急性脑卒中预后预测模型:机器学习与传统回归模型的比较

The Outcome Prediction Model of Acute Stroke:Comparison between Machine Learning and Traditional Regression Model

目的探讨多模态MRI的机器学习模型在预测急性脑卒中血管内治疗后预后中的价值并与传统回归模型进行比较。方法对216例在南京市第一医院就诊的急诊脑卒中患者的临床及影像资料进行回顾性分析。应用多元logistic回归分析筛选卒中预后的预测因子,并构建卒中预后模型。提取并筛选多模态MRI图像病变区影像组学特征,通过支持向量机分类器建立预测卒中预后模型。比较传统回归模型及机器学习模型的效能。结果logistic回归分析显示入院NIHSS评分、HIR、血管完全再通与卒中预后密切相关(P<0.05)。ROC分析显示logistic回归模型预测急性卒中预后的AUC为0.775。机器学习模型预测训练集患者预后的AUC达0.991;预测测试集患者预后的AUC达0.950。logistic回归模型与机器学习模型效能间存在统计学差异(Z=3.146;P<0.001)。结论应用机器学习算法可较为准确的预测急性脑卒中血管内治疗后预后,明显优于传统回归模型,可为临床治疗方案制定提供指导。

Objective To explore the value of machine learning model of multimodal MRI in predicting outcome of acute stroke after mechanical thrombectomy and compare with the traditional regression model.Methods Clinical and imaging data of 216 acute stroke patients in Nanjing First Hospital were retrospectively collected.Multiple logistic regression analysis was used to screen the independent predictors of stroke outcome and construct the prediction model of stroke outcome.The lesion features on multimodal MRI were extracted and screened.Support vector machine classifier was used to construct the prediction model.The performance of two model were compared using statistics analysis.Results Logistic regression analysis showed that NIHSS score on admission,HIR and complete recanalization were associated with stroke outcome(P<0.05).ROC showed that the AUC of statistical regression model for predicting the outcome of acute stroke was 0.775.The AUC of machine learning model for predicting outcome in the training set was 0.991;and the AUC of machine learning model for predicting outcome in the test set was 0.950.There was statistical difference between statistical regression model and machine learning model(Z=3.146;P<0.001).Conclusion The application of machine learning algorithm can accurately predict the outcome of acute stroke after mechanical thrombectomy,which is significantly better than the traditional regression model,and can provide guidance for the formulation of clinical treatment plan.