基于机器学习算法构建鼻胃管患者发生误吸的风险预测模型

Construction of a risk predictive model for aspiration in patients with a nasogastric tube based on machine learning algorithms

目的比较7种机器学习算法在预测鼻胃管(NGT)患者发生误吸风险方面的效能。方法回顾性选取2021年1月至2022年12月合肥市第二人民医院高压氧医学科收治的352例NGT患者为研究对象,将患者按照7∶3的比例随机分为训练集246例和验证集106例。采用Boruta算法进行独立特征变量选择,并运用逻辑回归、随机森林、决策树、k-近邻树、轻量级梯度提升机、支持向量机、极端梯度提升7种机器算法分别建立预测模型,通过ROC曲线评估模型的区分能力,结合AUC值、准确度、灵敏度、特异度、阳性预测值和阴性预测值等指标选择最优模型,在选择最优模型后,结合随机森林的变量重要性图和shapley加法解释值解释关键特征对误吸风险的贡献。结果352例患者中,发生误吸102例,误吸发生率为28.98%。训练集和验证集构建了7种NGT患者误吸风险预测模型,均以随机森林模型最优。随机森林模型在训练集中的AUC、准确度、灵敏度、特异度、阳性预测值、阴性预测值均为1.000,在验证集中的AUC、准确度、灵敏度、特异度、阳性预测值、阴性预测值分别为0.977、0.934、0.882、0.962、0.845、0.985。结论基于机器学习算法成功建立了预测NGT患者误吸风险模型,其中随机森林模型呈现出良好的预测能力。

Objective To compare the performance of seven machine learning algorithms in predicting the risk of aspiration in patients with a nasogastric tube(NGT).Methods A retrospective analysis was conducted on 352 NGT patients admitted to Department of Hyperbaric Oxygen Medicine in the Second People's Hospital of Hefei between January 2021 and December 2022.Patients were randomly divided into a training set(246 cases)and a validation set(106 cases)at a 7∶3 ratio.Independent feature selection was performed using the Boruta algorithm,and seven machine learning algorithms(logistic regression,random forest,decision tree,k-nearest neighbors,light gradient boosting machine,support vector machine,and extreme gradient boosting)were used to develop predictive models.The discriminatory ability of the models was evaluated by the ROC curve,and the best model was selected based on the AUC value,accuracy,sensitivity,specificity,positive predictive value,and negative predictive value.After selecting the optimal model,the variable importance plot from the random forest and shapley additive explanations values were used to explain the contribution of key features to the risk of aspiration.Results Among the 352 patients,102 cases of aspiration occurred,with an incidence rate of 28.98%.Seven predictive models for NGT aspiration risk were constructed in the training and validation sets,with the random forest model demonstrating the best performance in both sets.In the training set,the AUC,accuracy,sensitivity,specificity,positive predictive value,and negative predictive value of the randome forest model were all 1.000.In the validation set,the values were 0.977,0.934,0.882,0.962,0.845,and 0.985,respectively.Conclusion Predictive models for the risk of aspiration in NGT patients were successfully established using machine learning algorithms,with the random forest model showing excellent predictive performance.