Figure 3

The identified panel of metabolites and its predictive performance. (a) Boxplots showing group differences and P values for identified metabolites across NCI (non-converters), NCI (converters), and MCI/AD for serum abundances. (b) ROC curves of metabolite models trained on the 70% training data and tested on the 30% testing data according to 100-times randomly training–testing splitting. (c) The ROC curve of the final metabolite model on the validation data. (d) RF scores of the final metabolite model across NCI (non-converters), NCI (converters), and MCI/AD. *P-value < 0.05, **P-value < 0.01, ***P-value < 0.001, Wilcoxon rank sum test. The optimal cutoff was determined by the Youden index. AD Alzheimer’s disease, AUC area under the receiver operating characteristic curve, NCI(C) NCI (converters), NCI(N) NCI (non-converters), CIs confidence intervals, MCI mild cognitive impairment, NS not significant, SE sensitivity, SP specificity.