Characterization of technical uncertainty in the classification of centroid-based multivariate assays

Multivariate assays using gene expression as their contributing factors, such as the centroid-based PAM50 Breast Cancer Intrinsic Classi er, are becoming commonly used in assisting treatment decisions in medicine, especially in oncology. Although physicians may rely on these multivariate assays for planning treatment, little is known about the e ects on the results of an assay due to the intrinsic error in the laboratory process and measuring its contributing factors. While we expect that classi cation of samples in proximity to one of the centroids de ning the tumor classes will be stable with respect to experimental errors in the gene expression measurements, what happens to the samples not in proximity to a single centroid is unknown. Results reported to the attending physician may be misleading because he or she is receiving no information about the probability for sample misclassi cation. Given the serious consequences due to ambiguous results in clinical classi cations, methods to measure the e ects of a multivariate assay's intrinsic errors need to be established and communicated to attending physicians. In this study, a method to characterize the technical uncertainty in the classi cation of centroid-based multivariate assays, is developed and described, using the PAM50 Breast Cancer Intrinsic Classi er as the model multivariate assay. Furthermore, the described method provides a general and individual classi cation con dence measurement that advances multivariate assays towards personalized healthcare by providing personalized con dence measurements on the assay's result. Finally, this study explores whether using parametric versus nonparametric distance measurements is most e ective when using a single gene expression platform, such as microarray or Real-time, quantitative Polymerase Chain Reaction.