NurseLLM: Specialized LLM for Nursing
- NurseLLM is a specialized large language model tailored for nursing, emphasizing holistic care, MCQ answering, and documentation through domain-specific adaptations.
- It employs rigorous data engineering techniques including synthetic MCQs, retrieval-augmented generation, and multi-agent orchestration to address nursing-specific challenges.
- Empirical evaluations show that NurseLLM outperforms general-purpose and medical LLMs on dedicated nursing benchmarks, enhancing operational workflows.
Searching arXiv for the core NurseLLM papers and closely related nursing-LLM work. NurseLLM denotes an emerging class of nursing-specialized LLM systems and, more specifically, the model introduced in “NurseLLM: The First Specialized LLM for Nursing” (Khondaker et al., 8 Oct 2025). In the literature, the term spans both a domain-adapted foundation model for nursing multiple-choice question answering and a broader architectural pattern in which LLMs are specialized for nursing documentation, outpatient reception, communication training, patient monitoring, and procedure-centered feedback. What unifies these systems is a shift away from physician-centric medical NLP toward nursing-specific knowledge, workflows, communication styles, and operational constraints (Khondaker et al., 8 Oct 2025).
1. Conceptual scope and disciplinary basis
NurseLLM emerged from the observation that generic medical LLMs are predominantly physician-oriented, whereas nursing practice is holistic and person-centered, emphasizing advocacy, education, psychosocial support, functional status, and coordination of care (Khondaker et al., 8 Oct 2025). The literature explicitly treats this distinction as substantive rather than terminological: nursing reasoning is framed as different from a purely biomedical focus on diagnosis and treatment, and existing medical benchmarks are described as inadequate for systematically probing nursing-specific judgment, care planning, patient education, or NCLEX-style reasoning (Khondaker et al., 8 Oct 2025).
This specialization gap is consequential because there are over 28 million nurses worldwide, yet the major open medical LLMs cited in the nursing literature are tuned primarily on physician-oriented exams, biomedical literature, and clinical notes rather than nursing curricula or licensing frameworks (Khondaker et al., 8 Oct 2025). As a result, current work on NurseLLM has developed along two complementary axes. One axis concerns domain-specialized base models trained or adapted for nursing knowledge tasks, especially MCQ answering. The other concerns workflow-specific systems that place LLMs inside nursing documentation, triage, counseling, education, and simulation pipelines. This suggests that NurseLLM is best understood not as a single architecture but as a nursing-centered model class with multiple deployment forms.
2. Data foundations and specialization strategies
The model-level instantiation of NurseLLM is built by fine-tuning Llama3-Med42-8B on 125K synthetic nursing MCQs with rationales and then applying MergeKit with the DARE merging strategy at a 60% merge proportion to mitigate catastrophic forgetting of Med42’s generic medical capabilities (Khondaker et al., 8 Oct 2025). Its target task is multiple-choice question answering in an NCLEX-like format, with the model expected both to select the correct option and to generate a coherent rationale (Khondaker et al., 8 Oct 2025). This specialization pipeline is paired with new nursing benchmarks intended for rigorous evaluation, and the reported outcome is that NurseLLM outperforms state-of-the-art general-purpose and medical-specialized LLMs of comparable size on those benchmarks (Khondaker et al., 8 Oct 2025).
Across the broader literature, several additional dataset-building strategies recur. “NursingPiles” is described as a multilayer Chinese nursing dataset centered on elderly care, combining 2,777,526 textbook tokens for incremental pre-training, 497,184 tokens of manuals and regulations for RAG, 17,580 QA pairs from research papers, 5,000 QA pairs from nursing safety and ethics materials, 5,000 QA pairs from open-source medical datasets, 1,000,000 simulated nursing dialogues, 500,000 psychology and clinical dialogues, and 2,510 image-text pairs (Sun et al., 2024). In outpatient reception, the Service Flow aware Medical Scenario Simulation framework produces SFMSS-CD with 2,000 first-visit and 400 follow-up dialogues grounded in authentic outpatient records and controlled by nurse, patient, and supervisor agents (Bao et al., 2024). In documentation research, SYNUR contributes 223 synthetic nurse dictations containing approximately 3,000 observations, while the ontology used in its generation expands to 193 concepts after LLM proposal and nurse validation (Corbeil et al., 7 Jul 2025).
| Resource | Purpose | Reported scale |
|---|---|---|
| NurseLLM synthetic nursing MCQ dataset | Nursing MCQ specialization | 125K synthetic nursing MCQs with rationales |
| NursingPiles | IPT, PEFT/SFT, RAG, future multimodality | 2,777,526 textbook tokens; 1,500,000 simulated dialogues; 2,510 image-text pairs |
| SFMSS-CD | Outpatient reception dialogue generation | 2,000 first-visit + 400 follow-up dialogues |
| SYNUR | Nurse dictation structuring | 223 dictations containing ~3000 observations |
These data constructions reveal three dominant specialization strategies. The first is supervised adaptation on nursing-native instructional data. The second is retrieval grounding through manuals, regulations, or ontology schemas. The third is synthetic scenario generation used to compensate for the scarcity and sensitivity of authentic nursing interactions. A plausible implication is that data engineering, rather than model scale alone, is the main bottleneck in NurseLLM development.
