Methodology for Evaluation of Diagnostic Performance

Abstract

The proliferation of expensive technology in diagnostic medicine demands objective, meaningful assessments of diagnostic performance. Receiver Operating Characteristic (ROC) analysis is now recognized widely as the best approach to the task of measuring and specifying diagnostic accuracy (Metz, 1978; Swets and Pickett, 1982; Beck and Schultz, 1986; Metz, 1986; Hanley, 1989; Zweig and Campbell, 1993), which is defined as the extent to which diagnoses agree with actual states of health or disease (Fryback and Thornbury, 1991; National Council on Radiation Protection and Measurements, 1995). The primary advantage of ROC analysis over alternative methodologies is that it separates differences among diagnostic decisions that are due to actual differences in discrimination capacity from those that are due to decision-threshold effects (e.g., ''under-reading'' or ''over-reading''). An ROC curve measures diagnostic accuracy by displaying True Positive Fraction (TPF: the fraction of patients actually having the disease in question that is diagnosed correctly as ''positive'') as a function of False Positive Fraction (FPF: the fraction of patients actually without the disease that is diagnosed incorrectly as ''positive''). Different points on the ROC curve--i.e., different compromises between the specificity and the sensitivity of a diagnostic test, for a given inherent accuracy--can be achieved by adopting different critical values of the diagnostic test's ''decision variable'' --e.g., the observer's degree of confidence that each case is positive or negative in a diagnostic image-reading task, or the numerical value of the result of a quantitative diagnostic test. ROC techniques have been used to measure and specify the diagnostic performance of medical imaging systems since the early 1970s, and the needs that arise in this application have spurred a variety of new methodological developments. In particular, substantial progress has been made in ROC curve fitting and in developing statistical tests to evaluate the significance of measured differences between ROC curves. These are especially important tasks in medical applications, because various practical issues usually limit the number of patients with clearly established diagnostic truth that can be included in any study that seeks to measure diagnostic performance objectively. Other progress has been made in relating ROC analysis to cost/benefit analysis, and in generalizing ROC methods to accommodate some diagnostic tasks where more than two decision alternatives are available. ROC analysis clearly provides the most rigorous and fruitful approach for such assessments but, like many other powerful techniques that provide useful insight concerning complex situations, it currently suffers from limitations, particularly in evaluation studies that involve small case samples. However, the potential of this relatively new analytic approach and the concepts on which it is based have not been fully explored. The research proposed here is designed to refine and supplement existing ROC methodology to increase both the accuracy and the precision of its results.