logistic回归、决策树和神经网络在脑卒中高危筛查中的性能比较

Comparison of screening group with high risk of stroke among logistic regression, decision trees and neural networks

目的利用logistic回归模型、决策树C 5.0和神经网络进行脑卒中高危筛查性能的比较,探讨不同模型在疾病中的应用效果。方法所有数据来自上海市心脑血管疾病监测系统和闵行区居民健康档案系统,采用病例对照研究方法,将上海市闵行区2014年1月1日至12月31日所有40岁以上、闵行区常住户籍的1 391例新发脑卒中患者纳入病例组,同时从居民电子档案系统和2014年闵行区居民危险因素调查中抽取≥40岁1 388名常住户籍居民作为对照组,用R统计软件包检测3种模型对脑卒中筛查的准确率。结果 logistic回归模型、决策树C 5.0和神经网络3种模型的精确度分别是71.3%、74.5%和73.5%,灵敏度分别为70.9%、74.8%和67.0%,特异度分别为71.6%、74.2%和80.0%。3种筛查模型的受试者工作特征曲线下面积(AUC)分别为0.801、0.822和0.805,在筛查效果上差异无统计学意义(P>0.05)。3种模型中均有统计学意义的影响因素为职业、是否患有高脂血症和高血压、是否有规律的体育锻炼、高血压家族史。结论 3种模型在筛查效果上无明显差异,筛查的危险因素有所差异,在研究筛查技术时,将不同的方法进行合理的比较,应该根据实际情况结合专业背景进行。

Objective To compare the effects on screening group with high risk of stroke among logistic regression, decision trees C5.0 and neural networks, and to explore the utilization effectiveness of different models. Methods All data were from Shanghai cardiovascular and cerebrovascular diseases monitoring system and Minhang e-health records system. Case control study was performed, 1 391 patients（〉40 years old） diagnosed with stroke from Jan. 1 of 2014 to Dec. 31 of 2014 served as the subjects, and1 388 disease-free residents（〉40 years old） served as the controls. The information was collected from the health system. R software was used to detect the accuracy of three methods for screening the risk factors of stroke. Results The accuracies of logistic regression, decision trees C 5.0 and neural networks were 71.3%, 74.5% and 73.5%, respectively. The sensitivities of 3 models were 70.9%,74.8% and 67.0%, respectively. The specificities of 3 models were 71.6%, 74.2% and 80.0%, respectively. The results showed the areas under ROC curve（AUC） of three models were 0.801, 0.822 and 0.805, respectively. There was no significant difference of screening effectiveness（P〉0.05）. The influencing factors of 3 models were occupation, hyperlipidemia, hypertension,regular exercise and the hypertension family history. Conclusion There is no significant difference of screening effectiveness, but there is significant difference of screening risk factors among three models. When studying the screening techniques, the different methods should be compared reasonably, according to actual condition and specialty background.