Reducing ovarian cancer mortality through screening: an impossible dream?

Although ovarian cancer is “only” the 10th most common cancer in women, it is the fifth-leading cause of cancer death (1). Sixty-five percent of ovarian cancers are diagnosed after the disease has spread within the peritoneal cavity (stage III) or distant organs (stage IV) (2). Because 5-year survival for localized disease is over 90% compared with 30% for distant disease (2), efforts at developing effective early-detection strategies for reducing ovarian cancer mortality (3-5) have been ongoing since the 1980s. Unfortunately, large randomized trials have repeatedly failed to show a significant reduction in mortality in screened patients (6,7). In this issue of the Journal, Ishizawa and colleagues (8), using data from the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial (6) and UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS) (7), use an innovative approach to gain insight into factors contributing to these disappointing results.

Using an approach initially validated using lung cancer screening data (9), the authors modeled cancer progression as a 3-state (no clinical disease, preclinical detectable disease, and clinical disease) continuous-time Markov chain. In the absence of screening, there are transition rates between no disease and preclinical disease (λ₁) and between preclinical disease and symptomatic disease leading to diagnosis (λ₂); in a screening setting, the preclinical state can be detected based on the sensitivity of the tests. Varying the values of λ_1, λ_2, and sensitivity results in different estimates of the number of screen-detected cases within a specified interval. By calibrating the value of the 3 parameters so that model estimates approximate reported cases from the trials, it is possible to derive an estimate for the mean sojourn time—the “window of opportunity” for effective early detection. The authors further extended this approach to separately estimate sojourn time between early-stage and advanced-stage cancers. Individual-level data from PLCO and published summary data from UKCTOCS were used; in addition, stage-specific and histology-specific incidence and survival data from Surveillance, Epidemiology, and End Results were used to estimate histology-specific survival hazard ratios relative to high-grade serous carcinoma, the most common subtype with the poorest survival, and subsequently estimate histology-specific sojourn times.

Using the PLCO data, estimated overall mean sojourn time was 2.1 years (95% confidence interval [CI] = 1.9  to 2.4 years), with a sensitivity over the 6 years of screening of 65.7% (95% CI = 60.2% to 71.2%). Sojourn time estimates did not differ when analyzed on the basis of screening modality (cancer antigen 125 plus ultrasound vs cancer antigen 125 alone), although sensitivity with multimodal screening (80.8%, 95% CI = 74.3% to 87.4%) was statistically significantly higher than cancer antigen 125 alone (36.2%, 95% CI = 29.6% to 43.2%). Estimated sensitivity for early-stage cancers (39.1%, 95% CI = 34.9% to 43.3%) was statistically signficantly lower than for advanced cancers (82.9%, 95% CI = 78.0% to 87.8%). Critically, estimated time in the early-stage preclinical state was 1.1 years (95% CI = 1.0 to 1.1 years).

Results using the published UKCTOCS data were similar: estimated mean sojourn times in the multimodal screening arm of 2.0 years (95% CI = 1.8 to 2.1 years) and 2.4 years (95% CI = 2.2 to 2.5 years) in the transvaginal ultrasound arm. Estimated sojourn times for aggressive histologic subtypes (including high-grade serous carcinoma) were significantly shorter (range = 0.9-1.8 years) than for histologic types with better prognosis (range = 2.9-6.6 years).

The population-level time course of cancer progression from an initial mutation until diagnosis with symptoms is an inherently “unknowable unknown,” but estimates of the unobserved rate of progression are necessary for developing models to compare the potential impact of different cancer-control strategies. A variety of approaches exist for inferring these rates based on population-level data; Ishizawa and colleagues (8) have demonstrated an innovative approach to deriving estimates from empiric data, for example. Although the authors note that the lack of individual-level data from UKCTOCS is a limitation, the fact that estimates of mean sojourn time on the basis of the individual-level data from PLCO are similar to those on the basis of UKCTOCS summary data provides additional confidence in the validity of their approach.

Neither trial demonstrated a significant reduction in ovarian cancer mortality among screened women (despite a stage shift in UKCTOCS). The analysis by Ishizawa and colleagues (8) provides some explanation for these findings. First, the mean sojourn time for ovarian cancer is short; as the authors point out, it is substantially shorter than estimates for other screen-detectable cancers, such as prostate, lung, breast, and colorectal cancers. Second, the sojourn time for early-stage preclinical disease was only 1 year, and sensitivity of all screening modalities was lower for early-stage disease. Third, the sojourn time for high-grade serous carcinoma, the most frequent cause of ovarian cancer death, was substantially shorter than for histologic subtypes with better prognoses. As the authors point out, these findings are consistent with our current understanding of the biology of high-grade serous carcinoma. Most of these cancers arise in the fallopian tube epithelium—there are no physical barriers to cancer cell migration to the surface of the ovary, and from the surface of the ovary to the surfaces of other organs within the peritoneum. Given that the biology of high-grade serous carcinoma is so different from that of cervical cancer, the paradigm for successful cancer mortality reduction through screening, it is not surprising that we do not yet have an “ovarian Pap test.”

