A domain-specific large language model for evidence-based nursing: development and multi-method evaluation of EviNurse

Large language models (LLMs) show promise for improving access to healthcare evidence, yet general-purpose LLMs often lack domain-specific knowledge to adequately support evidence-based nursing practice. We developed and evaluated EviNurse, a domain-specific LLM designed to facilitate timely access to high-quality nursing evidence and support evidence-based decision-making. We constructed NursData, a high-quality evidence-based nursing dataset organized according to the 5S evidence pyramid to support the training, retrieval, and evaluation of EviNurse. Based on Qwen3-32B, EviNurse was developed using supervised fine-tuning and retrieval-augmented generation to integrate domain-specific nursing evidence. EviNurse was evaluated using 3438 multiple-choice questions, 25 expert-rated short-answer questions, and a usability evaluation involving 316 nurses across all provinces in mainland China. EviNurse achieved 91.45% accuracy on the multiple-choice benchmark, ranking second among the evaluated models. In expert evaluation, EviNurse significantly outperformed general-purpose LLMs overall (P < 0.001), with particularly strong performance in reliability and satisfaction and across surgery, pediatrics, and emergency and critical care. Usability was favorable across all six dimensions, with mean scores ranging from 1.67 to 1.97. These findings suggest that integrating curated nursing evidence with domain adaptation and retrieval augmentation can enhance LLM performance in nursing and facilitate access to high-quality evidence for evidence-based nursing practice.