"Beyond the ROC Curves: A Statistical Approach to Infer on the Ordering of Biomarker Changes in Preclinical Alzheimer’s"
Hosted by the WUSTL Division of Biostatistics.
Abstract: In many irreversible chronic diseases such as Alzheimer’s, the preclinical stage, as defined by the time window during which no clinical symptoms are present but the underlying pathological processes of the diseases have already been active, is important because it likely offers an optimal time window for preventive and therapeutic intervention before it is too late. Because no clinical diagnosis of the diseases is available, biomarkers of different modalities are used to assess the preclinical changes of the diseases. In order to evaluate the ability of multi-modalities of biomarkers to track the preclinical disease change in the absence of a clinical diagnosis, we propose a nonparametric approach to assess the predictive ability of one biomarker against another more downstream biomarker. Because biomarkers are mostly continuous measures with very different measurement units and scales across modalities, the predictive ability of one biomarker against another must be invariant under monotonic transformations. We propose that, for each plausible threshold of a downstream biomarker Y that divides individuals into either positive or negative by their values, the Receiver Operating Characteristic (ROC) curves and the area under the curve (AUC) be used to summarize the power of an early biomarker X to predict the positivity of Y. When the prevalence of the positivity of Y varies with different thresholds, this process leads to a predictive accuracy curve (PAC) for biomarker X. We first propose a nonparametric estimator to the PAC as well as a pointwise confidence interval estimate, and then provide inferences on the difference between two PACs from two early biomarkers predicting the same downstream biomarker. Finally, we apply the proposed methodology to the biomarkers data collected by the Adult Children Study at Washington University on preclinical Alzheimer disease (AD), and estimate the relative PACs of several cerebrospinal fluid (CSF) and neuroimaging biomarkers in predicting a downstream biomarker or cognitive outcome. These results have the potential to shed light on the likely orderings of the biomarker changes during the preclinical stage of AD.
Copyright © 2009 | Contact Us | SitemapOffice of Neuroscience Research660 South Euclid Ave., Campus Box 8111, St. Louis, Missouri 63110Email: ONR@neuro.wustl.edu