|Home | About | Journals | Submit | Contact Us | Français|
Dr. Gadi: Division of Medical Oncology, Department of Medicine, University of Washington, 825 Eastlake Avenue East, Seattle, WA 98109.
Mr Markowitz and Dr Etzioni: Division of Public Health Sciences, Fred Hutchinson Cancer Research Center, 1100 Fairview Avenue North, M2-B230, PO Box 19024, Seattle, WA 98109.
Mammography trials, which are the primary sources of evidence for screening benefit, were conducted decades ago. Whether advances in systemic therapies have rendered previously observed benefits of screening less significant is unknown.
To compare the outcomes of breast cancer screening trials had they been conducted using contemporary systemic treatments with outcomes of trials conducted with previously used treatments.
Computer simulation model of 3 virtual screening trials with similar reductions in advanced-stage cancer cases but reflecting treatment patterns in 1975 (prechemotherapy era), 1999, or 2015 (treatment according to receptor status).
Meta-analyses of screening and treatment trials; study of dissemination of primary systemic treatments; SEER (Surveillance, Epidemiology, and End Results) registry.
U.S. women aged 50 to 74 years.
10 and 25 years.
Mammography, chemotherapy, tamoxifen, aromatase inhibitors, and trastuzumab.
Breast cancer mortality rate ratio (MRR) and absolute risk reduction (ARR) obtained by the difference in cumulative breast cancer mortality between control and screening groups.
At 10 years, screening in a 1975 trial yielded an MRR of 90% and an ARR of 5 deaths per 10 000 women. A 2015 screening trial yielded a 10-year MRR of 90% and an ARR of 3 deaths per 10 000 women.
Greater reductions in advanced-stage disease yielded a greater screening effect, but MRRs remained similar across trials. However, ARRs were consistently lower under contemporary treatments. When contemporary treatments were available only for early-stage cases, the MRR was 88%.
Disease models simplify reality and cannot capture all breast cancer subtypes.
Advances in systemic therapies for breast cancer have not substantively reduced the relative benefits of screening but have likely reduced the absolute benefits because of their positive effect on breast cancer survival.
University of Washington and National Cancer Institute.
Academic and public dialogue is polarized about whether the benefits of population screening programs outweigh their harms (1, 2). Policy panels are mandated to prioritize randomized clinical trials when developing guidelines, as indicated by the U.S. Preventive Services Task Force’s designation of randomized clinical trials as “level I evidence” in its hierarchy of research designs (3). When results from randomized clinical trials conflict, such as in breast and prostate cancer screening trials (4, 5), creating guidelines becomes controversial.
Recent studies (6–8) have highlighted a particularly intractable issue about screening benefit: The benefit of early detection of cancer is inextricably linked to the effectiveness of treatment in reducing cancer mortality. Recognizing this link between treatment and screening, some researchers (9–11) have hypothesized that contemporary treatments may have decreased the effect of screening mammography. Many of the mammography trials, which were initiated in the 1960s and 1970s (12), predated the advent of adjuvant multiagent chemotherapy and endocrine treatment that entered general use in the 1980s and 1990s and are now standards of care (13). The availability of these therapies is cited as a potential reason that the Canadian National Breast Screening Study, which was conducted in the 1980s, did not find a significant effect of screening on mortality (5, 14, 15), whereas meta-analyses of all trials estimated a 15% to 20% relative reduction in mortality (9, 10, 16). The recent Cochrane review (9) and Swiss Medical Board guideline (11) similarly cite advances in treatment as support for the hypothesis that the true mortality reduction due to mammography is lower than that suggested by meta-analyses.
Jüni and Zwahlen (10) recently proposed that “The only way to know [the effect of mammography screening] for certain is to initiate a new trial in the era of contemporary screening technologies and breast cancer therapies.” Unfortunately, the 20 years of follow-up needed to reliably observe mortality reductions attributable to screening (17) presents a problem of timing that no trial can circumvent. Treatments administered in a new trial may be outdated in 20 years (18). Moreover, predicting how treatment advances will affect screening is difficult. When treatments improve for tumors detected early in the process of disease progression, the benefit of screening will increase. Conversely, better treatments for more advanced disease will decrease the value of screening.
When empirical studies fall short of providing the evidence needed for policy development, modeling can be valuable. Models have been used to decouple the joint contributions of screening and treatment to trends in breast cancer mortality in the population setting (6, 19), but the effect of treatment advances on screening trial results has not been explicitly investigated. We developed a modeling framework to quantify how screening trial results change when treatments improve. We applied this framework to breast cancer to address the hypothesis that advances in treatment have altered the effect of screening on breast cancer mortality.
