摘要
目的验证并评价垂类大模型(通义千问Qwen2模型)预测重症监护病房(ICU)患者死亡风险的精准性,以探讨该模型的临床应用潜力。方法选取某三级甲等医院ICU收治的3526例患者作为研究对象,利用患者的临床数据、电子病历和实验室检查指标,采用Qwen2模型对死亡风险进行预测。通过大模型实现文本数据与结构化指标的特征融合,并进行模型的微调训练,从而提升预测性能。结果在死亡风险预测任务中,Qwen2模型准确率为96.68%,曲线下面积为87.93%,灵敏度为75.60%,特异性为99.84%,F1得分为96.49%。结论Qwen2模型能够有效预测ICU患者的死亡风险,具备较强的临床应用潜力,未来可通过多中心数据验证进一步提高模型的泛化能力。
Objective This study aims to exam and evaluate the accuracy of the mortality probability prediction of ICU patients using the vertical large language model(Qwen2),and to explore the potential for clinical application.Methods A total of 3526 patients from a tertiary hospital ICU in Sichuan.Data such as clinical information,electronic medical records,and laboratory test results were used to train the Qwen2 model for mortality probability prediction.Feature fusion of text data and structured indicators were achieved through the large language model,and fine-tuning training of the model was completed,improving its prediction performance.Results The Qwen2 model demonstrated excellent performance in predicting mortality probability,with an accuracy of 96.68%,AUC of 87.93%,sensitivity of 75.60%,specificity of 99.84%,and F1 score of 96.49%.Conclusion The Qwen2 model shows strong potential for clinical application,especially in mortality probability prediction.Model generalization can be enhanced by validating it with multi-center datasets.
作者
谢佳希
郑兵
杨豪
董婧
石锐
XIE Jiaxi;ZHENG Bing;YANG Hao;DONG Jing;SHI Rui(West China Hospital of Sichuan University,Chengdu 610041,Sichuan,China)
出处
《中国卫生信息管理杂志》
2025年第2期195-201,220,共8页
Chinese Journal of Health Informatics and Management
基金
四川省重大科技专项“医学诊疗大模型关键技术研发及应用示范”(2024ZDZX0017)。
关键词
垂类大模型
Qwen2
死亡风险
ICU
预测
vertical large language models
Qwen2
mortality probability
ICU
prediction
