ROC analysis using covariate balancing propensity scores with an application to biochemical predictors for thyroid cancer

Biomarker evaluation is important for diagnosing clinical diseases. Covariate adjusted receiver operating characteristic (ROC) regression has been used to identify significant biomarker candidates. Here, we show that the statistical significance of a biomarker can be affected by its prevalence. We propose a novel method that incorporates covariate prevalence information in the ROC regression. This approach is based on covariate balancing propensity scores proposed by Imai and van Dyk. Our method produces higher AUC values, demonstrating improved discrimination ability compared to direct ROC regression or unadjusted ROC analysis; this method can be used to improve biomarker development and can be implemented by an artificial intelligence (AI) system. Extensive simulation studies and data from a thyroid cancer study illustrate the advantages of our approach.