Diagnostic Test Accuracy and Clinical Decision Making

Selecting a diagnostic test is a complex and sometimes bewildering task. Systematic reviews of diagnostic test accuracy provide only some of the essential information that physicians need. Choice of a diagnostic test is embedded within a dynamic multistage screening and decision-making context in which past clinical assessments and review of personal and family medical history determine the need for testing. For diagnostic tests, it is important to understand not only the test characteristics, such as sensitivity and specificity, but also the relative benefits and harms from actions in response to test results for subgroups of patients and the individual patient. Authors of systematic reviews of diagnostic test accuracy need to see beyond reporting the summary receiver-operating characteristic (ROC) curves and summary measures of diagnostic accuracy. By examining the expected downstream harms and benefits of positive and negative test results, authors can link diagnostic accuracy to clinical decision making.

Approaching a systematic review of test accuracy within this context has substantial implications for framing study questions, setting parameters for literature searches, and selecting studies to include in the systematic review. Variation in study design, reference standards, sources of bias, and multiplicity of measures makes synthesizing the evidence from a set of studies of diagnostic accuracy more complicated than analyzing treatment effects from a set of randomized, controlled trials. The complexity of the problem and the lack of standards for reporting and synthesizing studies of diagnostic accuracy add to the cognitive burden on authors, reviewers, editors, and readers.

Summarizing diagnostic accuracy measures presents particular methodological challenges. Because studies often choose different cut-points to define an abnormal result, sensitivity and specificity may vary between studies, obscuring the underlying relationship between sensitivity and specificity. Statistical methods can reveal this underlying relationship. Simple regression of the log diagnostic odds …