基于支持向量机的急性出血性脑卒中早期预后模型的建立与评价

Establishment and evaluation of early prognosis models of acute intracerebral hemorrhage based on support vector machine

目的：比较支持向量机（support vector machine,SVM）和传统的Logistic回归构建的急性出血性脑卒中（intracerebral hemorrhage,ICH）早期预后判别模型的预测性能,探索急性ICH预后研究的新方法。方法 ：收集急性ICH患者339例,随访观察21 d时的临床转归情况。应用随机数字法以3∶1的比例分为两组,一组作为训练样本用于筛选变量和建立预测模型,计254例;另一组作为验证样本,用于评价模型预测效果,计85例。建模方法采用SVM和常规统计方法中的Logistic回归。结果：通过对85例ICH患者的预测判别验证,SVM1的预测分类能力在4个模型为最强,4个模型预测的准确率和Youden指数分别为：Logistic回归：72.9%（62.0%~81.7%）、0.441（0.249~0.633）;SVM1：82.4%（72.3%~89.5%）、0.632（0.465~0.799）;SVM2：78.8%（68.4%~86.6%）、0.557（0.379~0.735）;SVM3：78.8%（68.4%~86.6%）、0.563（0.385~0.741）。结论：采用SVM能较好地判断急性ICH患者的早期预后,其效能优于Logistic回归模型。

Objective：To compare the performance of predictive models which were established by support vector machine（SVM）and traditional logistic regression and to study the new method of early prognosis in the patients with ICH. Methods：Totally 339 patients with ICH were collected and followed up the clinical outcomes for 21 days. Using the random number method,the original sample was divided into two groups according to the proportion of 3 ∶1. One group（254 cases） was regarded as a training set for screening the variables and establishing the prediction model and the another group（85 cases） was used as validation set for evaluating the model effect. SVM and the conventional statistical methods of logistic regression were used to construct the predictive models. Results：Through the discriminant validation of the forecast of 85 patients with ICH,the predictive ability of SVM1 was the strongest in the four models. The accuracy and Youden index of four models were as follows,logistic regression：72.9%（62.0%~81.7%）,0.441（0.249 ~0.633）;SVM1：82.4%（72.3% ~89.5%）,0.632（0.465 ~0.799）;SVM2：78.8%（68.4% ~86.6%）,0.557（0.379 ~0.735）;SVM3：78.8%（68.4%~86.6%）,0.563（0.385~0.741）. Conclusion：The model based on SVM could better predict the early prognosis of the patients with ICH. The efficacy of SVM model is superior to that of logistic regression model.