3. Architectural patterns
The architecture space associated with NurseLLM is heterogeneous. The base-model line is exemplified by supervised fine-tuning plus model merging on top of Med42-8B for MCQ answering (Khondaker et al., 8 Oct 2025). A second line uses no LLM fine-tuning and instead relies on retrieval-augmented generation, prompt engineering, and ontology constraints. In “Retrieval-Augmented Generation Based Nurse Observation Extraction,” the core formulation is segment-level conditional generation of structured JSON observations from dictation text, with retrieval context comprising ontology schemas and few-shot examples, so that the generator models rather than alone (Hwang et al., 27 Mar 2026).
A third line combines domain adaptation and workflow orchestration. In the Chinese nursing and elderly-care framework, GLM4-Chat 9B and LLaMA 3.1 8B Instruct are adapted through incremental pre-training, parameter-efficient fine-tuning via LoRA, and LangChain-based chains for data extraction, diagnosis triggering, care-plan generation, monitoring, and health education (Sun et al., 2024). A fourth line is explicitly multi-agent: PIORS separates a patient-facing reception nurse, a query-generation module, a summarization module, and a HospInfo-Assistant that converts natural language requests into HIS operations (Bao et al., 2024). A fifth line extends NurseLLM into simulation and pedagogy: Adaptive-VP decomposes the runtime system into Evaluation, Dynamic Adjustment, Dialogue Generation, and Safety Monitoring modules, with patient behavior adapting to trainee communication quality (Lee et al., 31 May 2025). A sixth line is multimodal and embodied: RoboNurse-VLA treats the scrub nurse role as a Vision-Language-Action problem integrating YOLOv8, SAM 2, Llama 2, and discretized action tokens for robotic grasp-and-handover (Li et al., 2024).
Taken together, these works indicate that NurseLLM is not tied to a single training recipe. It may take the form of a decoder-only transformer specialized through SFT, a training-free RAG system, a LangChain-orchestrated assistant, a multi-agent HIS-connected receptionist, an adaptive virtual patient environment, or a VLA robotic controller. This suggests that “NurseLLM” functions as an application-layer category defined by nursing alignment rather than by one canonical network design.
4. Core application domains
A central application domain is structured documentation from speech. In the MEDIQA-SYNUR setting, nurse dictations are segmented into contiguous “clinical fact” units by GPT-5-mini, then each segment is enriched by top-10 ontology schemas retrieved with a BlueBERT and TF-IDF hybrid and top-15 semantically similar examples retrieved with a BlueBERT and BM25 hybrid, after which GPT-5.1 generates JSON observations constrained to the retrieved schema (Hwang et al., 27 Mar 2026). The output format uses fields such as id, name, value_type, and value, and hallucinations are reduced by restricting the observation space to the retrieved schema and enumerated value options (Hwang et al., 27 Mar 2026). A closely related study on structuring speech transcripts extends this logic to hospital flowsheet mapping and releases SYNUR as an open dataset for nurse observation extraction (Corbeil et al., 7 Jul 2025).
A second domain is dynamic nursing and elderly-care assistance. The Chinese framework positions the LLM as a knowledge engine, patient interaction agent, decision support tool, and core reasoning module in a dynamic nursing assistant that collects patient data, triggers nursing diagnoses, generates personalized interventions, and adjusts plans over time through feedback (Sun et al., 2024). Its workflow explicitly includes IoT or API-based data collection, natural-language data processing, diagnosis triggering, care-plan generation, continuous monitoring, stage transitions, follow-up education, and encrypted data management (Sun et al., 2024).
A third domain is outpatient reception and pre-consultation triage. PIORS places an LLM-based reception nurse in front of the doctor, with duties that include department guidance, pre-diagnosis information gathering, administrative consultation, and personalized service, while a collaborating HospInfo-Assistant handles patient archive operations and information retrieval through the hospital information system (Bao et al., 2024). The system therefore treats NurseLLM not as a standalone chatbot but as a reception workflow component integrated with institutional infrastructure (Bao et al., 2024).
A fourth domain is communication training and educational simulation. Adaptive-VP creates clinically grounded virtual patients whose communication style, complaint intensity, and responsiveness change as a function of trainee performance, allowing hostility or uncooperativeness to escalate when the nurse responds ineffectively and to soften when the nurse uses calm, empathic, or de-escalatory strategies (Lee et al., 31 May 2025). Related educational work uses a unified XAI-LLM pipeline for endotracheal suctioning activity recognition and feedback generation from video, with Gemini 2.5 Pro as a central reasoning engine for spatiotemporal labeling and explanation (Phan et al., 29 Jan 2026).
