CLINICALLY RELEVANT USE OF PARTIAL ROC CURVES IN PERFORMANCE COMPARISON STUDIES

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): The Receiver Operating Characteristic curve (ROC) is the most commonly used approach for assessing and comparing performance levels of diagnostic procedures that may involve the observer as an integral part of the system / practice. One of the more clinically relevant methods of ROC analysis is based on the part of the ROC curve which describes operating characteristics that are expected to be actually observable in clinical practice. A well known example of such an approach is the analysis based on partial area under the ROC curve (partial AUC) over the range of operating points with low false positive fraction (FPF). Although analyses based on a part of the ROC curve are frequently considered to be more relevant, these are not commonly employed due to several important issues. Two of these are the difficulty with a consistent definition of the actual range of interest and the potential loss of efficiency, namely higher likelihood of inconclusive results of a study or the need for a larger sample size for a future study. Our research proposal is primarily related to: 1) Defining the appropriate portion of interest of an ROC curve when estimated for a laboratory experiment performed outside of the actual clinical environment (given that the region of interest is known for the clinical ROC curve). Indeed the same decision threshold in a retrospective laboratory experiment is likely to have different operating characteristics (i.e. FPF and TPF) and therefore the direct application of the range of interest defined in terms of FPF is inappropriate and could lead to inconclusive and less, or totally irrelevant, results; 2) Empirical operating points in performance assessment studies typically span only the region of low FPF and therefore consideration of the entire ROC curve rather than the empirically supported part of the curve may not actually increase the statistical efficiency of comparisons of diagnostic modalities. The research we propose herein will help researchers use partial ROC analysis in a more relevant manner leading to more conclusive results, as well as provide guidelines on the potential statistical tradeoffs between partial and full ROC analyses in commonly encountered experimentally ascertained data, thereby addressing concerns about efficiency. Our proposed effort will provide researchers with three difference methods to define the clinically relevant portion on a laboratory-estimated ROC curve. We will also develop an approach for comparison of partial AUCs by combining strengths of both parametric and nonparametric approaches. This will provide a more interpretable and statistically efficient technique for this purpose. Last, we will investigate the relative efficiency of the partial and full AUC in different types of data commonly encountered experimentally where the empirical operating points span only the range of low FPF, and develop an approach that combines practical relevance of the partial AUC with statistical stability of the full AUC. All investigated analyses will be evaluated using extensive simulations, as well as an assessment of the possible effects, if any, of the proposed analyses on conclusions of previously performed and published large ROC type studies.