Decision support value of locally deployed large language model for personalized VTE prevention nursing in orthopedic postoperative patients
基于院内本地部署大语言模型的骨科术后VTE预防个性化护理决策辅助研究

摘要

目的 探究基于院内本地部署架构的大语言模型（LLMs）——素问医疗大模型，在骨科术后静脉血栓栓塞症（VTE）预防护理决策辅助里的应用价值，评估低年资护士护理决策正确率与护理计划措施个性化实施水平的成效。方法 回顾性选取2025年1月至12月十堰市太和医院骨科收治的择期关节置换术或者需要做骨折内固定术的60例患者病历，以骨科科室制定的个性化VTE预防护理措施为对照标准。选取2名工作年限≤3年的低年资骨科护士，借助素问医疗大模型进行VTE风险个性化评估、预防方案推荐、护理问题识别，记录决策结果与制定方案耗时；同时选取2名工作年限≥10年的高年资护士不使用模型完成相同工作。对比低年资护士运用模型前后所做出的决策是否正确、其个性化响应能力如何以及决策所需的时间以及主观满意度评分。结果 应用模型后，低年资护士个性化决策正确率由51.7%提升至78.3%（P<0.05），个性化响应度由41.2%提升至73.5%（P<0.01），总体决策正确率由58.3%提升至81.7%（P<0.05）；平均决策耗时由（9.2±2.4）min缩短至（4.8±1.6）min（P<0.01）。护士满意度评分值为（4.5±0.5）分，其中95%的护士表示，该模型有临床辅助价值。结论 院内本地部署的素问医疗大模型，能显著提升低年资骨科护士VTE预防护理决策的准确性与效率，从而提高护理方案个性化制定水平。本地部署模式能很好地保障医疗数据的安全，给大语言模型在骨科专科护理里安全地落地以及个性化应用提供参考。

Abstract

Objective To investigate the value of a locally deployed hospital-based large language model (LLM)—the Suwen Medical LLM—in assisting nursing decisions for the prevention of venous thromboembolism (VTE) after major orthopedic surgeries, and to evaluate its impact on the accuracy of nursing decisions made by junior nurses and the level of personalization in nursing care plans.
Methods A retrospective analysis was conducted on medical records of 60 patients who underwent elective joint arthroplasty or internal fixation of fractures in an orthopedic department of a tertiary hospital from January to December 2025. The orthopedic department formulated personalized VTE prevention nursing decisions, which were used as the control standard. Two junior orthopedic nurses (working experience ≤3 years) used the Suwen Medical LLM to perform three tasks: personalized VTE risk assessment, recommendation of personalized prevention plans, and identification of personalized nursing issues. The outcomes of the decision and the time spent on planning were recorded. Meanwhile, two senior nurses (working experience ≥10 years) performed the same tasks without using the model. Before and after the model was used by junior nurses, we compared the accuracy of decisions, the level of personalization, the decision-making time, and the subjective satisfaction scores.
Results After using the model, the accuracy of personalized decisions made by junior nurses increased from 51.7% to 78.3% (P<0.05), the level of personalization increased from 41.2% to 73.5% (P<0.01), and the overall decision accuracy increased from 58.3% to 81.7% (P<0.05). The average decision-making time decreased from (9.2±2.4) minutes to (4.8±1.6) minutes (P<0.01). The nurse satisfaction score was (4.5±0.5), and 95% of the nurses considered the model clinically valuable.
Conclusion   The locally deployed Suwen Medical LLM significantly improves the accuracy and efficiency of VTE prevention nursing decisions made by junior orthopedic nurses, as well as the level of personalization in nursing plans. The local deployment model effectively safeguards the security of medical data, offering a reference for the safe and personalized application of LLMs in orthopedic nursing.