Non-linear Analysis of Medical and Pharmaceutical Data using Neural Network and Generalized Additive Model

我们尝试将GAM、MARS、决策树、MART、分层人工神经网络（HANN）、Livingstone型人工神经网络（LANN）等几种非参数数据分析方法应用于医学和制药领域的数据分析。得到了合理的回归分析和主成分分析结果。首先，我们比较了Freedman等人使用理想人工数据开发的MARS和MART两种非参数非线性回归方法。MARS的拟合和预测性能出奇地好；MART的表现也相当不错。结果表明，MART是一种鲁棒的数据挖掘方法。因此，我们得出结论，对于噪声较小的数据应采用MARS，对于可能受噪声影响的数据分析应采用MART。然后，我们将这些方法包括GAM和HANN应用于一些流行病学数据集。例如，MARS a . More和MART方法应用于雌激素的药物流行病学研究，雌激素可能与子宫内膜癌的发生有关。结果表明，雌激素的活性形式与癌症的发生有关。虽然我们可以得到雌激素药物流行病学研究的一定信息，但在这种非参数模型中，要检验每个预测变量的显著性并不容易。因此，为了检验预测变量的显著性，我们引入了经我们小组修订的扩展移位检验方法。并将这些方法应用于食管癌发病与饮酒关系的临床流行病学研究。我们证实了饮酒影响食道癌的可能性。此外，我们还将LANN应用于气液色谱保留数据的非线性主成分分析。该方法的分类结果比线性主成分分析法更清晰。少

We have tried to adopt the several nonparametric data analysis methods, such as GAM, MARS, decision tree, MART, hierarchical artificial neural networks (HANN), and Livingstone type artificial neural networks (LANN), to the data in the field of medical and pharmaceutical sciences. And the reasonable results of regression and principal component analyses were obtained.First, we compared the two nonparametric nonlinear regression methods, MARS and MART, which were developed by Freedman et al., using ideal artificial data. MARS showed a surprisingly good fitting and prediction performances ; MART also showed quite good performances. And the results indicate that MART is a robust data mining method. Thus, we resulted that the MARS should be adopted for the data containing little noise and that the MART should be adopted for the data analyses which can be affected by noise.Then, we applied these methods including the GAM and the HANN to some epidemiological data sets. For example, the MARS a … More nd the MART methods were applied to the pharmacoepidemiological study of estrogen, which might be related to the onset of endometrial cancer. The results show that the active form of estrogen is related to the onset of the cancer. Although we could get a certain information of the pharmacoepidemiological study of estrogen, it was not so easy to test the significance of each prediction variables used in such nonparametric models. Thus, we introduced the extended shift test method, which was revised by our group, in order to test the significance of prediction variables. These methods were also applied to the clinical epidemiological study on the relation between onset of esophagus cancer and alcohol consumption. We confirmed the possibility that the alcohol consumption affect the esophagus cancer.In addition, we applied the LANN to the nonlinear principal component analyses of Gas-Liquid chromatographic retention data. The clearer classified result than the one obtained by linear PCA method could be obtained. Less