Estimating the optimal threshold for a diagnostic biomarker in case of complex biomarker distributions
1 Université de Lyon, F-69000 Lyon, France
2 Université Lyon 1, F-69622 Villeurbanne, France
3 CNRS, UMR5558, Laboratoire de Biométrie et Biologie Evolutive, F-69622 Villeurbanne, France
4 Hospices Civils de Lyon, Service de Biostatistique, 162 avenue Lacassagne, F-69003 Lyon, France
BMC Medical Informatics and Decision Making 2014, 14:53 doi:10.1186/1472-6947-14-53Published: 14 June 2014
Estimating the optimal threshold (and especially the confidence interval) of a quantitative biomarker to be used as a diagnostic test is essential for medical decision-making. This is often done with simple methods that are not always reliable. More advanced methods work well but only for biomarkers with very simple distributions. In fact, biomarker distributions are often complex because of a natural heterogeneity in marker expression and other heterogeneities due to various disease stages, laboratory equipments, etc. Methods are required to estimate a biomarker optimal threshold in case of heterogeneity and complex distributions.
A previously described Bayesian method developed for normally distributed biomarkers is applied to two flexible distributions; namely, a Student-t and a mixture of Dirichlet processes. Here, numerical studies assess the adequacy of the previous method with both distributions. Two applications are presented: the diagnosis of treatment failure after prostate cancer treated by ultrasound and the early diagnosis of cancers of the upper aerodigestive tract.
Bayesian inference provided reliable credible intervals in terms of bias and coverage probability. The two distributions analysed gave meaningful clinical interpretations in both applications.
Reliable methods can be used to estimate a biomarker optimal threshold, even in case of complex distributions.