Institute of Functional Genomics, University of Regensburg, Josef-Engert-Strasse 9, 93053 Regensburg, Germany.
J Proteome Res. 2012 Dec 7;11(12):6242-51. doi: 10.1021/pr3009034. Epub 2012 Nov 12.
Nontargeted metabolite fingerprinting is increasingly applied to biomedical classification. The choice of classification algorithm may have a considerable impact on outcome. In this study, employing nested cross-validation for assessing predictive performance, six binary classification algorithms in combination with different strategies for data-driven feature selection were systematically compared on five data sets of urine, serum, plasma, and milk one-dimensional fingerprints obtained by proton nuclear magnetic resonance (NMR) spectroscopy. Support Vector Machines and Random Forests combined with t-score-based feature filtering performed well on most data sets, whereas the performance of the other tested methods varied between data sets.
非靶向代谢物指纹分析越来越多地应用于生物医学分类。分类算法的选择可能对结果有相当大的影响。在这项研究中，我们采用嵌套交叉验证来评估预测性能，系统比较了 6 种二进制分类算法与不同的数据驱动特征选择策略相结合，对由质子核磁共振（NMR）光谱获得的尿液、血清、血浆和牛奶一维指纹图谱的 5 个数据集进行分类。支持向量机和随机森林与基于 t 分数的特征过滤相结合，在大多数数据集上表现良好，而其他测试方法的性能在数据集之间有所不同。
