case nurse-side condition achieved the highest open-question rate (OQ=21.5%) and was adopted as NURSE-ONE. PATIENT-ONE A/B/C Berelson analysis (κ=0.718, χ²=11.818, p=0.297) showed no significant category-distribution shift; the 735-case condition achieved the lowest response-length variability and highest information-collection rate (47.9%). Across both experiments, nurse-side LoRA modulated qualitative questioning style while patient-side LoRA controlled quantitative response stability—a functional asymmetry operating on orthogonal axes. All three coexistence modes were instantiated within the pre-registered latency threshold (≤3 s per turn).

Conclusions:

A five-step bidirectional fine-tuning pipeline produced two specialized models whose joint deployment constitutes a novel design pattern for nursing-education simulation. The functional asymmetry between qualitative nurse-side variation and quantitative patient-side stabilization underwrites the dyadic system's coherence; the resulting ecosystem supports three pedagogically distinct operating modes on a single shared substrate, runs entirely on consumer-grade hardware in Japanese, and is deployable in offline disaster-preparedness contexts where cloud-dependent alternatives are unusable.