To quantify uncertainty of and correlation between cervical cancer screening test sensitivity and specificity, for purposes of probabilistic sensitivity analysis.
Method:
We performed an abbreviated meta-analysis to determine the sensitivity and specificity of three cervical cancer screening tests in resource-limited settings: visual inspection with acetic acid (VIA), standard human papillomavirus (HPV) testing (Hybrid Capture 2, HC2) and rapid HPV testing (careHPV). Studies were identified from recent meta-analyses published in 2011-2012 (n=17), and from systematic PubMed searches (n=8). Selection criteria included: 1) all research subjects were healthy women with no gynecological symptoms; 2) reference standard was colposcopy or colposcopy and histology; and 3) disease threshold was cervical intraepithelial neoplasia grade II or higher (CIN2+) (n=20 studies, 2 with multiple sites). We then used count data (true positives, false positives, true negatives, false negatives) from included studies to inform a bivariate random effects model for each screening modality to generate logistic-normal and transformed distributions around sensitivity and specificity and conditional regression of logit specificity on sensitivity. We explored the impact of restricting to studies with quality control/independent review of histology.
Result:
Count data were derived for VIA (n=25 sites), HC2 (provider sampling, n=9 sites), and careHPV (provider and self-sampling, n=3 sites). Bivariate random effects models estimated sensitivity at 72% (95% CI: 65-78%) for VIA, 86% (95% CI: 72-93%) for HC2, and 79% (95% CI: 69-87%) for careHPV. Estimates for specificity were 87% (95% CI: 81-90%) for VIA, 90% (95% CI: 71-93%) for HC2, and 86% (95% CI: 83-89%) for careHPV. When we restricted to studies with quality control/independent review of histology, test sensitivity diminished for VIA (54%; 95% CI: 43-64%) and improved for HPV-based testing (91%; 95% CI: 80-96%). Using the covariance between logit sensitivity and specificity and the between-study variance in logit sensitivity from the bivariate models, we generated linear regressions of logit specificity conditional on sensitivity.
Conclusion:
Bivariate random effects models can be used to generate logistic-normal distributions for sensitivity and specificity. Standard output from the bivariate models can be used to inform linear regression of logit specificity conditional on sensitivity, facilitating probabilistic sensitivity analysis in microsimulation models that accounts for the correlation between parameters. Evaluation of cervical cancer screening test performance must incorporate quality control measures in histologic review.
See more of: The 35th Annual Meeting of the Society for Medical Decision Making