The authors discuss some of the implications of these findings for future efforts to develop effective ovarian cancer screening tests. A short sojourn time implies that screening intervals may need to be less than 1 year, and new tests should ideally have a higher sensitivity for early-stage disease than current modalities. Because a higher sensitivity would almost necessarily be associated with lower specificity, however, the combination of more frequent screening and lower specificity would inevitably lead to a high number of false-positive results—a particular problem for ovarian cancer, where definitive diagnosis requires surgery. Further mathematical modeling, informed by analyses such as this one, can be used to explore combinations of sensitivity, specificity, and screening frequency that could result in acceptable trade-offs of benefits (especially mortality reduction) and harms (such as false positives and unnecessary surgeries) for screening, either in the general population or in specific high-risk populations. As new screening tests and strategies are developed, those whose preliminary characteristics resulted in acceptable benefit/harm trade-offs could be considered for clinical trials, with mortality as the endpoint. As the UKCTOCS investigators point out, to date, findings of benefit for surrogate outcomes in ovarian cancer screening have not translated into mortality reduction (10).

Even if a strategy with characteristics that could potentially be acceptable were identified, however, another issue must be considered, both in study design and in implementation: Ovarian cancer incidence and incidence-based mortality have been declining significantly since the mid-1990s, and these declines are greatest for high-grade serous carcinoma (11). This decline in incidence may partially be attributable to several factors, including the increased use of contraceptive methods that are associated with lower risk and opportunistic salpingectomy, but age-specific data suggest a decline even in women born in the 1920s, who would not have benefited from oral contraceptives. Although the declining incidence in high-grade serous carcinoma is indisputably good news, it creates a potential problem for study design: Not accounting for potential cohort trends in cancer incidence could lower a study’s power. For example, Figure 1 shows the expected 5-year cancer incidence in women aged 55, 60, 65, and 70 years based on Surveillance, Epidemiology, and End Results cross-sectional data in 1993-1995, when the PLCO started enrollment, compared with observed estimates for each age cohort over 20 years of follow-up. Especially for the younger age groups, observed cancer incidence was lower than predicted based on cross-sectional data before beginning enrollment. If these trends continue, there are major feasibility issues for screening trials. Declining incidence will also increase the harm/benefit ratio (and decrease the cost-effectiveness) of any screening strategy.

Figure 1.

Difference between expected 5-year probability of first diagnosis of ovarian cancer, by age cohort, based on Surveillance, Epidemiology, and End Results cross-sectional estimates, 1993-1995, compared with observed age-specific incidence in subsequent years. Expected probabilities for the cumulative 5-year incidence starting at ages 55, 60, 65, and 70 years are based on cross-sectional age-specific incidence in 1993-1995, derived from DevCan (referent) (solid line). “Observed” lines plot 5-year cumulative probabilities for each age group for the years 2000-2002, 2009-2011, and 2015-2017. For example, “observed” incidence for women aged 55 years in 1993-1995 would be those of 60-year-old women in 1999-2001, 65-year-old women in 2003-2005, and 70-year-old women in 2009-2011. Unanticipated cohort effects could affect the precision of estimates of screening test effectiveness.

Given the inherent challenges to ovarian cancer screening and the continuing decline in incidence and mortality, the relative value of additional efforts to develop screening modalities compared with more optimal utilization of primary prevention strategies must be considered. In addition, better understanding of the causes of the ongoing decline in incidence at the population level is needed. First, to the extent that the decline is attributable to interventions such as contraception and salpingectomy, estimating the potential impact of optimizing access is crucial. Second, because there may be age-period-cohort trends in other exposures (both risk reducing and risk increasing) that have contributed to the declining incidence, estimates of those effects are needed to help with predictions of future trends.

Randomized trials with mortality as the primary endpoint will always be needed to prove screening effectiveness, but mathematical models, such as those developed and used by the National Cancer Institute–sponsored Cancer Intervention and Surveillance Modeling Network, are also necessary for estimating the population-level impact of implementing screening. The work by Ishizawa and colleagues (8) illustrates how modeling can provide insight into trial results. Moving forward, more modeling work will be critical for helping us understand the feasibility of future screening strategies for ovarian cancer, from the perspective of pragmatic study design and population-level implementation, and inform decisions about the optimal balance of resources devoted to primary prevention, screening, and improved treatment to continue and perhaps expedite the ongoing decline in mortality.

Data availability

The data used to generate the figure in this editorial are available from Surveillance, Epidemiology, and End Results (seer.cancer.gov) and were accessed using DevCan, version 6.7.5, software (DevCan: Probability of Developing or Dying of Cancer Software; Statistical Research and Applications Branch, National Cancer Institute, 2007 [http://srab.cancer.gov/devcan/]).

Author contributions

Evan R. Myers, MD, MPH (Conceptualization; Data curation; Software; Validation; Visualization; Writing—original draft; Writing—review & editing).

Funding

No funding was used for this editorial.

Conflicts of interest

Dr Myers has no disclosures.