Table 2.

A classification model that is predictive of skin sensitization was built using a support vector machine (SVM) and metrics identified through various feature selection methods. Classification performance in terms of accuracy, sensitivity, and specificity was determined by 10-fold cross validation. By utilizing all of the 27 cytokines from the Bioplex without performing any feature selection to build a predictive model, the accuracy, sensitivity and specificity of the classifier was very poor. Performing feature selection by ranking the margin distances from the SVM and p-values determined from ANOVA identified IL-12, IL-9, VEGF, IFN-γ as a molecular signature to build the classification model. This classification model performed superiorly as compared to a model built using features selected using hierarchical cluster analysis or by ranking the accuracies computed from the SVM. Data analyzed by all feature selection methods included all treatment conditions (untreated, vehicle, SA, IE, and PPD) and all of their respective concentrations for N = 3 independent replicates.

Feature selection method	Metrics	Accuracy (%)	Sensitivity (%)	Specificity (%)
No feature selection	All 27 cytokines from bioplex	75.00	67.00	83.00
P-values	IL-12, IL-9, VEGF, IFN-γ	92.00	92.00	92.00
Hierarchical cluster analysis	IL-4, IL-9, IL-12, VEGF	91.67	91.67	91.67
SVM margin distance	IL-12, IL-9, VEGF, IFN-γ	92.00	92.00	92.00
SVM accuracy	IL-12, IL-9, VEGF, PDGF	91.30	90.91	91.67