利用简单床旁指标和可解释机器学习算法建立老年危重症患者死亡预测模型
该研究回顾性纳入2017年6月至2020年5月塔城地区人民医院ICU收治的≥65岁老年危重症患者392例(存活组341例、死亡组51例),构建基线数据集(B)及B+APACHE II(BA)、B+GCS(BG)、B+SOFA(BS)四个数据集,用Logistic回归(LR)、极端梯度提升(XGboost)和梯度提升决策树(GBDT)三种算法训练12个预后模型。基于B数据集的模型整体表现落后,LR模型在BA数据集上表现最佳,AUC 0.767(95%CI 0.692-0.836)、准确度0.875(95%CI 0.837-0.903)、F1评分0.190;该模型最重要的3个变量为入ICU后首个24小时晶体液量、心率和平均动脉压,列线图显示总分150~230分区间适用。作者认为,包含简单床旁参数联合APACHE II评分的可解释机器学习模型可有效识别老年危重症患者的死亡风险。
要点
- 392例≥65岁老年危重症患者中死亡51例、存活341例,死亡前三位病因为休克、脑出血和慢性阻塞性肺疾病。
- 12个预后模型中,基于基线数据集(年龄、体温、心率、脉搏血氧饱和度、呼吸频率、平均动脉压、尿量、输液量、晶体液量)的模型整体表现落后。
- LR模型在B+APACHE II数据集上表现最佳:AUC 0.767(95%CI 0.692-0.836)、准确度0.875(95%CI 0.837-0.903)、F1评分0.190。
- 该最佳模型中重要性前3位的变量依次为入ICU后首个24小时晶体液量、心率、平均动脉压。
- 模型列线图显示总分150~230分区间为适用区间。
适用人群与场景
- 2017年6月至2020年5月在塔城地区人民医院ICU住院、年龄≥65岁的老年危重症患者,回顾性纳入。
- 按住院结局分为存活组(341例)和死亡组(51例)。
诊断标准与关键阈值
- 模型列线图总分150~230分为适用区间。
研究设计与关键结果
- 回顾性纳入392例老年危重症患者,存活组341例、死亡组51例;两组在心率、脉搏血氧饱和度、平均动脉压、输液量、晶体液量和病因分布上差异有统计学意义。
- 死亡前三位病因为休克、脑出血和慢性阻塞性肺疾病。
- 共训练12个预后模型(LR、XGboost、GBDT三种算法×B、BA、BG、BS四个数据集);基于B数据集的模型整体表现落后。
- LR模型在BA数据集上表现最佳:AUC 0.767(95%CI 0.692-0.836)、准确度0.875(95%CI 0.837-0.903)、F1评分0.190。
- 该模型重要性前3位变量为入ICU后首个24小时晶体液量、心率、平均动脉压;列线图总分150~230分为适用区间。
证据类型、等级与局限
- 证据类型:回顾性研究,基于可解释机器学习算法(LR、XGboost、GBDT)构建死亡预测模型,并以SHAP法绘制特征重要性直方图。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为,包含简单床旁参数联合APACHE II评分的可解释机器学习模型可有效识别老年危重症患者的死亡风险。
- 该模型基于单中心回顾性数据,原文未据此给出推荐。
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 explore the feasibility of incorporating simple bedside indicators into death predictive model for elderly critically ill patients based on interpretability machine learning algorithms, providing a new scheme for clinical disease assessment.
METHODS: Elderly critically ill patients aged ≥ 65 years who were hospitalized in the intensive care unit (ICU) of Tacheng People's Hospital of Ili Kazak Autonomous Prefecture from June 2017 to May 2020 were retrospectively selected. Basic parameters including demographic characteristics, basic vital signs and fluid intake and output within 24 hours after admission, as well acute physiology and chronic health evaluation II (APACHE II), Glasgow coma score (GCS) and sequential organ failure assessment (SOFA) were also collected. According to outcomes in hospital, patients were divided into survival group and death group. Four datasets were constructed respectively, namely baseline dataset (B), including age, body temperature, heart rate, pulse oxygen saturation, respiratory rate, mean arterial pressure, urine output volume, infusion volume, and crystal solution volume; B+APACHE II dataset (BA), B+GCS dataset (BG), and B+SOFA dataset (BS). Then three machine learning algorithms, Logistic regression (LR), extreme gradient boosting (XGboost) and gradient boosting decision tree (GBDT) were used to develop the corresponding mortality predictive models within four datasets. The feature importance histogram of each prediction model was drawn by SHapley additive explanation (SHAP) method. The area under curve (AUC), accuracy and F1 score of each model were compared to determine the optimal prediction model and then illuminate the nomogram.
RESULTS: A total of 392 patients were collected, including 341 in the survival group and 51 in the death group. There were statistically significant differences in heart rate, pulse oxygen saturation, mean arterial pressure, infusion volume, crystal solution volume, and etiological distribution between the two groups. The top three causes of death were shock, cerebral hemorrhage, and chronic obstructive pulmonary disease. Among the 12 prognostic models trained by three machine learning algorithms, overall performance of prognostic models based on B dataset was behind, whereas the LR model trained by BA dataset achieved the best performance than others with AUC of 0.767 [95% confidence interval (95%CI) was 0.692-0.836], accuracy of 0.875 (95%CI was 0.837-0.903) and F1 score of 0.190. The top 3 variables in this model were crystal solution volume with first 24 hours, heart rate and mean arterial pressure. The nomogram of the model showed that the total score between 150 and 230 were advisable.
CONCLUSION: The interpretable machine learning model including simple bedside parameters combined with APACHE II score could effectively identify the risk of death in elderly patients with critically illness.
参考来源
- Meng Y, Li J, Shan X, et al. [Establishing of mortality predictive model for elderly critically ill patients using simple bedside indicators and interpretable machine learning algorithms]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2025. PMID: 40017366. doi:10.3760/cma.j.cn121430-20240729-00640
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。