文章摘要
基于机器学习构建术后咽喉疼痛风险的预测模型
Prediction model for postoperative sore throat established based on machine learning
  
DOI：10.12089/jca.2025.05.005
中文关键词: 气管插管  术后咽喉疼痛  机器学习  随机森林  风险预测模型
英文关键词: Intubation, intratracheal  Postoperative sore throat  Machine learning  Random forest  Risk prediction model
基金项目:
作者单位E-mail
王锐 620000,四川省眉山市,眉山市中医医院麻醉科  
郑雷雷 贵州中医药大学第二附属医院麻醉科  
余璇 遵义医科大学第二附属医院麻醉科  
张益 遵义医科大学第二附属医院麻醉科  
易斌 陆军军医大学第一附属医院麻醉科  
黄桂华 遵义医科大学第三附属医院麻醉科(现在北京积水潭医院贵州医院麻醉科) 435141387@qq.com 
摘要点击次数: 527
全文下载次数: 141
中文摘要:
      
目的：利用机器学习构建术后咽喉疼痛（POST）的风险预测模型。
方法：回顾性收集2022年9月至2023年6月行全身麻醉气管插管成年患者757例，男327例，女430例，年龄≥18岁，ASA Ⅰ—Ⅲ级。根据术后24 h内是否发生POST将患者分为两组：POST组（P组）和无POST组（NP组）。收集患者术前、术中以及术后随访信息，将数据集随机划分成80%的训练集与20%的测试集，采用20折交叉验证以及网格搜索的方式在训练集上进行模型开发，测试集用于内部验证，分别建立逻辑回归（LR）、随机森林（RF）、自适应提升算法（AdaBoost）和极端梯度提升算法（XGBoost）等4种机器学习模型。采用受试者工作特征（ROC）曲线下面积（AUC）、准确性、敏感性、特异性、精确率-召回率曲线（PRC）下面积（AUPRC）、Brier分数评估模型的性能。采用夏普利加法解释（SHAP）对最佳性能的模型进行解释分析。
结果：有221例（29.1%）患者发生POST。LR、RF、AdaBoost和XGBoost 4种机器学习模型的AUC分别为0.89（95%CI 0.87～0.90）、0.90（95%CI 0.87～0.91）、0.86（95%CI 0.81～0.89）和0.88（95%CI 0.86～0.91）。RF模型在测试集中表现最优，ROC曲线准确性84.0%，敏感性77.0%，特异性86.0%，Brier分数0.13，AUPRC 0.81，AUC 0.90。SHAP模型解释分析结果显示，拔管时导管套囊带血迹该变量对RF的贡献度最大，其余依次为留置气管导管时间、术中输液量、性别、年龄、气管插管次数>1次、ASA分级、留置胃管。
结论：本研究基于LR、RF、AdaBoost和XGBoost算法构建4种POST风险预测模型，其中RF性能最佳。
英文摘要:
      
Objective: Utilize machine learning to construct a risk prediction model for postoperative sore throat.
Methods: A total of 757 adult patients who underwent tracheal intubation under general anesthesia from September 2022 to June 2023 were retrospectively collected, 327 males and 430 females, aged ≥ 18 years, ASA physical status Ⅰ-Ⅲ. The patients were divided into two groups according to whether POST happened within 24 hours after operation: POST group (group P) and non-POST group(group NP). The preoperative, intraoperative and postoperative follow-up information of the patients was collected. The data set was randomly divided into a training set of 80% and a test set of 20%. The 20-fold cross-validation and grid search methods were used to develop the model on the training set. The test set was used for internal validation. Four models, namely logistic regression (LR), random forest (RF), adaptive boosting (AdaBoost), and extreme gradient boosting (XGBoost), were established respectively. The performance of the four models was compared using the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, area under the precision-recall curve (AUPRC), and Brier score. In addition, the Shapley additive explanations (SHAP) was used to conduct explanatory analysis on the model with the best performance.
Results: A total of 221 patients (29.1%) had POST. The AUC values of the four machine learning models, namely LR, RF, AdaBoost, and XGBoost, were 0.89 (95% CI 0.87-0.90), 0.90 (95% CI 0.87-0.91), 0.86 (95% CI 0.81-0.89), and 0.88 (95% CI 0.86-0.91), respectively. The RF model performed the best in the test set, with ROC curve accuracy of 84.0%, sensitivity of 77.0%, specificity of 86.0%, Brier score of 0.13, AUPRC of 0.81, and AUC of 0.90. The results of the SHAP showed that the variable of blood in the tracheal tube had the greatest contribution to RF, followed in order by the time of indwelling tracheal tube, amount of intraoperative fluids, gender, age, the number of intubations greater than one, ASA physical status, and indwelling gastrostomy tube.
Conclusion: In this study, we constructed a risk prediction model for POST based on four algorithms, including LR, RF, AdaBoost, and XGBoost, RF had the best performance.
查看全文   查看/发表评论  下载PDF阅读器
关闭