We used a microsimulation approach to implement a state-transition model, where the states were healthy, screen-detected or clinically detected breast cancer, and death (Figure). Breast cancer may be detected at either an early or an advanced stage; the time from diagnosis to death depends on the stage at detection and the treatments received. The effect of screening is to shift some cases that would have been detected later to an earlier stage at diagnosis, resulting in a decrease in advanced-stage incidence. The effect of treatment is to change the disease-specific survival given the stage at diagnosis.
We used the model to project mortality results corresponding to 3 hypothetical screening trials, which reflected the same effect of screening on disease stage but different treatment distributions. The resulting virtual trials were a 1975 trial with treatments available in the U.S. population at that time, a 1999 trial with treatment frequencies in the U.S. population from that year, and a 2015 trial that reflected optimal use of contemporary treatments for the specific types of breast tumors modeled. For each virtual trial, we simulated a population of 200 000 women aged 50 years who were randomly assigned to screening and control groups. Our objective was to compare the projected effect of screening on disease-specific mortality across the 3 virtual trials.
In the remainder of this section, we describe the data and methods used to generate the dates of diagnosis and death for the control and screening groups in the virtual trials. Additional details are included in Appendix Table 1. The model was built and deployed using R, version 3.2.0 (R Foundation for Statistical Computing) (20). This study was exempt from institutional review board approval because it used only published or publicly available data.
The incidence of disease in the absence of screening (clinical incidence) was an anchor point for the model. We focused on clinically incident cases because they are the only cases that lead to disease-specific mortality, which was our primary outcome of interest. For these cases, we generated the following 3 variables.
This was based on age-specific rates of clinical diagnosis (within 5-year age categories) from the SEER (Surveillance, Epidemiology, and End Results) database for 1975 to 1979 (21), which represents the interval directly before the adoption of mammography in the United States. The empirical age-specific incidence was used and was kept constant across the virtual trials.
This was assigned as “early” (in situ or localized) or “advanced” (regional or distant). We used historical SEER breast cancer stage distributions from 1975 to 1979.
We included the breast cancer tumor receptor category because different treatments are used depending on estrogen receptor (ER) and human epidermal growth factor receptor 2 (HER2) status (22). Separately for early and advanced stages, we used SEER data on patients aged 50 to 75 years with breast cancer diagnosed in 2010 and known receptor status to simulate positive or negative/borderline ER and HER2 status.
We retained the number of patients with incident disease generated in the absence of screening and their simulated age at clinical diagnosis but reduced advanced-stage incidence by reclassifying a fraction of advanced-stage patients as early-stage at diagnosis. We used a fraction of 15% for our main analysis; this reflected the median decrease in the cumulative incidence of advanced-stage disease reported across 8 mammography trials (23). In the model, this “stage shift” under screening improves disease-specific survival.
Disease-specific survival was generated from the date of clinical diagnosis and depended on the stage of diagnosis and the treatment received. For baseline survival in the absence of systemic treatments, we used an exponential distribution fit to disease-specific survival from patients in SEER aged 50 to 75 years who were diagnosed between 1975 and 1979. Modern systemic treatments were incorporated by applying hazard ratios (HRs) to the baseline survival to modify the time from clinical diagnosis to cancer death. The age at death was generated as the lower of the age at cancer death and the age at other-cause death based on U.S. life tables from 2000 (24). With screening, a fraction of advanced-stage patients were reclassified as early-stage (“stage-shifted”) because of early detection. For these patients, disease-specific survival from clinical diagnosis to cancer death was regenerated from the distribution of early-stage survival times. Although screening leads to early detection, the effect of the earlier diagnosis was modeled as an extension of the survival from the original date of clinical diagnosis in the absence of screening to avoid lead-time bias (25, 26). For these patients, the survival time in the presence of screening was explicitly correlated with their survival time in the absence of screening (both were generated using the same percentile with their respective distributions).
Treatment distributions were based on a study of adjuvant treatment dissemination (27), which provided frequencies of treatment with multiagent chemotherapy and tamoxifen for ER-positive tumors between the 1970s and 2000. Therefore, in the 1975 virtual trial, we used treatment patterns from 1975, when chemotherapy was made available for only a fraction of patients with advanced disease (Table 1). The 1999 virtual trial reflected the shift toward using ER status to target adjuvant tamoxifen and greater use of chemotherapy. The 2015 virtual trial approximated a contemporary scenario in which combination chemotherapy and aromatase inhibitor therapy were used for ER-positive patients and chemotherapy alone was used for ER-negative patients. In addition, HER2-positive patients with either ER status received trastuzumab. Within each virtual trial, both the control and screening groups received the same stage-specific distributions of treatment.
