基于Logistic回归与XGBoost构建缺血性卒中院内复发风险预测模型的初步比较研究

目的基于Logistic回归和XGBoost方法构建缺血性卒中院内复发风险预测模型,并进行初步比较。方法利用中国国家卒中登记Ⅱ(China National Stoke RegistryⅡ,CNSRⅡ)数据库中按医嘱离院的缺血性卒中患者数据,分别基于Logistic回归和XGBoost方法构建缺血性卒中院内复发风险预测模型。备选的预测因子包括人口学特征、卒中严重程度、既往病史、用药史以及临床测量指标。模型的评价指标包括ROC曲线下面积(area under the cure,AUC)、校准截距、校准斜率以及Brier得分。所有统计分析均在R(3.6.2版)中完成。结果最终纳入17227例符合条件的患者,平均年龄64.72±11.84岁,女性6317例(36.7%),发病前mRS评分为0或1分的病例14482例(84.1%),入院NIHSS评分4(2~6)分,院内卒中复发444例(2.6%)。预测模型识别的前三位强预测因子,在Logistic回归中为发病前mRS评分、心房颤动及卒中史;在XGBoost中为发病前mRS评分、心房颤动及总胆固醇。Logistic回归预测模型与XGBoost预测模型的AUC无显著差异(0.63,95%CI 0.58~0.68 vs 0.64,95%CI 0.59~0.68,P=0.9229)。Logistic预测模型校准截距、校准斜率以及Brier得分分别为-0.81、0.76和0.03;XGBoost预测模型的校准截距、校准斜率以及Brier得分分别为-1.37、1.20和0.38。Logistic预测模型校准度更好。结论利用CNSRⅡ数据构建的缺血性卒中院内复发风险预测模型应用中,基于XGBoost方法构建的预测模型相比Logistic回归构建的预测模型的区分度没有显著差异,但校准度略低。

Comparison of Prediction Models for In-hospital Stroke Recurrence in Patients with Ischemic Stroke Based on Logistic Regression and XGBoost Methods

Objective To compare prediction models for in-hospital stroke recurrence in patients with ischemic
stroke based on logistic regression and XGBoost methods.
Methods Data of ischemic stroke inpatients discharged according to medical advice from China
National Stroke Registry Ⅱ (CNSR Ⅱ) database were retrospectively analyzed. Logistic regression
and XGBoost methods were used to develop a model for predicting in-hospital stroke recurrence.
Candidate predictors included demographic characteristics, stroke severity, medical history,
medication history, and clinical measure indicators. The performance measures of the predictive
models included the area under the receiver operating characteristic curve (AUC), calibration intercept,
calibration slope and Brier score. All statistical analysis was performed using R (version 3.6.2).
Results A total of 17 227 eligible patients were included in this analysis. The mean age was
64.72±11.84 years, and 6317 (36.7%) cases were females. A total of 14 482 (84.1%) patients had amRS score of 0 or 1 point before symptoms onset, and the NIHSS score at admission was 4 (2-6).
A total of 444 (2.6%) patients had recurrent stroke during hospitalization. The three leading strong
predictors were mRS score, atrial fibrillation and stroke history in logistic regression model, and mRS
score, atrial fibrillation and total cholesterol in XGBoost model. No significant difference was found
in AUC between logistic regression model and XGBoost model (0.63, 95%CI 0.58-0.68 vs 0.64,
95%CI 0.59-0.68, P =0.9229). The calibration intercept, calibration slope and Brier score in logistic
regression model were -0.81, 0.76 and 0.03, respectively; and were -1.37, 1.20 and 0.38 in XGBoost
model. Logistic regression model had better calibration than XGBoost model.
Conclusions No significant difference was found in discrimination between logistic-based
prediction model and XGBoost-based prediction model for in-hospital stroke recurrence constructed
using data of CNSR II, while logistic-based prediction model had better calibration.