基于机器学习构建重症监护病房脓毒性休克患者早期急性肾损伤风险预测模型
该研究回顾性分析了2015年4月至2019年6月济宁医学院附属医院ICU收治的303例脓毒性休克患者的临床数据,按入住ICU 7天内是否发生AKI分为AKI组(153例)和非AKI组(150例),AKI发生率为50.50%。随机选取70%病例作为训练集构建XGBoost模型,剩余30%作为验证集,通过ROC曲线评估预测能力,并与APACHE II、SOFA、PCT等进行比较。结果显示XGBoost模型预测AKI风险的AUC为0.816,敏感度73.3%,特异度71.7%,准确率72.5%,性能显著优于常用临床评分,校准曲线拟合优度更高(校准曲线评分最低,为0.205)。作者认为该模型能更准确地预测脓毒性休克患者AKI风险,有助于在预测预后的同时制定合适的诊断、治疗和随访策略。
要点
- ICU脓毒性休克患者7天内AKI发生率为50.50%(153/303),AKI组APACHE II评分、SOFA评分、血乳酸、去甲肾上腺素剂量、机械通气比例均高于非AKI组,且心率更快。
- XGBoost模型预测AKI风险的AUC为0.816,敏感度73.3%,特异度71.7%,准确率72.5%,性能显著优于APACHE II评分、SOFA评分和PCT。
- 模型排序前10的特征依次为:入ICU时血肌酐水平、去甲肾上腺素使用、饮酒史、白蛋白、血钠、C反应蛋白、乳酸、体重指数、血小板计数、血尿素氮水平。
- 模型校准曲线拟合优度高于其他评分(校准曲线评分最低,为0.205),提示预测值与实际风险一致性更好。
- 作者认为,与常用临床评分相比,XGBoost模型能更准确地预测脓毒性休克患者AKI风险,有助于制定诊断、治疗和随访策略。
适用人群与场景
- 2015年4月至2019年6月于济宁医学院附属医院ICU住院的脓毒性休克患者,共303例,其中AKI组153例、非AKI组150例。
诊断标准与关键阈值
- 根据入住ICU 7天内是否发生急性肾损伤(AKI)分为AKI组和非AKI组。
研究设计与关键结果
- 研究类型为回顾性分析,样本量303例,单中心(济宁医学院附属医院ICU),人群为脓毒性休克患者;按7天内是否发生AKI分组,70%病例随机作为训练集构建XGBoost模型,30%作为验证集。
- AKI发生率为50.50%(153/303);与非AKI组相比,AKI组APACHE II评分、SOFA评分、血乳酸(Lac)更高,去甲肾上腺素(NE)剂量更高,机械通气比例更高,心率更快。
- XGBoost模型预测AKI风险的AUC为0.816,敏感度73.3%,特异度71.7%,准确率72.5%。
- 与APACHE II评分、SOFA评分和PCT相比,XGBoost模型性能显著提高;校准曲线显示XGBoost模型拟合优度高于其他评分(校准曲线评分最低,为0.205)。
- XGBoost模型前10位特征依次为:入ICU时血肌酐(SCr)水平、NE使用、饮酒史、白蛋白、血钠、C反应蛋白(CRP)、Lac、体重指数(BMI)、血小板计数(PLT)、血尿素氮(BUN)水平。
证据类型、等级与局限
- 证据类型:回顾性分析(单中心)。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为,与常用临床评分相比,XGBoost模型能更准确地预测脓毒性休克患者AKI风险,有助于在预测预后的同时制定合适的诊断、治疗和随访策略。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开
OBJECTIVE: To analyze the risk factors of acute kidney injury (AKI) in patients with septic shock in intensive care unit (ICU), construct a predictive model, and explore the predictive value of the predictive model.
METHODS: The clinical data of patients with septic shock who were hospitalized in the ICU of the Affiliated Hospital of Jining Medical College from April 2015 to June 2019 were retrospectively analyzed. According to whether the patients had AKI within 7 days of admission to the ICU, they were divided into AKI group and non-AKI group. 70% of the cases were randomly selected as the training set for building the model, and the remaining 30% of the cases were used as the validation set. XGBoost model was used to integrate relevant parameters to predict the risk of AKI in patients with septic shock. The predictive ability was assessed through receiver operator characteristic curve (ROC curve), and was correlated with acute physiology and chronic health evaluation II (APACHE II), sequential organ failure assessment (SOFA), procalcitonin (PCT) and other comparative verification models to verify the predictive value.
RESULTS: A total of 303 patients with septic shock were enrolled, including 153 patients with AKI and 150 patients without AKI. The incidence of AKI was 50.50%. Compared with the non-AKI group, the AKI group had higher APACHE II score, SOFA score and blood lactate (Lac), higher dose of norepinephrine (NE), higher proportion of mechanical ventilation, and tachycardiac. In the XGBoost prediction model of AKI risk in septic shock patients, the top 10 features were serum creatinine (SCr) level at ICU admission, NE use, drinking history, albumin, serum sodium, C-reactive protein (CRP), Lac, body mass index (BMI), platelet count (PLT), and blood urea nitrogen (BUN) levels. Area under the ROC curve (AUC) of the XGBoost model for predicting the risk of AKI in patients with septic shock was 0.816, with a sensitivity of 73.3%, a specificity of 71.7%, and an accuracy of 72.5%. Compared with the APACHE II score, SOFA score and PCT, the performance of the model improved significantly. The calibration curve of the model showed that the goodness of fit of the XGBoost model was higher than the other scores (the calibration curve had the lowest score, with a score of 0.205).
CONCLUSIONS: Compared with the commonly used clinical scores, the XGBoost model can more accurately predict the risk of AKI in patients with septic shock, which helps to make appropriate diagnosis, treatment and follow-up strategies while predicting the prognosis of patients.
参考来源
- Zhang S, Tang S, Rong S, et al. [Construction of a predictive model for early acute kidney injury risk in intensive care unit septic shock patients based on machine learning]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2022. PMID: 35574741. doi:10.3760/cma.j.cn121430-20211126-01790
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。