Diagnostic Test Accuracy Review

A diagnostic test accuracy (DTA) review is a specialized type of systematic review that evaluates the performance of one or more diagnostic tests by synthesizing data on sensitivity, specificity, and other accuracy metrics across multiple primary studies. DTA reviews use distinct methods from intervention reviews, including paired meta analysis of sensitivity and specificity, summary receiver operating characteristic (SROC) curves, and investigation of sources of heterogeneity in test performance. Cochrane DTA reviews follow a dedicated handbook that addresses the unique methodological challenges of diagnostic evidence synthesis.

Individual diagnostic accuracy studies are conducted in specific clinical settings, with particular patient populations, and with varying reference standards. Their results may not generalize to other contexts. DTA reviews address this limitation by pooling evidence across studies, producing summary estimates of test performance and examining how accuracy varies according to patient characteristics, test protocols, and reference standards. Guideline developers and health technology assessment bodies rely on DTA reviews to determine whether a test should be recommended for clinical use, to compare competing tests, and to identify gaps in the evidence base.

The process begins, like any systematic review, with a focused research question, often framed using the Joanna Briggs Institute's PIRD framework (Population, Index test, Reference test, Diagnosis of interest). A comprehensive literature search, including bibliographic databases, trial registries, and grey literature, identifies potentially relevant studies. Screening and inclusion follow predefined eligibility criteria, and risk of bias and concerns about applicability are assessed with a tool specifically designed for diagnostic accuracy studies: the widely used QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2), or its successor QUADAS-3, published in February 2026, which its developers now recommend.

Data extraction captures 2 by 2 table data (true positives, false positives, true negatives, false negatives) from each study, along with study characteristics that may explain variation in test performance. Because sensitivity and specificity are linked through the test threshold (a threshold that raises sensitivity lowers specificity, and vice versa), they tend to be negatively correlated across studies, so pooling each one separately with standard meta analytic methods designed for single outcomes ignores this relationship.

Instead, DTA meta analyses use bivariate models that jointly estimate summary sensitivity and specificity while accounting for their correlation and between study heterogeneity. The bivariate model produces a summary point (the estimated sensitivity and specificity) with a confidence region (reflecting precision) and a prediction region (reflecting expected variation across settings). Alternatively, the hierarchical SROC (HSROC) model estimates a summary ROC curve, which suits studies that used different thresholds; a single summary point is meaningful only for studies that share a common threshold. Without covariates, the bivariate and HSROC models are mathematically equivalent. Software such as MetaDTA, mada (in R), and the user written metandi command for Stata implement these models.

Investigation of heterogeneity is a critical component. Sources of variation may include patient spectrum (disease severity, comorbidities), test protocol (cutoff values, sample handling), reference standard (the method used to confirm diagnosis), and study design features. Subgroup analyses or meta regression can explore these sources.

A DTA review is appropriate when multiple primary studies have evaluated the accuracy of a diagnostic test and there is a need for summary estimates or comparison across studies. It is particularly valuable when individual studies show variable results and the reasons for variation need to be explored, when a new test is being compared to existing alternatives, or when clinical guidelines need evidence on test performance. It is not appropriate when the research question concerns the clinical impact of testing (whether testing improves patient outcomes), which requires a different review design focused on randomized test and treat trials.

A frequent error is applying standard meta analytic methods designed for single effect sizes (such as pooling sensitivity estimates independently) without accounting for the correlation between sensitivity and specificity. Another mistake is combining studies that used different test thresholds without adjusting for threshold effects, which can produce misleading summary estimates. Failing to use a tool designed for diagnostic accuracy studies, such as QUADAS-2 or its successor QUADAS-3, or substituting generic quality tools designed for intervention studies, misses bias domains specific to diagnostic studies. Pooling accuracy data across studies with different reference standards introduces reference standard heterogeneity that can bias summary estimates. Finally, some reviewers present summary sensitivity and specificity without the prediction region, giving readers an overly precise impression of expected test performance in new settings.

A hypothetical DTA review evaluates the accuracy of procalcitonin for diagnosing bacteremia (bacteria in the bloodstream) in adults with suspected sepsis, tested in emergency departments or intensive care units. The review includes 34 studies with a total of 12,400 patients, all using blood culture as the reference standard. Because the studies used different procalcitonin cutoffs, the review does not pool them into a single summary point. For the studies using a cutoff of 0.5 ng/mL, the bivariate meta analysis produces a summary sensitivity of 77% (95% CI: 72 to 81%) and a summary specificity of 79% (95% CI: 74 to 83%). The prediction region is substantially wider, indicating that accuracy varies considerably across settings. Meta regression shows that sensitivity at this cutoff is higher in intensive care unit populations (84%) than in emergency department populations (71%). Studies using a cutoff of 2.0 ng/mL report lower sensitivity but higher specificity than those using 0.5 ng/mL, so an HSROC analysis is used to summarize accuracy across all the cutoffs.