Lasso-logistic模型在医院下呼吸道感染预测中的应用

目的建立住院患者医院下呼吸道感染预测模型,构建新的、简单的风险评分方法。方法以2014年多家医院感染调查数据为训练集,建立住院患者医院下呼吸道感染的Lasso-logistic回归预测模型,选择贝叶斯信息准则(BIC)最小模型为最终模型,将回归系数放大相同倍数建立评分方法,以2015、2016年调查数据为验证集,并与文献建立的风险评分方法进行比较。结果 Lasso过程共进行360步,第24步时BIC最小(6 690.4),正则化参数λ=130.8。风险评分方法包含17个条目,数量是文献风险评分方法的1/4,DeLong’s检验显示,两评分方法验证集受试者工作特征曲线下面积(AUC)差异无统计学意义(Z=0.371,P=0.710),决策曲线几乎重合,净重新分类指数为-0.0149,差异无统计学意义(Z=-1.301,P=0.193),整体鉴别指数为0.006,改善差异有统计学意义(P=0.014)。结论利用Lasso-logistic回归模型建立了住院患者医院下呼吸道感染风险简单评分方法,该方法的条目相对简洁,预测效果准确。

Application of Lasso-logistic model in prediction of healthcare-associated lower respiratory tract infection

Objective To develop a predictive model for healthcare-associated lower respiratory tract infection(HA-LRTI) in hospitalized patients, and establish a simple risk scoring method. Methods Survey data of healthcare-associated infection(HAI)in a few hospitals in 2014 was as training dataset, a Lasso-logistic regression model for predicting HA-LRTI in hospitalized patients was established, minimum model of Bayesian information criterion (BIC) was chosen as the final model, scoring method was established by magnifying regression coefficient by the same scale, survey data of 2015 and 2016 were used as the validation dataset, and was compared with risk scoring method established in the literatures. Results Among the 360 steps of Lasso, smallest BIC (6 690.4) occurred at step 24 with regularization parameter λ=130.8. The risk scoring method consisted 17 items, which was 1/4 of the amount of literature risk scoring method, DeLong's test showed that there was no significant difference in area under the curve of receiver operating characteristic between two scoring methods (Z=0.371,P=0.710), decision curve analysis almost overlaid, the net reclassification index was -0.0149, with no significant difference (Z=-1.301,P=0.193), the integrated discrimination index was 0.006, and difference was significant (P=0.014). Conclusion Lasso-logistic regression model established a simple scoring method of HA-LRTI risk for inpatients, the items of the method is relatively concise and the predictive effect is accurate.