BP神经网络修剪算法筛选白血病预后危险因素

Application of Pruning Algorithms in BP Networks to Screen the Factors Affecting the Leukemia Prognosis

目的 通过单层BP网络的修剪算法 ,进行白血病预后危险因素筛选 ,并讨论修剪算法在医学统计中的应用及其与逐步logistic回归的联系。 方法 对上海市 1985 - 1995年间部分初发白血病病人的 6个月预后及可能的影响因素进行分析 ,分别用修剪算法和逐步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 利用修剪算法 ,可得到与逐步logistic回归相同的BP模型结构 ;应用不同的修剪参数得到含不同连接的BP网络模型 ,最终稳定于含 10个连接的模型。所有修剪的BP网络对测试集的判别效果均好于逐步logistic回归。 结论 修剪算法可以进行变量筛选 ,并可应用于弱影响因素的探索。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 。

Purpose: To apply the pruning algorithms in the risk factors' screening that may influence the leukemia prognosis in six months, 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: Information of leukemia patient during 1985 - 1995 in Shanghai was collected, including prognosis in six months and some possible risk factors. Logistic model and feedback- propagation networks with pruning were used in classification and predication of the prognosis. Results: After pruning, one BP network model with the same factors as logistic regression has been gotten. Several BP network models with different links are generated when using different pruning parameter. All the BP network models give better predication of test data than that of logistic model. A BP network model with ten input units generated. Conclusions: The pruning algorithms in single layer BP network may be used to variable screening and finding feebleness relationship between the input unit and output unit. The weights after pruning can reflect the effect of input unit to output units, just like the coefficient in logistic model.