Systematic review of model-based cervical screening evaluations

Most studies included a cost-effectiveness analysis (129/153) and investigated the introduction of new screening technologies (72/153), with fewer focusing exclusively in alternative strategies using already-adopted technologies (47/153), and even fewer on the introduction of screening programmes where non-existent (34/153). Evaluations of the introduction of a screening technology were generally favourable to its adoption, with LBC recommended over conventional cytology (18/27), HPV DNA recommended over cytology for primary screening (15/17), rapid HPV DNA (3/3) or self-sampling (2/4) recommended for primary screening, and HPV DNA (9/10) or genotyping (1/1) recommended for triage of equivocal results.

Overall, our findings are in line with those of previous reviews of cost-effectiveness analyses [5,8-11] and post-vaccination analyses in the context of developed countries with existing screening programs [40], which mostly recommend the introduction of HPV DNA primary screening in high-resource settings and the revision of screening policies towards the introduction of HPV DNA primary testing.

As Nahvijou and colleagues also found, [10] there is a discrepancy between guidelines and model-based evaluations regarding more recent technologies. Generally, current HIC screening guidelines ( Summary of cervical screening guidelines provided in Additional file 8) are aligned with the overall findings of evaluations of cytology-based strategies; however, most concluded lacking sufficient evidence on the effectiveness of HPV DNA testing for primary screening to support its implementation, [45] with only a few countries, such as Australia, the Netherlands and the USA, recommending it at the moment.

Most of the global cervical cancer burden lies in low- and middle-income countries without organised screening programmes [46] However, as noted in previous reviews, [5,11] only a small proportion of studies in our review (34/153) addressed these settings, with the vast majority (33/34) supporting the existence of a screening programme. Indeed, over half the studies (80/153) were set in just 4 HIC – the USA, the UK, the Netherlands, or Canada. More evaluations focused on the regions with the greatest cervical cancer burden may have greater influence in driving adoption of screening technologies where they are most needed.

Currently several molecular biomarkers are being investigated for their potential to be integrated alongside cytology and HPV DNA testing in screening algorithms. However, no model-based study was found in this review on these emerging screening technologies. Only a few studies analysed more recent technologies as rapid HPV DNA testing, self-sampled HPV DNA testing, or HPV 16/18 DNA genotyping. No study on rapid HPV DNA testing was found in a low-income setting either.

Some molecular-based tests are thought to have the potential to improve cytology’s accuracy and reproducibility (e.g. p16 immunostaining), while other are thought to be promising alternatives to cytology (e.g. HPV DNA testing, HPV mRNA testing, p16/ki-67 dual immunostaining, or methylation markers) as they can be subject to automated quantification [47]. The clinical utility of HPV DNA testing has been shown, [2] and it has recently been introduced in primary screening in a few HIC, e.g. the Netherlands and Ontario [48]. These recent developments in screening technologies may suggest a transition to a fully molecular-based screening approach. However, the population-level effectiveness and cost-effectiveness behind many of the molecular technologies is still unexplored. For most biomarkers there is currently only cross-sectional evidence of their potential accuracy [3]. HPV mRNA testing for instance has been recently approved by the U.S. Food and Drug Administration for screening women over 30 years in combination with cytology, despite evidence from longitudinal trials of its improved accuracy in the detection of CIN2+ lesions who do not regress be not yet available [49]. Mathematical models are a key tool to allow results from trials and observational studies of these technologies to be extrapolated to explore their long-term impact in population-based screening programmes.

Another aspect of research that can be explored via mathematical modelling is the interaction between vaccination and screening. Vaccinating adolescent girls has been found likely to be cost-effective even in settings with existing screening programmes [40,50]. However, vaccination is expected to decrease the incidence of cervical abnormalities and eventually cancer [51]. Hence the positive predictive value of cytology will decrease, as will the effectiveness of most screening modalities. [43,52] In order to assist in population level policy making, future analyses in settings with vaccination will need to account for its impact on existing and prospect screening programmes. This is particularly true if a 9-valent HPV vaccine is successful in trials, as it is projected to ultimately prevent 90% of invasive cervical cancers [53].

Also, most models of screening in post-vaccination settings relied on a static infection structure. This may be suitable for comparing alternate screening strategies in a setting in which disease prevalence is constant, but would not capture the long-term changes in HPV prevalence, in settings with successful national HPV vaccination programmes [54] such as the UK, Australia and Portugal. Dynamic transmission models are particularly important now that a 9-valent HPV vaccine has shown high immunogenicity and efficacy in clinical trials [55]. This will have further implications on cervical screening since vaccinated girls will have a very low risk of infection with an oncogenic HPV type and hence risk of cervical cancer. The few dynamic models compared alternative cytology-based strategies [56,57] or strategies with rapid HPV DNA testing versus vaccination only or alongside vaccination [26]. Their overall results were consistent with those of static models in that screening strategies alongside vaccination maximise health outcomes. However, it can take many years for the direct and indirect impact of vaccination to be observed in surveillance data, so dynamic models will be increasingly important to explore changes to screening as the first vaccinated cohorts enter the age of screening eligibility.

Model calibration to observed setting-specific data has become more common; however it is still not routinely used. As most natural history parameters governing the progress of cervical abnormalities are very difficult to measure directly, model calibration enables their estimation based on observable outcomes such as abnormal screening results. This is generally a more reliable approach than making assumptions on parameters based on limited studies, often in unrepresentative populations [41,58]. Even the studies reporting having calibrated these parameters to outcome data often gave few details about the goodness-of-fit measure used and very rarely provided details on other aspects of calibration, such as the selection of calibration targets, parameter search strategies, and convergence criteria used. Detailed reporting of the calibration process should be common practice for reproducibility purposes [59]. Also, there should be an indication of uncertainty in the parameter estimates used and how it is incorporated to judge the sensitivity of model predictions to the data sources used.

This review is subjected to limitations. We focused on models used to assess the impact of alternative screening strategies, and excluded model-based studies assessing vaccination strategies, including those modelling screening strategies alongside vaccination that did not compare different screening strategies. Because of the volume and diversity of the relevant modelling literature, we did not critically appraise the quality of individual studies, but instead focused on providing an overview of the main approaches and conclusions of the models. Further work is needed to critically review modelling literature that addresses specific questions (such as the choice between cytological and DNA-based screening methods) in more detail. The main strength of our work lies in providing a broad overview of the vast literature over a long time period, and in identifying key conclusions that are common across models as well as gaps in the methodology and scope of current models.