Meta-analysis of diagnostic test accuracy
What question does this help answer?
Across the available studies, how accurately does this test identify the target condition, how much does its accuracy vary between settings, and why?
When to use it
Use it within a systematic review of diagnostic accuracy when studies evaluate the same index test for the same target condition against an acceptable reference standard, and report the counts of true and false positives and negatives or enough to reconstruct them.
When not to use it
Do not pool studies that used different target conditions, reference standards that are not comparable, or spectrum-distorting designs such as recruiting obvious cases and healthy volunteers, without examining what that does to the estimate. Accuracy alone also does not show that using a test improves decisions or patient outcomes.
Why the two measures are modelled together
Sensitivity and specificity trade off: a study that calls more results positive gains sensitivity and loses specificity. Studies differ in their thresholds, explicitly or implicitly, so the two estimates are correlated across studies. Pooling each separately ignores that correlation and can produce a pair of summary values that describes no real threshold. Hierarchical models that analyse them jointly, the bivariate model and the hierarchical summary ROC model, are closely related and account for this.
What the threshold trade-off looks like
Key assumptions and design requirements
- The target condition, index test and reference standard are defined precisely enough that the included studies answer the same question.
- Risk of bias and applicability are assessed per study with a tool designed for diagnostic accuracy, such as QUADAS-2.
- The model accounts for the correlation between sensitivity and specificity and for between-study variation in both.
- Where studies used different thresholds, the review says whether it summarises a single threshold or the summary curve across thresholds, because they answer different questions.
- Sources of heterogeneity, such as setting, prevalence or reference standard, are explored with prespecified covariates rather than after inspecting the results.
Why Mallard does not run it
Mallard does not perform diagnostic meta-analyses or size systematic reviews. Its diagnostic 2×2 calculator summarises the counts from a single study and is not a substitute for a hierarchical synthesis across studies.
Clinical example
A review team plans to summarise the accuracy of a point-of-care ultrasound finding for a condition in emergency department patients. They define the reference standard in advance, plan QUADAS-2 assessment, choose a bivariate model, and prespecify exploring whether accuracy differs between studies with consecutive recruitment and those with case-control sampling. No pooled accuracy is shown; the example illustrates planning choices only.
What to watch out for
Sensitivity and specificity move together as the threshold for a positive result varies, so pooling each separately ignores that correlation and can describe a combination no study reported. The summary depends on the thresholds, populations and reference standards of the included studies, and a pooled estimate across very different settings may apply to none of them.
Before using this in your workspace
Record the target condition, index test and threshold handling, the acceptable reference standards, the risk-of-bias and applicability tool, the hierarchical model chosen, the prespecified sources of heterogeneity, and how the results will be translated into expected numbers of true and false results at plausible prevalences.
- Bivariate analysis of sensitivity and specificity produces informative summary measures in diagnostic reviews (opens in a new tab) · J Clin Epidemiol · doi 10.1016/j.jclinepi.2005.02.022
- A unification of models for meta-analysis of diagnostic accuracy studies (opens in a new tab) · Biostatistics · doi 10.1093/biostatistics/kxl004
- Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies: The PRISMA-DTA Statement (opens in a new tab) · JAMA · doi 10.1001/jama.2017.19163