A new evaluation measure of diagnostic tests based on modified area under the receiver operating characteristic curve

摘要

In this paper, a new diagnostic/classification measure based on modified AUC (MAUC) is proposed. By this measure, we penalize the margin of features between diseased and non-diseased groups in AUC. It's threshold independent, and under normal distribution assumption, we can prove that higher MAUC always means higher PAUC (within relatively low FPR) when AUCs are close to each other. Our simulations and experiment about prostate cancer MS data can also help to demonstrate it.