BP神经网络在代谢综合征影响因素分析中的应用

Application of Back-propagation Neural Network in Analysis of Influencing Factors for Metabolic Syndrome

目的探讨BP神经网络筛选疾病相关因素及构建疾病预测模型的作用,并与Logistic回归模型的分析结果进行对比,为更加准确地运用神经网络方法解决医学实际问题提供科学依据。方法利用MATLAB软件中的神经网络工具箱,建立代谢综合征相关影响因素的BP神经网络模型,通过计算平均影响值对影响因素进行筛选,依据ROC曲线下面积对比BP神经网络与Logistic回归分析所构建的疾病预测模型的效果。结果利用BP神经网络所构建的疾病预测模型,通过平均影响值算法筛选变量后的预测效果要好于未筛选的效果,且通过平均影响值算法所筛选的影响因素与Logistic回归分析基本一致,两者预测效果差异无统计学意义(AUCBP=0.837,AUCLogistic=0.841,u=0.3310,P=0.7406)。结论运用BP神经网络的平均影响值算法可实现对疾病相关因素的筛选及构建疾病预测模型,可在流行病学病因探索的研究中发挥与Logistic回归分析同样的作用。

Objective To explore the role of utilization of back-propagation（BP） neural network for screening the related factors of disease and constructing predictive models,to compare with the prediction with Logistic regression model, and to provide the scientific evidence for further correct implementation of neural network to solve practical problems in medical practice. Methods We built the predictive model of metabolic syndrome related risk factors by invoking BP neutral network toolbox from MATLAB.We selected the risk factors according to mean impact value,and then compared the performance of predictive models between BP neutral network and Logistic regression by calculating their area under ROC curves respectively. Results The predictive model of BP neutral network with mean impact value algorithm involved showed more reliable and robust results than the one without the algorithm;moreover,the risk factors selected by the former method were practically consistent with the results produced by Logistic regression,and no statistically significant difference was found（AUCBP=0.837,AUCLogistic=0.841,u=0.3310,P=0.7406）. Conclusions We can use BP neutral network based on mean impact value to select the influence variables and to construct disease forecast model.The results suggest that it plays as good role as the general Logistic regression in the use of cause exploration in epidemiological studies.