目的

基于可解释机器学习模型分析机械通气患者呼吸机相关性肺炎（ventilatorassociated pneumonia，VAP）风险。

方法

选取2022 年1 月至2024 年12 月我院收治的机械通气患者77 例。 发生VAP 19 例分为观察组和未发生VAP 58 例分为对照组，收集临床资料，采用LASSO 回归分析筛选与VAP 相关特征，运用逻辑回归（logistic regression，LR）、支持向量机（support vector machine，SVM）、多层感知器（multi-layer perceptron，MLP）及极端梯度提升（eXtreme Gradient Boosting，xGBoost）4 种机器学习（machine learning，ML）算法构建预测，采用SHAP 方法分析可解释性。

结果

单因素分析显示，观察组急性生理学及慢性健康状况评分系统（acute physiology and chronic health evaluation Ⅱ，APACHE Ⅱ）（21.53±2.50）分、机械通气时间（10.37±3.30）d 高于对照组APACHE Ⅱ（18.59±2.80）分、机械通气时间（8.38±2.55）d，观察组格拉斯哥评分（10.79±2.07）分、白蛋白水平（30.08±3.79）g/L 低于对照组格拉斯哥评分（12.14±1.94）分、白蛋白水平（34.22±5.20）g/L，观察组慢性阻塞性肺疾病、糖尿病、使用镇静剂、抗生素种类与对照组差异有统计学意义（P＜0.05）。 LASSO 回归分析显示，APACHE Ⅱ评分、机械通气时间、白蛋白是机械通气患者发生VAP 的风险因素。 基于风险因素构建4 个ML，xGBoost 曲线下面积（area under the curve，AUC）为0.882，准确性为81.30%，敏感度为84.00%，特异度为82.00%。SHAP 结果显示，VAP 风险贡献度前4 的特征分别为APACHE Ⅱ评分、机械通气时间、白蛋白水平。

结论

xGBoost 敏感度和特异度高，作为一种辅助诊断工具快速、准确判断机械通气患者发生VAP 风险，为临床决策提供参考。

Objective

To analyze the risk of ventilator-associated pneumonia （VAP） in patients with mechanical ventilation based on an interpretable machine learning model.

Methods

All of 77 patients with mechanical ventilation in our hospital from January 2022 to December 2024 were selected. 19 cases with VAP were divided into the observation group and 58 cases without VAP were divided into the control group. Clinical data were collected，and the features related to VAP were screened by LASSO regression analysis. Logistic regression （LR） and support vector machine were used. SVM，multi-layer perceptron （MLP） and eXtreme Gradient Boosting （xGBoost） were four machine learning（ML） algorithms to construct the prediction model，and SHAP was used to analyze the interpretability of the model.

Results

Univariate analysis showed that acute physiology and chronic health evaluation Ⅱ，APACHEⅡ） （21.53±2.50） minutes and mechanical ventilation time （10.37±3.30） d were higher than those of control group （18.59±2.80） minutes and mechanical ventilation time （8.38±2.55） d. Glasgow score （10.79±2.07） and albumin level （30.08±3.79） g/L in the observation group were lower than those in the control group （12.14±1.94） and albumin level （34.22±5.20） g/L. There were significant differences in age，chronic obstructive pulmonary disease，diabetes，sedative use and antibiotic type between observation group and control group （P ＜0.05）. LASSO regression results showed that age，APACHE Ⅱscore，mechanical ventilation time and albumin were risk factors for VAP in patients with mechanical ventilation. Four ML models were constructed based on the above four risk factors，and the area under the curve（AUC） of xGBoost model was 0.882，the accuracy was 81.30%，the sensitivity was 84.00%，and the specificity was 82.00%. SHAP results showed that the top 4 characteristics of VAP risk contribution were APACHE Ⅱscore，age，mechanical ventilation time and albumin.

Conclusion

xGBoost model has high sensitivity and specificity，which can be used as an auxiliary diagnostic tool to quickly and accurately determine the risk of VAP in patients with mechanical ventilation，and provide reference for clinical decision-making.