Figure 2
The 15 candidate serum metabolic biomarkers selected by adaptive LASSO multinomial regression and random forest algorithms. (a) Plots for adaptive LASSO multinomial regression coefficients over different values of the penalty parameter λ. (b) Cross-validation plots for the penalty parameter λ. The dashed line left represents the minimum λ. The 29 candidate metabolic signatures mapping the minimum λ (0.0.4) were subjected to the next analysis. (c) The confusion matrix of the internal cross-validation set shows 29 HC, 38 UA, and 75 AMI are correctly classified by the adaptive LASSO multinomial algorithm. The darker the color represents, the more correctly it is classified. (d) The correlation plots between the number of random forest trees and the model classification error. The error stabilized when using 500 trees. (e) The top 30 discriminant metabolic signatures are ranked in descending order of importance to the accuracy of the RF classifier for HC, UA, and AMI. The bar lengths indicate the importance of the signature. The insert represents a tenfold cross-validation error as a function of the optimal number of input signatures used to fit the RF classifier. The number of signatures against the cross-validation error curve reaches the inflection point when using 47 signatures. (f) The confusion matrix of the training set (n = 113) based RF classifier shows 94 subjects are correctly classified. (g) The multinomial receiver-operating characteristic (ROC) curves are used to distinguish HC, UA, and AMI in the internal test set (n = 37) based RF classifier. (h) The Venn diagram shows the shared 15 candidate serum biomarkers selected by adaptive LASSO multinomial regression and RF algorithms.
