基于生命体征时序数据的创伤致死性大出血伤情动态预测模型开发及验证OA北大核心CSTPCD
Development and validation of dynamic prediction models using vital signs time series data for fatal massive hemorrhage in trauma
目的 基于生命体征时序数据和机器学习算法建立创伤致死性大出血伤情动态预测模型.方法 回顾性分析重症监护医疗信息(MIMIC-Ⅳ)数据库2008-2019年7522例创伤伤员的生命体征时序数据,并按照创伤后是否发生致死性大出血事件分为致死性大出血组(n=283)与非致死性大出血组(n=7239).采用逻辑回归(LR)、支持向量机(SVM)、随机森林(RF)、自适应提升(AdaBoost)、门控循环单元(GRU)、门控循环单元-D(GRU-D)共6种机器学习算法开发创伤致死性大出血伤情动态预测模型,对创伤伤员未来T小时(T=1、2、3)发生致死性大出血伤情的风险进行动态预测.通过准确率、敏感度、特异度、阳性预测值、阴性预测值、约登指数以及受试者工作特征(ROC)曲线下面积(AUC)评估模型性能.基于解放军总医院创伤数据库对模型进行外部验证.结果 MIMIC-Ⅳ数据集中,基于GRU-D算法开发的一组动态预测模型效果最优,预测未来1、2和3 h发生致死性大出血的AUC分别为0.946±0.029、0.940±0.032和0.943±0.034,且差异无统计学意义(P=0.905).创伤数据集中,GRU-D模型取得了最佳外部验证效果,预测未来1、2和3 h发生致死性大出血的AUC分别为0.779±0.013、0.780±0.008和0.778±0.009,且差异无统计学意义(P=0.181).该组模型已部署在公开的网页计算器和医院急诊科信息系统中,便于公众和医护人员使用和验证.结论 成功开发并验证了一组动态预测模型,可对创伤致死性大出血伤情进行早期诊断和动态预测.
Objective To establish a dynamic prediction model of fatal massive hemorrhage in trauma based on the vital signs time series data and machine learning algorithms.Methods Retrospectively analyze the vital signs time series data of 7522 patients with trauma in the Medical Information Mart for Intensive Care-Ⅳ(MIMIC-Ⅳ)database from 2008 to 2019.According to the occurrence of posttraumatic fatal massive hemorrhage,the patients were divided into two groups:fatal massive hemorrhage group(n=283)and non-fatal massive hemorrhage group(n=7239).Six machine learning algorithms,including logistic regression(LR),support vector machine(SVM),random forests(RF),adaptive boosting(AdaBoost),gated recurrent unit(GRU),and GRU-D were used to develop a dynamic prediction models of fatal massive hemorrhage in trauma.The probability of fatal massive hemorrhage in the following 1,2,and 3 h was dynamically predicted.The performance of the models was evaluated by accuracy,sensitivity,specificity,positive predictive value,negative predictive value,Youden index,and area under receiver operating characteristic curve(AUC).The models were externally validated based on the trauma database of the Chinese PLA General Hospital.Results In the MIMIC-Ⅳ database,the set of dynamic prediction models based on the GRU-D algorithm was the best.The AUC for predicting fatal major bleeding in the next 1,2,and 3 h were 0.946±0.029,0.940±0.032,and 0.943±0.034,respectively,and there was no significant difference(P=0.905).In the trauma dataset,GRU-D model achieved the best external validation effect.The AUC for predicting fatal major bleeding in the next 1,2,and 3 h were 0.779±0.013,0.780±0.008,and 0.778±0.009,respectively,and there was no significant difference(P=0.181).This set of models was deployed in a public web calculator and hospital emergency department information system,which is convenient for the public and medical staff to use and validate the model.Conclusion A set of dynamic prediction models has been successfully developed and validated,which is greatly significant for the early diagnosis and dynamic prediction of fatal massive hemorrhage in trauma.
郭程娱;龚明慧;沈翘楚;韩辉;王若琳;张红亮;王俊康;李春平;黎檀实
南开大学医学院,天津 300071||解放军总医院第一医学中心急诊科,北京 100853清华大学软件学院,北京 100083解放军总医院第一医学中心急诊科,北京 100853
临床医学
创伤大出血机器学习辅助诊断
wounds and injuriesmassive hemorrhagemachine learningassistant diagnosis
《解放军医学杂志》 2024 (006)
629-635 / 7
This work was supported by the National Key Research and Development Program of China(2020YFC1512702) 国家重点研发计划(2020YFC1512702)
评论