Interpreting diagnostic accuracy studies for patient care

Abstract

Diagnostic test accuracy studies need to provide evidence in a comprehensible and intuitive format that facilitates choice of test for clinicians, their patients, and healthcare providers. Results should be reported in the context of clinical management decisions made at clinically sensible and important thresholds, preferably in terms of patients. For comparisons of tests, differences in true positive and false positive diagnoses should be reported, and it is important that any overall measures of diagnostic accuracy should incorporate relative misclassification costs to account for the fact that false negative and false positive diagnoses are rarely clinically equivalent. Measures need to be interpreted at a disease prevalence that reflects the real clinical situation. Analyses based on net benefit measures achieve these aims. In contrast, methods based on ROC AUC often incorporate thresholds that are clinically nonsensical, do not account for disease prevalence, and cannot account for the differing clinical implications of false negative and false positive diagnoses. We therefore caution researchers against solely reporting ROC AUC measures when summarising diagnostic performance, and caution healthcare providers against using ROC AUC alone to inform decisions regarding diagnostic performance. We recommend that diagnostic accuracy is presented by using paired measures with clinical context or using net benefit measures with their associated paired measures. © BMJ Publishing Group Ltd 2012.