可解释的机器学习模型用于预测远期脑缺血事件

Interpretable machine learning model to predict long-term cerebral ischemic events

目的 利用可解释的机器学习模型,探讨临床特征及颈动脉斑块成分预测远期脑缺血事件的可能性及具体特征的重要性。方法 研究入组206例急性前循环脑梗死患者,半自动软件测量颈动脉斑块成分,记录临床特征,采用随机森林模型进行训练及检验数据。通过SHAP模型解释包对模型预测结果进行解读。结果 206例患者中,女性86例,平均年龄(62.5±11.9)岁,共有128个脑区出现新发缺血性脑卒中,160个脑区出现脑白质疏松进展。随机森林模型预测新发缺血性脑卒中的准确率为94.9%和90.8%;预测脑白质疏松进展的准确率为94.2%及85.0%。年龄、收缩压及糖尿病史是远期脑缺血事件最重要的特征。结论 可解释的机器学习可量化特征重要度及SHAP预测值,较好地评估单个样本远期脑缺血事件的风险.

Objective To investigate the possibility and significance of clinical features and carotid plaque components in predicting long-term ischemic events using interpretable machine learning model. Methods 206 patients with acute cerebral infarction caused by the occlusion of anterior circulation were included in this study. Computed tomography angiography(CTA) imaging features were accessed by semiautomatic analytic software and clinical features were recorded. Random forest model was applied to training and testing dataset, respectively. SHAP package was used to explain the prediction results. Results Of 206 cases, 86 was female. Mean age of all was(62.5±11.9) years. There were 128 hemispheres with new cerebral ischemic stroke and 160 hemipheres with progressed leukoaraiosis. The accuracies of prediction to ischemic stroke were 94.9% and 90.8%, while the accuracies of prediction to progressed leukoaraiosis were 94.2% and 85.0%. SHAP values indicated that age, systolic blood pressure and diabetes were the most important features to predict long-term ischemic events. Conclusion Interpretable machine learning model can quantify the feature importance and SHAP predicted value, and can be used to predict the risk of long-term cerebral ischemic events of single samples.