基于机器学习算法的危重症患者压力性损伤风险预测模型的构建与验证
该前瞻性队列研究纳入杭州市第一人民医院北院区ICU 2023年1月至2024年3月的350例危重症患者,其中102例(29.1%)发生压力性损伤。采用Lasso回归筛选变量、多因素Logistic回归识别独立危险因素,并构建XG-Boost机器学习预测模型。模型在训练集AUC为0.896(95%CI 0.863-0.929),验证集AUC为0.835(95%CI 0.761-0.908),校准曲线显示预测概率与实际结局一致性良好,DCA提示在较宽阈值概率范围内有临床效用。SHAP分析显示ICU住院时长、机械通气、中性粒细胞计数、意识状态、血糖和使用暖被为重要性递减的预测特征。
要点
- 350例ICU危重症患者中102例(29.1%)发生压力性损伤,发生率较高。
- 多因素Logistic回归显示机械通气(OR=2.338,95%CI 1.768-3.089,P=0.002)、使用暖被(OR=1.772,95%CI 1.341-2.338,P=0.039)、ICU住院时间延长(OR=1.081,95%CI 1.067-1.097,P<0.001)、中性粒细胞计数升高(OR=1.044,95%CI 1.022-1.067,P=0.036)、血糖升高(OR=1.0
- XG-Boost模型预测压力性损伤的AUC训练集0.896(95%CI 0.863-0.929)、验证集0.835(95%CI 0.761-0.908),校准曲线一致性良好,DCA示较宽阈值概率范围内有临床效用。
- SHAP特征重要性降序为:ICU住院时长、机械通气、中性粒细胞计数、意识状态、血糖、使用暖被。
适用人群与场景
- 杭州市第一人民医院北院区ICU 2023年1月至2024年3月收治的危重症患者,采用方便抽样纳入,共350例。
研究设计与关键结果
- 前瞻性队列研究,纳入350例ICU危重症患者,其中102例(29.1%)发生压力性损伤。
- 压力性损伤组与非压力性损伤组在意识状态、机械通气、镇静剂使用、ICU住院时长、Braden评分、使用暖被、白细胞计数、中性粒细胞计数、血糖和乳酸水平方面差异均有统计学意义(均P<0.05)。
- Lasso回归筛选出6个预测变量:意识状态、机械通气、使用暖被、ICU住院时长、中性粒细胞计数、血糖。
- 多因素Logistic回归显示机械通气(OR=2.338,95%CI 1.768-3.089,P=0.002)、使用暖被(OR=1.772,95%CI 1.341-2.338,P=0.039)、ICU住院时间延长(OR=1.081,95%CI 1.067-1.097,P<0.001)、中性粒细胞计数升高(OR=1.044,95%CI 1.022-1.067,P=0.036)、血糖升高(OR=1.062,95%CI 1.031-1.094,P=0.027)为独立危险因素。
- 基于6个风险因素构建XG-Boost预测模型,ROC曲线分析示训练集AUC为0.896(95%CI 0.863-0.929),验证集AUC为0.835(95%CI 0.761-0.908)。
- 校准曲线显示预测概率与实际结局一致性良好,DCA提示模型在较宽阈值概率范围内有临床效用。
- SHAP分析特征重要性降序为:ICU住院时长、机械通气、中性粒细胞计数、意识状态、血糖、使用暖被。
证据类型、等级与局限
- 证据类型:前瞻性队列研究。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为,构建的XG-Boost机器学习模型在预测危重症患者压力性损伤发生方面具有高性能,识别关键预测因素有助于临床风险评估和干预。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「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: Predict the occurrence of pressure injury in critically ill patients by using machine learning models and conducting internal validation.
METHODS: A prospective cohort study was conducted. Critically ill patients admitted to the intensive care unit (ICU) of the North District of Hangzhou First People's Hospital from January 2023 to March 2024 were enrolled using convenience sampling. Patients were divided into two groups based on the occurrence of pressure injury during their ICU stay, and the differences in pressure injury related indicators were compared between the groups. The dataset was randomly divided into a training set (75%) and a validation set (25%). Feature selection was performed using the Lasso regression. Independent risk factors were then identified via multivariate Logistic regression analysis. An extreme gradient boosting (XG-Boost) machine learning model was developed to predict pressure injury risk. The model's performance was comprehensively evaluated using receiver operator characteristic curve (ROC curve), calibration curve, and clinical decision curve analysis (DCA). The Shapley Additive exPlanations (SHAP) method was used to rank feature importance.
RESULTS: A total of 350 critically ill patients were included, of whom 102 (29.1%) developed pressure injuries. There were statistically significant differences in consciousness status, mechanical ventilation, sedative use, length of ICU stay, Braden score, use of warm blankets, white blood cell count, neutrophil count, blood glucose, and lactate level between the pressure injury and non-pressure injury groups (all P<0.05). Lasso regression analysis identified six predictive variables: consciousness status, mechanical ventilation, use of warm blankets, length of ICU stay, neutrophil count, and blood glucose. Multivariate Logistic regression analysis subsequently revealed that mechanical ventilation, use of warm blankets, prolonged ICU stay, elevated neutrophil count, and elevated blood glucose were independent risk factors for pressure injuries [mechanical ventilation: odds ratio (OR)=2.338, 95% confidence interval (95%CI) was 1.768-3.089, P=0.002; use of warm blankets: OR=1.772, 95%CI was 1.341-2.338, P=0.039; prolonged ICU stay: OR=1.081, 95%CI was 1.067-1.097, P<0.001; elevated neutrophil count: OR=1.044, 95%CI was 1.022-1.067, P=0.036; elevated blood glucose: OR=1.062, 95%CI was 1.031-1.094, P=0.027]. Based on these six risk factors, a predictive model was constructed using the XG-Boost method. The ROC curve analysis demonstrated the model has high predictive performance, with an area under the curve (AUC) of 0.896 (95%CI was 0.863-0.929) in the training set and 0.835 (95%CI was 0.761-0.908) in the validation set. The calibration curve indicated good agreement between predicted probabilities and actual outcomes. DCA further suggested that the model had clinical utility across a wide range of threshold probabilities. SHAP analysis ranked feature importance in descending order as follows: length of ICU stay, mechanical ventilation, neutrophil count, consciousness status, blood glucose, and use of warm blankets.
CONCLUSIONS: The constructed XG-Boost machine learning model has high performance in predicting the occurrence of pressure injury in critically ill patients. Identify key predictive factors can aid clinical risk assessment and intervention.
参考来源
- Lyu W, Zhu M, Lu T, et al. [Construction and validation of risk prediction model of pressure injury in critically ill patients based on machine learning algorithm]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2026. PMID: 41876243. doi:10.3760/cma.j.cn121430-20250319-00278
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。