The Impact of Prospective versus Retrospective Study Design on Diagnostic Test Accuracy in Medical Imaging: A...

Diagnostic test accuracy (DTA) estimates may be influenced by whether data are collected prospectively or retrospectively, but previous meta-epidemiological evaluations across diverse diagnostic tests have reported inconsistent findings. Radiology provides a distinctive context because imaging tests are often based on protocolised acquisition and stored image datasets. We evaluated whether prospective versus retrospective study design was associated with reported sensitivity and specificity in radiology DTA studies. Secondary analyses examined study-level factors as contextual contributors to heterogeneity.

We included systematic reviews with meta-analyses published in radiology-related journals during the five years preceding the search date and containing at least 10 primary medical imaging DTA studies. Primary studies were classified as prospective or retrospective according to design labels reported in the included reviews. Complete 2 × 2 contingency-table data were extracted. A Bayesian three-level bivariate generalised linear mixed model jointly modelled sensitivity and specificity using binomial likelihoods while accounting for clustering of primary studies within systematic reviews. Effects were reported as posterior mean logit differences with 95% credible intervals, comparing retrospective with prospective studies.

We included 50 reviews comprising 790 primary studies, of which 332 (42.0%) were prospective and 458 (58.0%) were retrospective. Design-effect estimates were small for sensitivity (posterior mean logit difference 0.13; 95% CrI -0.07 to 0.32) and specificity (-0.05; 95% CrI -0.27 to 0.17), with both credible intervals including zero. Substantial heterogeneity was observed at the review and primary-study levels. In contextual secondary analyses, higher prevalence was associated with higher sensitivity and lower specificity, while larger sample size was associated with higher specificity. All remaining coefficients had credible intervals including zero.

In this sample of radiology DTA studies, we found no clear evidence that prospective versus retrospective study design was associated with reported sensitivity or specificity after accounting for clustering within systematic reviews. These findings support assessing specific sources of bias, study conduct and imaging context rather than relying on prospective and retrospective classification alone. Whether design-related effects vary according to diagnostic-test characteristics, including standardisation and operator dependence, requires direct evaluation across other diagnostic fields.