BP神经网络与Logistic回归分度诊断模型效能的比较研究  

Comparative study of back propagation neural networks and Logistic regression models in graded diagnose

作　　者：王慧敏[1] 沈婷[1] 曾智杰[1] 邹相华 姜傥[1] 

机构地区：[1]中山大学附属第一医院检验科,广东广州510080 [2]中山大学数学与计算机科学学院应用数学系,广东广州510275

出　　处：《热带医学杂志》2016年第3期316-319,339,共5页Journal of Tropical Medicine

摘　　要：目的以子痫前期分度为例,比较反向传播(BP)神经网络模型与传统Logistic回归模型在复杂性疾病中的拟合效果及诊断效能。方法 2008年1月至2013年12月中山大学附属第一医院孕妇区收治的145例符合子痫前期诊断标准及实验入选标准的患者,其中轻度87例、重度58例;随机分组后,分成127例训练集和18例测试集以建立妊娠晚期子痫前期分度诊断的Logistic回归模型及BP神经网络诊断模型,并对模型的诊断效能进行比较,探讨子痫前期孕前和孕期高危因素。结果入选模型的变量有妊娠早期纤维蛋白原(Fbg)、血小板计数(PLT)、平均血小板体积(MPV)以及妊娠晚期尿蛋白;训练集中BP神经网络模型的一致率为80.30%,灵敏度为74.50%,特异度为84.21%,均高于Logistic回归模型的74.80%、58.82%、82.89%,且差异均有统计学意义(P<0.05)。BP神经网络模型的ROC曲线下面积为(0.887±0.029),大于Logistic回归模型的(0.823±0.036);在最佳诊断值界值的BP神经网络Youden index为67.0%,仍高于Logistic回归模型的54.1%,差异均有统计学意义(P<0.05)。结论 BP神经网络模型在妊娠晚期子痫前期分度中的拟合效果优于Logistic回归,更适合用于复杂性疾病多因素分析的研究;妊娠早期Fbg、PLT、MPV以及妊娠晚期尿蛋白与妊娠晚期子痫前期分度有关。Objective To compare the diagnostic efficiency between back propagation（BP） neural networks and Logistic regression models in grading complex diseases,illustrated by the example of preeclampsia. Methods 145 cases were selected from data of patients with preeclampsia during January 2008 and December 2013 in our hospital,including 87 mild cases and 58 severe cases. Then modeled BP ANNs and Logistic regression after 127 cases randomized to training set and the rest testing set and analyzed their efficiency in discriminating preeclampsia. Results Fibrinogen（Fbg） in early pregnancy,platelet count（PLT）,meam platelet volume（MPV） and late pregnancy proteinuriawere enrolled in models. The consistent rate,sensitivity and specificity of BP ANNs were 80.30%,74.50% and 84.21%,respectively,which were higher than those of Logistic regression（P〈0.05）. The area under the ROC curve of the BP ANNs model was（0.887±0.029）,more than（0.823±0.036） of the Logistic regression; in the best diagnostic value,the Youden index of BP ANNs was 67%,still above 54.1% of the Logistic regression model. All the differences were statistically significant（P〈0.05）.Conclusion BP ANNs has better simulating effect in division of preeclampsia than Logistic regression and more suitable for multifactorial analysis of complex disease; Fbg in early pregnancy,PLT,MPV and late pregnancy proteinuriaare associated with preeclampsia discrimination in late pregnancy.

关 键 词：BP神经网络 LOGISTIC回归 子痫前期 

分 类 号：R714.425[医药卫生—妇产科学]