Measures of Performance and Clinical Superiority Thresholds for 'Test-and-treat' Predictive Biomarkers.

Abstract

Predictive biomarkers are intended to predict an individual's expected response to specific treatments. These are an important component of precision medicine. We explore measures of biomarker performance that are based on the expected probability of response to individual treatment conditional on biomarker status. We show how these measures can be used to establish thresholds at which testing strategies will be clinically superior. We used a decision model to compare expected probabilities of response of treat-all and test-and-treat strategies. Based on this, R-Shiny-based apps were developed which produce plots of the threshold positive and negative predictive values or sensitivities and specificities above which a 'test-and-treat' strategy will outperform a 'treat-all' strategy. We present a case study using data on the use of RAS status to predict response to panitumumab in metastatic colorectal cancer. Where a companion diagnostic is predictive of response to one of the treatments being compared, it is possible to estimate threshold sensitivities and specificities above which a testing strategy will outperform a treat-all strategy, based only on the odds ratio of response. Where negative and positive predictive values were used, the threshold depended on the prevalence of the biomarker-positive patients. These intuitive performance measures for predictive biomarkers, based on expected response to individual treatments, can be used to identify promising candidate companion diagnostic tests and indicate the potential magnitude of the net benefit of testing.