基于机器学习的心源性休克患者院内死亡的预测研究

Research on Prediction Model of Cardiogenic Shock in-Hospital Death Based on Machine Learning

目的:构建心源性休克患者院内死亡预测模型。方法:利用MIMIC-III数据库,收集人口特征、实验室检查、合并症等87个指标并进行特征选择后,使用随机森林、Logistic回归、XGBoost、卷积神经网络算法,构建预测模型。用敏感性、特异性、曲线下面积、准确性来比较这4种模型的性能。结果:在这项研究的804名患者中,有209名患者(26%)出现院内死亡。在预测院内死亡时,四个模型的接收器工作特征曲线(ROC)的曲线下面积(AUC)在0.757至0.848的范围内。在所有模型中,XGBoost的灵敏度最高(87.3%),特异性(81%)和准确性最高(84.6%)。结论:机器学习算法可以准确预测心源性休克患者院内死亡率,尤其是XGBoost模型。

Objective: To construct a prediction model for in-hospital death in patients with cardiogenic shock. Methods: Using the MIMIC-III database, 87 indicators of demographic characteristics, laboratory tests, and comorbidities were collected and then feature selection. Logistic regression, random for-est, XGBoost, and convolutional neural network algorithms were employed to build models. Sensi-tivity, specificity, accuracy, and area under the curve were applied to access the performance. Re-sult: Among 804 patients enrolled in this study, 209 patients (26%) died in hospital. In the predic-tion of the in-hospital death, the areas under the curve (AUCs) of the receiver operating characteris-tic curves (ROCs) of the four models ranged from 0.757 to 0.848. Among all models, XGBoost achieved the highest sensitivity (87.3%), specificity (81%) and accuracy (84.6%). Conclusion: Ma-chine learning algorithms can accurately predict the in-hospital mortality of patients with cardio-genic shock, especially the XGBoost model.