Presentation (CTI)

OBJECTIVES: Meta-analysis of diagnostic test accuracy requires specialized methods to jointly synthesize sensitivity and specificity while accounting for theircorrelation across studies. Bayesian approaches offer a flexible framework for combining evidence and quantifying uncertainty. This study aimed to conduct aBayesian bivariate meta-analysis of diagnostic accuracy using study-level sensitivity and specificity estimates derived under a common diagnostic threshold.
METHODS: Hypothetical diagnostic accuracy data were generated for ten studies, including counts of true positives, false positives, false negatives, and truenegatives. A Bayesian bivariate random-effects model was implemented in WinBUGS via R (R2WinBUGS package; version 4.5.0), following NICE TechnicalSupport Document 25 guidance. Between-study heterogeneity was modeled using vague priors (mean 0, variance 2). Posterior inference was performed usingMarkov Chain Monte Carlo simulation with 2,000 burn-in iterations and 80,000 sampling iterations across three chains.
RESULTS: The pooled posterior median sensitivity was 0.95 (95% credible interval [CrI]: 0.91-0.97), while the pooled specificity was 0.67 (95% CrI: 0.48-0.82). Theestimated between-study correlation between sensitivity and false-positive fraction was 0.41 (95% CrI: −0.40 to 0.89), indicating moderate correlation andsupporting joint modelling of diagnostic accuracy measures.
CONCLUSIONS: Bayesian bivariate random-effects models provide a robust framework for meta-analysis of diagnostic accuracy by jointly estimating sensitivity and specificity and appropriately capturing between-study correlation and uncertainty. This approach enhances interpretability and supports evidence synthesis for diagnostic decision making.