纳入脑小血管病总体负担的急性缺血性脑卒中患者1年内复发的预测模型构建

Construction of a Prediction Model for One-Year Recurrence in Acute Ischemic Stroke Patients Included in the Overall Burden of Cerebral Small Vessel Disease

摘要: 目的：探讨脑小血管病(Cerebral Small Vessel Disease, CSVD)总体负担与复发性缺血性卒中的关系，并比较不同机器学习算法构建急性缺血性卒中患者1年内复发风险的预测模型。方法：收集2020年12月~2022年1月期间住院的临床资料完整的527例急性缺血性卒中患者，对其进行CSVD总体负担评分，在此基础上采用最小绝对值收敛和选择算子算法(Least Absolute Shrinkage and Selection Operator, LASSO)回归进行特征筛选。而后基于Logistic回归模型及机器学习算法[构建以自适应提升(Adaptive Boosting, Ada Boost)、极端梯度提升(Extreme Gradient Boosting, XG Boost)、梯度提升回归(Gradient Boosting, GBRT)模型、随机森林模型]构建纳入脑小血管病总体负担的急性缺血性卒中患者1年内复发的预测模型。使用AUC (Area under Curve)值以及校准度(Brier分数)指标来评估模型效果。结果：随访1年时，复发组89例，未复发组438例，两组在年龄、高血压、糖尿病、空腹血糖、高密度脂蛋白、服用抗血小板药物情况、服用调脂药物、CSVD总体负担、WMH、CMB、EPVS、腔隙方面均具有统计学差异，LASSO回归选择出年龄、卒中病史、高血压、糖尿病、高脂血症、服用抗血小板药物情况、脑动脉狭窄评级、CSVD总体负担评分、WMH评级、CMB评级作为缺血性卒中复发的预测因子。各模型区分度均高于0.75，其中在Ada Boost模型区分度(AUC: 0.82, 95% CI 0.796~0.825)较其他模型较高，各模型的Brier分数均小于0.20。结论：CSVD总体负担与复发性缺血性卒中存在相关性，在模型中具有良好的预测价值，机器学习算法中Ada Boost模型有较好的预测效能。

Abstract: Objective: Exploring the relationship between overall burden of cerebral small vessel disease (CSVD) and recurrent ischemic stroke, and developing a prediction model for recurrence within 1 year in acute ischemic stroke patients based on machine learning algorithms. Methods: By collecting complete clinical data from 527 cases of acute ischemic stroke patients hospitalized between December 2020 and January 2022, the overall burden of cerebral small vessel disease (CSVD) will be assessed. Based on this, feature selection will be performed using the least absolute shrinkage and selection operator (LASSO) regression method. Subsequently, a predictive model for the risk of recurrence within 1 year in acute ischemic stroke patients will be developed using Logistic regression and machine learning algorithms, including adaptive boosting (Ada Boost), extreme gradient boosting (XG Boost), gradient boosting regression trees (GBRT), and random forest. The performance of the models will be evaluated using the area under curve (AUC) value and Brier score. Results: At the 1-year follow-up, there were 89 cases of recurrence and 438 cases without recurrence. The two groups showed statistical differences in age, hypertension, diabetes, fasting blood glucose, high-density lipoprotein, use of antiplatelet drugs, reuse of lipid-lowering drugs, overall burden of CSVD, white matter hyperintensities (WMH), cerebral microbleeds (CMB), enlarged perivascular spaces (EPVS), and lacunar infarcts. LASSO regression selected age, history of stroke, hypertension, diabetes, dyslipidemia, use of antiplatelet drugs, intracranial artery stenosis grade, CSVD burden score, WMH grade, and CMB grade as predictive factors for recurrence of ischemic stroke. The discriminative ability of each model was higher than 0.75, with the Ada Boost model showing the highest discrimination (AUC: 0.82, 95% CI 0.796~0.825) compared to other models. The Brier scores of all models were less than 0.20. Conclusion: There is a correlation between the overall burden of cerebral small vessel disease (CSVD) and recurrent ischemic stroke, indicating good predictive value in the model. Among machine learning algorithms, the Ada Boost model demonstrates better predictive performance.