修剪算法在医学多因素分析中的应用探讨  

Approach of the Pruning Algorithms Using in Medicinal Multi-variable Analysis

作　　者：邓伟[1] 金丕焕[1] 

机构地区：[1]复旦大学公共卫生学院卫生统计与社会医学教研室,200032

出　　处：《中国卫生统计》2002年第5期286-288,共3页Chinese Journal of Health Statistics

摘　　要：目的　本文主要讨论修剪的单层BP网络的应用及其与逐步logistic回归的联系。 方法　搜集 2 0 5 2例急性阑尾手术术后病人的预后及可能的影响因素 ,分别用ANN和逐步logistic回归拟合不同的模型 ,利用ROC曲线下面积比较各个模型的判别和预测效果。结果　通过修剪 ,可得到与逐步logistic回归相同的ANN模型结构 ;应用不同的修剪参数得到不同的ANN网络模型 (Net1-5 )并可趋于稳定的结构。对测试集的判别效果 ,逐步logistic回归 <未修剪的ANN<修剪的ANN。结论　修剪的单层BP网络的权重系数与逐步logistic回归的回归系数相同 ,具有流行病学含义 ,且其判别效果好于逐步logistic回归。对修剪结果的分析提示我们 ,修剪算法可以应用于弱影响因素的探索。Objective The main object is to discuss the using of pruning neural networks in medicine analysis and find out the relationship between pruning neural network model and stepwise logistic model.Methods Total of 2052 appendectomy patients' information is collected, including prognosis and possible effect factors. Logistic model and feedback-propagation networks with pruning were used to classification and predication of the appendectomy prognosis.Results The area under the receiver operating characteristic curve of pruning network model and logistic model are 0.845 and above 0.85 respectively.Conclusion The performance of pruning network model is very similar to logistic model. The weights of pruning network model can reflect the effect of input unit to output units, just like the coefficient in logistic model. It may use pruning to find some feebleness relationship between the input unit and output unit.

关 键 词：人工神经网络 修剪算法 逐步logistic回归 

分 类 号：R195[医药卫生—卫生统计学]