Figure 4

Random Forest (RF) classification. An RF classifier based on two-thirds of the sample population (n=58) was first created to rank the top 300 masses differentiating controls from PC patients (all p<0.001, FDR corrected) based on their contribution to the classification accuracy and the Gini index (See Additional file 3). The top 11 masses (A) based on these criteria were selected and a classifier created to predict the identity of the blinded test set samples as shown in B. C, ROC curve (with AUC ±95% CI) based on the RF predicted PC probabilities for each of the 32 blinded samples.