基于人工神经网络及logistic回归的耐碳青霉烯类革兰阴性菌感染预测模型研究

Prediction model of carbapenem-resistant Gram-negative bacteria infection based on artificial neural network and logistic regression

目的 构建基于人工神经网络与logistic回归模型的耐碳青霉烯类革兰阴性菌(CR-GNB)感染预测模型。方法 回顾性收集某院2020年1月—2021年3月革兰阴性菌感染患者1 824例为研究对象,根据患者临床资料分别用人工神经网络和logistic回归算法构建CR-GNB感染预测模型,从区分度和校准度两个方面比较模型的预测效能。结果 共发生医院感染2 392例,CR-GNB感染504例,占所有医院感染的21.07%。logistic回归模型ROC曲线下面积(AUC)为0.837(0.794~0.880),在截断值为0.255时,灵敏度为73.3%,特异度为79.3%,模型预测准确率为77.5%,H-L拟合优度检验:χ^(2)=12.225,P=0.141;人工神经网络模型的AUC为0.850(0.828~0.872),在截断值为0.251时,灵敏度为80.1%,特异度为78.7%,模型预测准确率为78.8%,H-L拟合优度检验:χ^(2)=11.363,P=0.182。结论 两种模型在区分度和校准度方面均有较好的表现,人工神经网络稍优于logistic回归模型。

Objective To construct a prediction model of carbapenem-resistant gram-negative bacteria(CR-GNB)infection based on artificial neural network and logistic regression model.Methods 1824 patients with gram-negative bacterial infection from January 2020 to March 2021 in our hospital were retrospectively collected as the study population,and the prediction model of CR-GNB infection was constructed using artificial neural network and logistic regression algorithm respectively based on the patients'clinical data,and the prediction efficiency of the models was compared in terms of discrimination and calibration.Results A total of 2392 cases of nosocomial infections occurred,and 504 cases of CR-GNB infection accounted for 21.07%of all nosocomial infections.The area under the ROC curve(AUC)of the logistic regression model was 0.837(0.794-0.880),with a sensitivity of 73.3%,a specificity of 79.3%,a model prediction accuracy of 77.5%,and H-L goodness-of-fit test:(χ^(2)=12.225,P=0.141 at a cut-off value of 0.255;the AUC of the artificial neural network model was 0.850(0.828~0.872),with a sensitivity of 80.1%,a specificity of 78.7%,a model prediction accuracy of 78.8%,and H-L goodness-of-fit test:χ^(2)=11.363,P=0.182 at a cut-off value of 0.251.Conclusion Both models have good performance in terms of discrimination and calibration,and the artificial neural network is slightly better than the logistic regression model.