Recent Developments in Biomarker Evaluation with Applications in Lung and Ovarian Cancer
MetadataShow full item record
In the field of diagnostic studies, ROC curve and related summary measures are commonlyused for the purpose of biomarker evaluation under binary classification. In practice,many disease processes naturally involve more than two classes and there exist manypossible orderings among these disease classes. Among multi-class orderings, simple (i.e.ordinal) ordering and tree (or umbrella) ordering are most widely studied in statisticalliterature. This thesis research consists of three components: 1) we propose a newmeasure, i.e. the integrated false negative rate under tree ordering (ITFNR) as anadditional diagnostic measure besides the AUC under tree ordering (TAUC), and presentthe idea of using (TAUC, ITFNR) to evaluate the diagnostic accuracy under tree or umbrellaordering; 2) we investigate the issues arising from the commonly used pooling strategy inbiomarker evaluation for differentiating two major groups (e.g. non-cancer group andcancer group) where each major group involves at least one subclass, and propose newappropriate diagnostic measures for the purpose; 3) to overcome the shortcomingspossessed by existing diagnostic measures under ordinal ordering, with K stages, wepropose a novel idea for developing diagnostic accuracy measures by partioning thewhole classification problem into K – 1 steps. The resulting new diagnostic accuracymeasures have flexibility accommodating weights at each step of classification, henceenabling differentiation among certain biomarkers for which the existing measures fail totell the difference. Datasets from lung and ovarian cancer studies are analyzed using theproposed measures and methods.