A fifth domain is reflective and narrative documentation. Work on ICU diary writing does not present a complete implementation, but it frames LLMs as collaborative writing partners that can help nurses decide what to write, translate clinical events into everyday language, align with prior diary style, and potentially retrieve contextual information such as recent documentation or non-sensitive external context (Freire et al., 2024). This suggests that NurseLLM research also includes narrative, relational, and psychosocial nursing tasks, not only extraction and decision support.
5. Empirical performance and benchmark evidence
The strongest direct benchmark claim for the specialized base model is that NurseLLM outperforms state-of-the-art general-purpose and medical-specialized LLMs of comparable size on newly introduced nursing benchmarks (Khondaker et al., 8 Oct 2025). In documentation extraction, the RAG-based nurse observation system reports a development-set improvement from F1 with few-shot examples only to F1 when ontology schema retrieval is added, and its best MEDIQA-SYNUR test configuration—GPT-5.1 as generator with GPT-5-mini as segmentor—achieves precision , recall , and F1 (Hwang et al., 27 Mar 2026). In the related speech-structuring study, few-shot prompting and RAG yield F1 scores in the high 80s to low 90s on proprietary hospital datasets and SYNUR, with GPT-4.1 reaching 92.4 average few-shot F1 across H1, H2, H3, and SYNUR (Corbeil et al., 7 Jul 2025).
The Chinese nursing and elderly-care framework reports that GLM4 + IPT + SFT reaches precision , recall , F1 , and accuracy 0 on nursing exam benchmarks, outperforming GPT-4o on those metrics (Sun et al., 2024). PIORS-Nurse reports automatic evaluation results of accuracy 1, Overall Score 2, Info Score 3, average turn number 4, and average turn length 5, and it is described as outperforming GPT-4o, Qwen2-7B, Llama3-8B, HuaTuoGPT2-13B, and an SF-ablated fine-tuned baseline in that outpatient reception setting (Bao et al., 2024). In communication training, Adaptive-VP shows that its LLM-based communication skill evaluation distinguishes experienced from novice nurses with Mann–Whitney 6 and 7, and that dynamic VPs significantly improve Role Fidelity and Conversational Realism over static VPs in linear mixed-effects analyses (Lee et al., 31 May 2025). In procedural education, the endotracheal suctioning framework reports mean accuracy 8 and mean F1 9 for Prompt A, compared with accuracy 0 and F1 1 for the Ngo et al. baseline, with statistically significant improvements (Phan et al., 29 Jan 2026).
These results collectively indicate that nursing specialization can be operationalized in several ways: benchmark-focused model adaptation, schema-constrained RAG, task-grounded orchestration, and multimodal simulation. They also show that evaluation in this area is unusually heterogeneous, spanning MCQ accuracy, extraction F1, rubric-based human judgments, interaction efficiency, and activity-recognition metrics.
6. Limitations, safety, and future directions
The literature repeatedly emphasizes that present NurseLLM systems remain bounded by data availability, deployment constraints, and safety requirements. In observation extraction, the MEDIQA-SYNUR work notes data consistency issues such as “37.5℃” sometimes being represented as “375,” dependence on ontology and memory-bank coverage, lack of explicit temporal reasoning, and uncertainty about generalization from benchmark data to real hospital dictations that may contain different styles, noise, and PHI (Hwang et al., 27 Mar 2026). The SYNUR work likewise stresses that synthetic dictations do not fully capture local jargon or full flowsheet complexity, and that hallucinated or omitted observations remain a deployment risk (Corbeil et al., 7 Jul 2025).
At the system level, many NurseLLM implementations are geographically and institutionally narrow. NursingPiles is Chinese-focused, PIORS is trained on data from a single hospital context in China, and Adaptive-VP is validated in a South Korean communication-training setting rather than across multilingual or multinational nursing environments (Sun et al., 2024). Several systems depend on proprietary models or infrastructures, including GPT-5 variants, GPT-4o, Claude-3.5 Sonnet, and Gemini 2.5 Pro, which complicates reproducibility and on-premise deployment (Hwang et al., 27 Mar 2026). The ICU diary paper adds a distinct socio-technical caution: even when LLMs can reduce time and cognitive barriers, they raise questions about authenticity, authorship, hallucination, privacy, and the effect of machine assistance on nurse–patient–family relationships (Freire et al., 2024).
Future work in this area points toward more explicit multi-agent specialization, broader real-world validation, and stronger governance. The NurseLLM paper itself explores the role of reasoning and multi-agent collaboration systems in nursing (Khondaker et al., 8 Oct 2025). Outside the nursing-specific model paper, expertise-aware multi-LLM recruitment has been proposed for medical decision-making through an expertise table, dynamic selection of models by department and difficulty, and confidence- plus adversarial-driven collaboration (Bao et al., 19 Aug 2025). This suggests that future NurseLLM systems may combine nursing-specialized base models with expertise-aware orchestration, retrieval-grounded local protocols, and human-in-the-loop review. The recurrent conclusion across the literature is that nursing specialization improves relevance, but safe deployment still depends on auditability, constrained outputs, privacy-preserving infrastructure, and explicit accommodation of the distinctive epistemic and relational character of nursing work.