Comprehensive comparison of classifiers for metabolic profiling analysis

Metabonomics is an emerging field providing insight into physiological processes and difference. Besides conventional PCA, PLS and OPLS approaches, more and more machine learning classifiers are likely to become the supplements for metabolic profiling data analysis. A comprehensive comparison of PLS, support vector machine (SVM, with linear and quadratic kernels), linear discriminant analysis (LDA), and random forest (RF) was reported applying on clinical metabonomics data. The accuracy of these classifiers was tested by 7-fold and holdout Cross Validation. Their stability and over fitting were evaluated by holdout Cross Validation and permutation (repeated 100 times). Their prediction ability was investigated by ROC curve, and their sensitivity on irrelevant variables was studied by variable ranking combining selection step by step. The overall performance of RF and SVM (linear kernel) is superior to the others. Some selected variables are of significance for further research on metabolic difference.