Presentation (CTI)

OBJECTIVES: Meta-analysis of diagnostic test accuracy differs from usual meta-analysis and is useful for increasing validity by combining data from multiple studies. This study aims to perform a meta-analysis of a single estimate of sensitivity and specificity per study using a Bayesian approach by considering the same threshold for all the studies included.
METHODS: The hypothetical data were generated with ten studies. The dataset included true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) counts per study. The analysis was implemented in WinBUGS through R (R2WinBUGS package), version 4.5.0. The bivariate random-effect model was used to estimate sensitivity and specificity across studies as per NICE TSD 25. The vague prior (0,2) was used for between-study variation. The posterior distributions were obtained using the Markov Chain Monte Carlo (MCMC) algorithm with 2000 burn-ins and 80,000 iterations of three chains.
RESULTS: The pooled estimate, that is, posterior median and 95% credible interval (CrI) of sensitivity was 0.95 (0.91, 0.97) whereas the pooled estimate of specificity was 0.67 (0.48, 0.82). Additionally, the between-study correlation estimated was 0.41 (-0.40, 0.89), suggesting moderate correlation between sensitivity and FP fraction.
CONCLUSIONS: The bivariate random-effect model helps to depict joint uncertainty in the estimated average sensitivity and specificity, which makes it a flexible and robust method for diagnostic meta-analysis while accounting for correlation.