Building Classification Models with Combined Biomarker Tests: Application to Early Detection of Liver Cancer.

Abstract

Early detection of hepatocellular carcinoma (HCC) is critical for the effective treatment. Alpha fetoprotein (AFP) serum level is currently used for HCC screening, but the cutoff of the AFP test has limited sensitivity (~50%), indicating a high false negative rate. We have successfully demonstrated that cancer derived DNA biomarkers can be detected in urine of patients with cancer and can be used for the early detection of cancer (Jain et al., 2015; Lin et al., 2011; Song et al., 2012; Su, Lin, Song, & Jain, 2014; Su, Wang, Norton, Brenner, & Block, 2008). By combining urine biomarkers (uBMK) values and serum AFP (sAFP) level, a new classification model has been proposed for more efficient HCC screening. Several criterions have been discussed to optimal the cutoff for uBMK score and sAFP score. A joint distribution of sAFP and uBMK with point mass has been fitted using maximum likelihood method. Numerical results show that the sAFP data and uBMK data are very well described by proposed model. A tree-structured sequential test can be optimized by selecting the cutoffs. Bootstrap simulations also show the robust classification results with the optimal cutoff.