摘要
目的建立并验证基于机器学习算法的大面积烧伤患者新发脓毒性休克预测模型,以提升风险识别性能。方法回顾性分析单中心2012至2023年间收治的大面积烧伤队列的临床数据,按7∶3随机划分训练集和验证集,使用最小绝对值选择与收缩算子回归算法筛选预测因子,在训练集中分别使用支持向量机(SVM)、logistic回归(LR)、随机森林(RF)和决策树(DT)四种机器学习算法构建新发脓毒性休克预测模型,并在验证集中评估最优模型性能。结果共纳入220例大面积烧伤病人,其中训练集和验证集分别为154和66人,共计32人合并新发脓毒性休克。在训练集中,筛选得到烧伤占总体表面积百分比、谷丙转氨酶、血清肌酐、血钠、降钙素原、热损伤类型、缺乏液体复苏以及血培养阳性8项新发脓毒性休克的预测因子。在验证集中,随机森林模型预测区分度最佳,AUC为0.92(0.83~1.00),优于SOFA评分[0.86(0.75~0.97)],且模型通过校准度检验。结论基于随机森林算法的新发脓毒性休克预测模型表现出良好的预测区分度和校准度,在早期识别脓毒性休克和风险分层应用中成为有潜力的辅助决策工具。
Objective To develop and validate a novel machine-learning-based prediction model for new-onset septic shock in patients with extensive burns,aiming to improve the performance of risk stratification.Methods A retrospective analysis of the clinical data of patients with extensive burns admitted between 2012 and 2023 was conducted.The patients were randomly divided the training set and validation according to 7∶3 Set.Predictors were screened by minimum absolute value selection and theshrinkage operator regression algorithms.The four machine learning algorithms,i.e.,Support Vector Machine(SVM),Logistic Regression(LR),Random Forest(RF)and Decision Tree(DT),were used in the training set to develop the new-onset septic shock prediction model.The performance of the model was evaluated in the validation set.Results A total of 220 patients with extensive burns were included,154 in the training set and 66 in the validation set.Among them,32 patients developed new-onset septic shock.Eight predictive factors for outcome were identified in the training set,including total burn surface area,aspartate transaminase,serum creatinine,serum sodium,procalcitonin,type of thermal injury,insufficient fluid resuscitation,and positive blood culture.In the validation set,the RF model had the best prediction discrimination,with an AUC of 0.92(0.83-1.00),which was better than the SOFA score 0.86(0.75-0.97),and the model passed the calibration test.Conclusion The new-onset septic shock prediction model based on the RF algorithm has been established and validated.The model showed excellent discriminative and calibration performance,making it a promising adjunctive decision support tool for early septic shock identification and risk stratification.
作者
邱侃
李江虹
叶勇
王贤正
谭谦
QIU Kan;LI Jianghong;YE Yong;WANG Xianzheng;TAN Qian(Medical School of Nanjing University,Nanjing 210093,Jiangsu,China;Department of Burns and Plastic Surgery,Anqing Petrochemical Hospital of Nanjing Drum Tower Hospital Group,Anqing 246000,Anhui,China)
出处
《医学研究与战创伤救治》
CAS
北大核心
2023年第12期1299-1304,共6页
Journal of Medical Research & Combat Trauma Care