机器学习模型预测心脏外科手术患者术后谵妄的有效性

Effectiveness of machine learning models in predicting postoperative delirium in patients undergoing cardiac surgery

目的基于机器学习算法建立心脏外科手术患者术后谵妄(POD)风险预测模型,并验证其有效性。方法选择2021年5-12月择期行心脏外科手术患者710例,男468例,女242例,年龄≥18岁,ASAⅠ-Ⅳ级。采用医疗电子信息系统收集患者资料。记录术前抑郁筛查量表(PHQ-9)评分、广泛性焦虑障碍量表(GAD-7)评分以及简易精神状态检查量表(MMSE)。将采集的整个数据集按照训练集(75%)和测试集(25%)的比例进行划分,其中训练集和测试集中POD发生率相同。建立6种机器学习模型,包括梯度提升决策树(GBDT)、支持向量机(SVM)、随机森林(RF)、逻辑回归(LogR)、K最邻近法(KNN)、深度神经网络(DNN),基于这6种算法以5折交叉验证的方式对训练数据集的数据进行模型的学习训练,通过测试数据集的数据对模型性能进行验证。基于准确率、精确率、召回率、F1分数、受试者工作特征曲线(ROC)及ROC曲线下面积(AUC)比较不同模型的有效性,并找出适合本研究数据框架的最佳模型。结果有151例(21.3%)心脏外科手术患者中发生POD。本研究进行了6个机器学习模型的性能比较,在使用全部特征作为潜在风险因素的条件下,GBDT的AUC为0.86(95%CI 0.82~0.89),SVM的AUC为0.79(95%CI 0.76~0.83),RF的AUC为0.85(95%CI 0.83~0.87),LogR的AUC为0.67(95%CI 0.63~0.70),KNN的AUC为0.67(95%CI 0.63~0.69),DNN的AUC为0.78(95%CI 0.74~0.82)。结论机器学习算法开发的预测模型可用于心脏外科手术后POD的预测,其中GBDT和RF表现出了较好的机器学习效能,适合于本研究数据框架,更有可能提高POD预测的准确性。特征工程可进行患者数据的可视化处理,以筛选心脏外科手术发生POD的风险因素。

Objective Risk prediction model for postoperative delirium(POD)in cardiac surgery patients was established based on machine learning(ML)algorithm,and its effectiveness was verified.Methods A total of 710 patients with elective heart surgery from May 2021 to December 2021 were selected,468 males and 242 females,aged≥18 years,ASA physical statusⅠ-Ⅳ.Data were collected through the medical electronic information system.Preoperative follow-up questionnaires included the PHQ-9 depression status score,GAD-7 anxiety status score and the assessment of the brief mental state examination scale(MMSE).The whole data set were divided according to the proportion of training set(75%)and test set(25%),in which the proportion of POD in training set and test set was equal.Six kinds of machine learning models were established,including gradient lifting decision tree(GBDT),support vector machine(SVM),random forest(RF),logistic regression(LogR),K-nearest neighbor method(KNN),and deep neural networks(DNN).Based on these six algorithms,the model learning and training of the data in the training data set were carried out in the way of 5-fold cross-validation.The model performance was verified by the data of the test data set.The effectiveness of different models was compared.Based on the effectiveness of accuracy,precision,recall,F1-score,receiver operating characteristic curve(ROC)and area under curve(AUC),the best model suitable for the data framework of this study was found.Results POD occurred in 151(21.3%)cardiac surgical patients after cardiac surgery.The performance of six machine learning models was compared.Under the condition that all features were used as potential risk factors,the AUC of GBDT was 0.86(95%CI 0.82-0.89),the AUC of SVM was 0.79(95%CI 0.76-0.83),the AUC of RF was 0.85(95%CI 0.83-0.87),the AUC of LogR was 0.67(95%CI 0.63-0.70),the AUC of KNN was 0.67(95%CI 0.63-0.69),and the AUC of DNN was 0.78(95%CI 0.74-0.82).Conclusion The prediction model developed by machine learning algorithm can be used to predict POD after cardiac surgery,wherein GBDT and RF show better machine learning efficiency,which is suitable for the data framework of this study and more likely to improve the accuracy of POD prediction.The feature engineering allows visualization of patient data to screen for risk factors for POD in cardiac surgery.