Deciding about Screening

As shown in an accompanying article, the benefits of any broadly based screening program for ovarian cancer are limited more by the low prevalence of disease than by the potential benefits of treating early disease or the performance of currently available tests. Screening for such diseases is best restricted to special populations at markedly increased risk. In low-prevalence populations, false-positive results and the costs and risks of the workups they induce are magnified. In determining the marginal cost-effectiveness ratio of such programs, failure to consider even apparently minor factors can markedly affect the conclusions. In the case of screening for ovarian cancer, factoring in the discomfort and potential inconvenience of venipuncture and transvaginal sonography can eliminate the small benefit of the program. Taking the time and energy required to organize relevant data in constructing a decision analytic model can provide important insights into the usefulness of such screening programs.

Screening for preclinical disease can lead to early intervention and thereby improve health outcomes and possibly save resources. Because screening is rarely free of cost and often carries risk, either from the test itself or from the subsequent workup it can induce, we must decide when and for whom screening is appropriate. A formal decision analysis can help structure the problem, organize data, elucidate tradeoffs, and estimate benefits and costs.

In this issue of Annals, Schapira, Matchar, and Young [1] analyze the effect of screening 40-year-old women for early ovarian cancer. They conclude that screening would, on average, prolong life expectancy of each woman screened by only 14 hours and thus would not be an effective health policy. Although these investigators focused on the characteristics of screening tests and the risks of laparotomy, they correctly identify the crux of the problemthe low prevalence of ovarian cancer. Even assuming perfect test performance and ignoring the risks and morbidities of laparotomy, their model suggests that the benefit of screening remains trivialonly 1.4 daysa toss-up [2]. Recall that modifying lifestyle and risk factors for coronary heart disease is estimated to prolong the life expectancy of a 35-year-old by up to 1 year [3]; annual Papanicolaou smears for cervical cancer could add 3 months to a woman's life [4]; and regular screening for breast cancer may increase the life expectancy of 50-year-old women by 2 months [5]. Even if we required an average gain of only 30 days before endorsing a program, screening for ovarian cancer would be recommended only in populations where disease prevalence exceeds 1400/100 000, 50 times higher than the prevalence in unselected 40-year-old women.

The conclusion that population-based screening for ovarian cancer would provide little benefit is all the more surprising because early intervention in ovarian cancer can markedly improve prognosis. A screening program using CA 125 and transvaginal sonography will extend the life expectancy of those patients who have preclinical disease by 11.6 years; a perfect screening test would extend their life expectancy by 26 years. However, the benefits for these women must be balanced against the adverse consequences of screening for the remaining women. Because these tests are imperfect, the number of patients submitted to workup because of false-positive results would far exceed the number of true positives. Screening 40-year-old women for ovarian cancer with CA 125 as a single test (specificity 97.6%) produces 340 false-positive results for each additional early cancer detected. Combining blood CA 125 levels with transvaginal sonography increases the specificity to 99.95% and produces only eight false-positive results for each additional early cancer detected. Clearly, maintaining very high specificity must be a critical goal in any program that screens low-prevalence populations [6]. From another perspective, knowing that a screening program could increase the life expectancy of a patient with occult ovarian cancer by 11.6 years, we might ask how much an unaffected woman would lose. On average, her life expectancy would be shortened by 25 minutes, reflecting the rare complication of laparotomy when results of both CA 125 level and the transvaginal sonogram are falsely positive.

Other authors have recently concluded that screening for ovarian cancer is not appropriate unless groups at particularly high risk can be identified, because the performance of current tests is relatively poor [7] and the benefit of treatment is limited [8]. Why did Annals devote space to yet another paper that reaches the same conclusion? Despite assuming better test performance and perhaps best-case efficacy for early treatment, Schapira and coworkers conclude that the benefits would be trivial and would remain so even if better screening tests and better treatments are developed because of the low prevalence of preclinical ovarian cancer.

The same logic would apply to screening for many low-prevalence diseases. In adults, no therapy, no matter how efficacious, can improve life expectancy beyond that of a healthy person of the same age. Even when considering diseases that are uniformly fatal in the short-term, the upper bounds of the average benefit of a perfect therapy varies from 5 years (at age 88) to 50 years (at age 20). Using a perfect test and a riskless workup to screen for an immediately fatal disease with a prevalence of 1/1000 could provide a maximum benefit to each person screened of only 2 to 18 days, even if detecting preclinical disease increased survival to full life expectancy. Using imperfect tests and workups with risk further limits the benefit of screening.

Toss-ups present a variety of problems for both the clinician and the policy analyst. No analysis, formal or informal, can consider all factors. We explicitly or implicitly edit out minor issues. but when the expected difference between strategies is very small (14 hours in this case), such minor omissions can change our conclusions. For example, the authors of the current analysis ignore the morbidity of the venipuncture, the potential discomfort and embarrassment of the transvaginal sonogram, and the anxiety of waiting for test results. They also ignore the fact that most patients would prefer to defer risks, such as the potential complications of laparotomy, as long as possible. Some of these minor adjustments apply only to the 2.4% of the unaffected population who have a falsely elevated CA 125 level, whereas others apply to everyone screened. Explicit consideration of these factors will markedly diminish the expected 14-hour gain in survival and may wipe it out altogether.

Many health policy analyses look at the health benefit of a proposed program first. If that benefit is negative or very small, analysts sometimes do not even consider the economics. Such partial analyses can, however, be misleading because economic factors can be the driving force, particularly if early detection markedly decreases the cost of treating patients with disease (as is the case for many neonatal screening programs). Although Schapira and coworkers have not undertaken a formal cost-effectiveness analysis, an informal back of the envelope analysis may be helpful. The actual variable costs for a CA 125 serologic test and a transvaginal sonogram are approximately $20 and $90, respectively (although charges are higher). If sonograms are performed only in patients with positive serologic results, the direct cost of screening one patient will be approximately $22. If only patients with positive results of both tests undergo laparotomy, the total of direct and induced costs will be approximately $25. Even if shifting patients from late to early ovarian cancer provides no savings in the cost of care, spending $25 to save 14 life-hours, when applied to a population, is equivalent to spending $15 000 to save 1 year of life. If we discount future benefits at 5%/year, the cost effectiveness ratio rises to $30 000 per year of life saved. Comparable ratios are $30 000 to $100 000 per year of life saved for annual breast cancer screening [5] and $15 000 to $30 000 per year of life saved for cervical cancer screening every 4 years [4]. If we could be confident that screening for ovarian cancer would improve life expectancy at a cost of $15 000 to $50 000 per life year saved, it would not be an unreasonable policy. but suggesting a program, no matter how apparently cost-effective, with an expected gain in life expectancy of only 14 hours per patient screened is troubling because no model can ever be complete. If small omissions wipe out the gain, then the cost-effectiveness ratio would become infinitely high.

Was it worth performing a decision analysis on a question so well studied, or at least so broadly commented on in recent literature? The two main products of a formal analysis are recommendations and insights. After reading Schapira and coworkers' analysis, our conclusions about ovarian cancer screening should remain unchangedsuch programs cannot be recommended as public policy. but we can now appreciate that the driving force is the low prevalence of disease, not the particular characteristics of the CA 125 serologic analysis or transvaginal sonography. Screening for ovarian cancer should be limited to women at markedly increased risk. The dynamics of screening for ovarian cancer and other low-prevalence diseases are now clearer because Schapira and coworkers have taken the time and energy required to organize the data, estimate the effects of early detection, and provide a model that can be easily explored. The benefits of such careful analyses are clear and well worth the cost.

• Copyright 2004 by the American College of Physicians