通讯作者: 徐华, hzzxyy_006@163.com

DOI：10.12201/bmr.202602.00094

A predictive model and performance evaluation for acute respiratory distress syndrome in the elderly patients with severe trauma based on interpretable machine learning

• 摘要：目的 构建一种基于机器学习算法的预测模型，用于评估老年严重创伤并发急性呼吸窘迫综合征（acute respiratory distress syndrome，ARDS）风险。方法 选取2020年1月至2025年10月湖州市中心医院收治的326例老年严重创伤患者作为研究对象，按照7:3随机分为建模组（70%）和验证组（30%）。基于建模组数据，通过单因素分析、Lasso回归、多因素Logistic回归分析患者并发ARDS的独立危险因素，分别使用随机森林（random forest，RF）、梯度提升（gradient boosting machine，GBM）、人工神经网络（artificial neural network，ANN）、逻辑回归（logistic regression，LR）4种机器学习算法构建模型，运用验证组数据评估模型效能。此外，运用沙普利可加性解释（shapley additive explanation，SHAP）识别危险因素贡献程度。结果 基于建模组数据，单因素分析、Lasso回归共筛选出5个与ARDS发生相关的特征变量，进一步纳入多因素Logistic回归分析，结果显示，格拉斯哥昏迷评分（glasgow coma score，GCS）≤8分、胸部简明损伤定级标准(abbreviated injury scale，AIS)≥3分、白细胞计数（white blood cell，WBC）、乳酸（lactic acid，Lac）、血糖（glucose，Glu）是老年严重创伤并发ARDS的独立危险因素。基于上述变量，构建4种机器学习模型，其中RF模型在建模组、验证组展示最优性能，AUC分别为0.877、0.838；F1分数分别为0.750、0.557。校准曲线显示，RF模型在建模组、验证组中预测概率与实际概率一致性较好。决策曲线显示，RF模型在两组间均具有较广的临床净收益。此外，SHAP分析显示，Glu、Lac、WBC依次为RF模型预测ARDS的前三大重要特征变量。结论 基于RF模型构建的老年严重创伤并发ARDS预测模型具有较好的预测能力，运用SHAP方法提升了RF模型可解释性，有助于指导临床实践。

  关键词： 老年；严重创伤；机器学习；急性呼吸窘迫综合征；预测模型; ; ; ; 

   

  Abstract: Objective To construct a predictive model based on machine learning algorithms for evaluating the risk of acute respiratory distress syndrome (ARDS) in elderly patients with severe trauma. Methods A total of 326 elderly patients with severe trauma admitted to Huzhou Central Hospital from January 2020 to October 2025 were selected as the research subjects. According to a ratio of 7 to 3, they were randomly divided into a modeling group (70%) and a validation group (30%). Based on modeling group data, univariate analysis, lasso regression and multivariate logistic regression were used to analyze the risk factors for ARDS in patients. Four machine learning algorithms, namely random forest (RF), gradient boosting machine (GBM), artificial neural network (ANN) and logistic regression (LR) were used to construct the models. The validation group data was used to evaluate the performance of models. In addition, Shapley additive explanation (SHAP) was used to identify the contribution of risk factors. ResultsBased on modeling group data, a total of 5 feature variables related to ARDS were selected through univariate analysis and lasso regression.Further incorporating multiple logistic regression analysis, the results showed that Glasgow Coma Score (GCS)≤8 points, chestAbbreviated Injury Scale (AIS)≥3 points, White Blood Cell (WBC), Lactic Acid (Lac) and Glucose (Glu) were independent risk factors for ARDS in elderly patients with severe trauma.Based on the above five variables, four machine learning models were constructed, among which the RF model showed the best performance in the modeling group and validation group, AUC were 0.877 and 0.838, F1 scores were 0.750 and 0.557. The calibration curve showed that the RF model had good consistency between the predicted probability and the actual probability in the modeling and validation groups. The decision curve showed that the RF model had a wide clinical net benefit between both groups. In addition, SHAP analysis showed that Glu, Lac and WBC were the top three important feature variables for RF model prediction of ARDS. Conclusion The predictive model for ARDS in elderly patients with severe trauma based on RF model had good predictive ability, and SHAP method improved the interpretability of RF model, which helps to guide clinical practice.

  Key words: Elderly; Severe trauma; Machine learning; Acute respiratory distress syndrome; Predictive model

  提交时间：2026-02-24

  版权声明：作者本人独立拥有该论文的版权，预印本系统仅拥有论文的永久保存权利。任何人未经允许不得重复使用。
• html

• 图表

• 吴苏龙, 曾文美, 郑玥, 黎毅敏. 人中性粒细胞多肽对急性呼吸窘迫综合征患者病情及预后评估作用. 2024. doi: 10.12201/bmr.202410.00060

  何源, 杨树杰, 曾英, 易海英, 黄杰, 甘斌, 魏贤, 唐艳林. 微创给药在极早产儿呼吸窘迫综合征的疗效观察. 2024. doi: 10.12201/bmr.202410.00018

  刘娇娇, 郭连梅, 孙晓红, 张喆, 田野. 基于机器学习的出血性卒中患者下肢深静脉血栓预测模型构建. 2025. doi: 10.12201/bmr.202510.00023

  段馨悦, 王忠. 血清sTREM2与老年急性冠脉综合征的相关性研究. 2024. doi: 10.12201/bmr.202412.00086

  苏燕凤, 洪素茹, 陈钰霜, 吴夏阳. 结合机器学习模型的早产儿ICU死亡风险评估与可解释性分析. 2025. doi: 10.12201/bmr.202503.00066

  肖飞龙, 程江华, 丁颖威, 毛云. CLAUS方案与床旁超声快速管理方案在急诊室重症患者急性呼吸困难病因诊断中的对比. 2024. doi: 10.12201/bmr.202412.00087

  杨林. 冠状动脉周围脂肪衰减指数联合实验室指标预测急性冠脉综合征的研究. 2024. doi: 10.12201/bmr.202409.00001

  唐诗诗, 李宇轩, 唐圣晟, 刘庆华, 周毅. 基于SARIMA-LSTM模型的肾综合征出血热发病率预测研究. 2024. doi: 10.12201/bmr.202407.00046

  芮晨. 基于机器学习的2型糖尿病合并高血压影响因素模型对比与解释性分析. 2025. doi: 10.12201/bmr.202511.00058

  姚泓哲, 郭鑫龙, 杨林瀛. FAP/ALB、NLPR对老年CAP严重程度的预测价值. 2024. doi: 10.12201/bmr.202412.00057

• 序号	提交日期	编号	操作
  1	2026-01-12	10.12201/bmr.202602.00094V1	下载

• 公开评论  匿名评论  仅发给作者 提交评论