Fig. 3
CCS performance on randomized subsets of testing cohort and validation cohort. a Receiver operating characteristic (ROC) plots for 150 runs of the EpiSwitchâ„¢ logistic classifier. Data for the 59 patient cohort was randomised 150 times using the WEKA sample randomisation function. This reorders the data prior to splitting in developing the training set, ensuring that the same starting point for the classifier is not used and allowing multiple accuracy calculations for the same data. The average area under the curve (AUC) for the 150 classifier runs was 0.90. The plot is the average ROC from the 150 test results. b Receiver operating characteristic (ROC) plots for the EpiSwitchâ„¢ logistic classifier run on the blinded validation cohort of 19 RA patients. The classifier had a sensitivity of 75.0% and specificity of 85.7% with an AUC of 0.91