Collaborative Research: Empirical Likelihood Based Statistical Methods for Diagnostic Systems

University Of Washington, Seattle WA

Investigators

Abstract

The receiver operating characteristic (ROC) curve methodology is the statistical methodology for assessment of the accuracy of prediction rules. It is commonly used in a wide range of scientific fields such as signal detection theory, medical imaging, weather forecasting, real and false alarms testing for anti-terrorist system, and diagnostic medicine. Motivated by assessing diagnostic accuracy of Dermoscope, prostate cancer screening test, and audio test and the need to find tractable and easily implemented solution to complex statistical problems in ROC studies, in this project the investigators develop new semi-parametric and nonparametric statistical methods for ROC analysis, particularly on the statistical inferences for the partial areas or the full areas under the ROC curves (AUC) and ROC and AUC regression models. In this project, the investigators extend the traditional empirical likelihood (EL) to the setting of ROC analysis in three directions: (1) to develop EL-based semi-parametric methods for the partial areas under the ROC curves, (2) to develop EL-based nonparametric methods for the partial or full areas under the ROC curves, and (3) to develop EL-based statistical methods for ROC and AUC regression models. The new methods are expected to be more robust and more accurate than the existing methods in evaluation of competing diagnostic systems. They have potentially better small sample performances than the existing methods. A software is developed for the actual use of the newly proposed methods for ROC analysis. The methods developed from this project are great addition to existing medical diagnostic testing and applicable to a wide range of scientific fields. High-tech diagnostic systems of various kinds around us have been made tremendous progress in the last two decades. Many new diagnostic systems have been used to reveal diseases in people, malfunctions in nuclear power plants, flaws in manufactured products, collision courses of aircraft, threatening activities of terrorists, etc. However, costs of diagnostic systems can be high. To select more accurate diagnostic system for widespread use, it is important to develop appropriate statistical methods for evaluating the diagnostic accuracy of competing systems. The use of attractive statistical methods like the semi-parametric and nonparametric methods for ROC analysis developed in this project is going to help diagnostic systems users make informed choices of the most reliable diagnostic systems. This may contribute to the reduction of health care costs in the long run.