基于机器学习和电阻抗断层成像的撤机结局预测方法研究

Prediction method for weaning outcomes based on machine learning and electrical impedance tomography

目的 :提出一种基于机器学习和电阻抗断层成像(electrical impedance tomography,EIT)预测撤机结局的方法。方法:首先,从来自30例患者的84个样本提取EIT图像特征,并将采用极致梯度提升(extreme gradient boosting,XGBoost)算法筛选出的重要特征作为模型的输入。其次,采用随机森林(random forest,RF)、支持向量机(support vector machines,SVM)、XGBoost、梯度提升决策树(gradient boosting decision tree,GBDT)、逻辑回归(logistic regression,LR)和决策树(tree)6种机器学习方法建立预测模型。在未平衡数据集、过采样平衡数据集和随机欠采样平衡数据集下通过AUC、准确率、敏感度和特异度指标评估模型的预测性能。结果:在AUC、准确率和特异度3个方面,过采样平衡数据集和随机欠采样平衡数据集下的模型性能均优于未平衡数据集(P<0.05);但在敏感度方面,过采样平衡数据集与未平衡数据集下的模型性能比较差异无统计学意义(P>0.05),随机欠采样平衡数据集下的模型性能比未平衡数据集下的模型性能有所下降(P<0.05)。在过采样平衡数据集和随机欠采样平衡数据集下模型性能比较差异无统计学意义(P>0.05)。基于XGBoost建立的模型在过采样平衡后的数据集中综合性能最佳,AUC为0.769、准确率为0.808、敏感度为0.938、特异性度为0.600。结论:基于机器学习和EIT能够较好地预测长时间机械通气患者的撤机结局,可以为临床医生判断合适的撤机时机提供辅助决策支持。

Objective To propose a method for predicting weaning outcomes based on machine learning and electrical impedance tomography(EIT).Methods Firstly,EIT image features were extracted from a total of 84 samples from 30 patients,and the important features screened with the extreme gradient boosting(XGBoost)algorithm were used as inputs to the model.Secondly,the prediction model was built with six machine learning methods,namely random forest(RF),support vector machines(SVM),XGBoost,gradient boosting decision tree(GBDT),logistic regression(LR)and decision tree(tree).Then the prediction model had its prediction performance evaluated by AUC,accuracy,sensitivity and specificity under imbalanced dataset,over-sampling balanced dataset and random under-sampling balanced dataset.Results In terms of AUC,accuracy and specificity,the model under the over-sampling balanced dataset and the random under-sampling balanced dataset behaved better than that under the imbalanced dataset(P<0.05);in terms of sensitivity,the difference in model performance between the over-sampling balanced dataset and the imbalanced dataset was not statistically significant(P>0.05),and the model performance under the random under-sampling balanced dataset decreased when compared with that under the imbalanced dataset(P<0.05).There were no significant differences between the model performance under the over-sampling balanced dataset and that under the random under-sampling balanced dataset(P>0.05).The model based on XGBoost behaved the best under the over-sampling balanced dataset,with an AUC of 0.769,an accuracy of 0.808,a sensitivity of 0.938 and a specificity of 0.600.Conclusion The method based on machine learning and EIT predicts weaning outcomes of patients with prolonged mechanical ventilation,and thus can be used for auxiliary decision support for clinicians to determine the appropriate timing of weaning.[Chinese Medical Equipment Journal,2023,44(10):1-6]