Figure 3

AUC analysis on the entire metabolite set (bars in the left), and on a set of the top ten ranked metabolites using four common feature selection methods, i.e. two sample t-test (P-value), the unpaired Biomarker Identifier (uBI), ReliefF, and Information gain (IG) on MCADD data (bars in the right). Red bars represent the predictive value expressed by the AUC of selected analyte sets, determined on a single derivation cohort with cross validation and blue bars without cross-validation. Interestingly, using the entire metabolite set (43 analytes) for distinguishing between the two groups, the discriminatory ability dropped from AUC = 1.0 (without cross validation) to AUC = 0.51 after 10-fold cross validation, thus indicating no discrimination between the cohorts. On the selected subset, the AUC dropped by 15% to 25% after cross validation, demonstrating weak predictive value and thus low generalizability of the selected subset in this experiment.