Treatment efficacies were based on published meta-analyses. For chemotherapy and tamoxifen, we used results from the Early Breast Cancer Trialists’ Collaborative Group (28, 29) (see Appendix Table 1 for more details). For trastuzumab, Cochrane reviews (30, 31) reported HRs for overall survival; we used these HRs as proxies for the effect on disease-specific survival. For aromatase inhibitor therapy, we multiplied the tamoxifen HR by 0.89 based on the reduction in breast cancer mortality after recurrence observed in 2 major trials comparing aromatase inhibitors and tamoxifen (32). Finally, we assumed multiplicative independent effects on the HR scale for combination therapies.
The main outcome of interest was the cumulative number of breast cancer deaths in the control and screening groups of each virtual trial. We assessed the effect of screening by using the mortality rate ratio (MRR) (the ratio of cumulative mortality between the screening and control groups) and the absolute risk reduction (ARR) (the absolute amount by which screening reduced cumulative breast cancer mortality). We report these statistics over 10 and 25 years. In addition, we assessed the combined effect of screening and treatment across virtual trials by computing MRRs and ARRs for each group of each trial relative to the control group of the 1975 trial. Along with our “within-trial” results reflecting screening efficacy, we also present “across-trial” results reflecting the effect of screening and treatment changes.
To externally validate the model, we used the SEER program’s DevCan software, version 6.7.3 (33), which provides estimates of the chance of dying of breast cancer over a specified interval beginning at a specified age. The software does not provide results before 2000. We used the DevCan 25-year mortality estimates based on breast cancer mortality rates for 2000 to 2002 to validate our 1999 trial and the estimates based on mortality rates for 2009 to 2011 to validate the 2015 trial.
To investigate the sensitivity of our modeled outcomes to changes in key assumptions, we explored 4 alternatives to our main analysis. First, we assumed Weibull rather than exponential distributions for baseline survival by stage. The 2-parameter Weibull distribution can capture an increasing or decreasing risk for disease-specific death over time, as opposed to the exponential distribution, which assumes a constant mortality risk. Second, treatment efficacies per stage were varied independently, from an HR of 0.25 (very beneficial) to 1 (no benefit), to represent treatment advances that affect the 2 stages in different ways. We assumed that all detected cases were treated and varied the effect of treatment on stage-specific survival. Third, the decrease in advanced-stage incidence was increased from 15% to 50%, which could represent increased early-stage detection due to better technology or more frequent screening or biopsy. Finally, the breast cancer mortality rate among early-stage patients in the screened group was reduced by 30% to reflect improvement in survival corresponding to a within-stage shift. This would allow a benefit for patients detected at an early stage with screening even if they would have been detected at an early stage in the absence of screening. This benefit was not applied to patients shifted by screening from an advanced stage to an early stage.
This study was supported in part by a Comparative Effectiveness Research dissertation award from the University of Washington Pharmaceutical Outcomes Research and Policy Program and by the National Cancer Institute. The funding sources had no role in the design, conduct, and analysis of the study or in the decision to submit the manuscript for publication.
For 2000 to 2002, the DevCan (33) estimate of the cumulative 25-year breast cancer mortality rate for a population of women aged 50 years without breast cancer was 1.4%; our 1999 trial estimate for women aged 50 years having screening was 1.67%. For 2009 to 2011, the DevCan estimate for women aged 50 years without breast cancer was 1.2%; our 2015 trial estimate for women aged 50 years undergoing screening was 1.33% (Appendix Table 2).
Table 1 presents the modeled frequency of systemic treatment by disease stage and ER or HER2 status in the 1975, 1999, and 2015 virtual trials. Appendix Table 1 summarizes our remaining input parameters and their sources. We estimated that, in the absence of screening, 49.6% of patients would be detected at an early stage and the remainder would be detected at an advanced stage. Baseline annual breast cancer excess mortality rates for early- and advanced-stage patients were set to 2.0% and 10.7%, respectively, based on observed disease-specific survival for early- and advanced-stage patients diagnosed between 1975 and 1979.
Our main analysis assumed a 15% reduction in advanced-stage incidence associated with screening. Table 2 presents the projected cumulative mortality over 10 years for the 3 virtual trials, by trial group. The MRR for the 1975 trial was 0.90, reflecting a 10% reduction in the disease-specific mortality risk over this follow-up period. Within-trial MRRs for the 1999 and 2015 trials were almost identical (0.91 and 0.90, respectively), showing almost no effect of the improved mortality benefits of systemic therapies on screening efficacy. Corresponding absolute mortality reductions ranged from 5 (1975 trial) to 3 (2015 trial) deaths per 10 000 women. Thus, although changes in treatment across the 3 trial settings affected survival, they did not substantively alter the relative mortality benefit associated with screening but may have been associated with reductions in the absolute mortality benefit.
The survival benefits of treatment advances were readily apparent from mortality comparisons across trials. The 10-year MRR comparing the control groups of the 2015 and 1975 trials was 0.63, reflecting a 37% reduction in the risk for disease-specific death due to treatment advances during this time. The 10-year MRR comparing the screening group of the 2015 trial with the control group of the 1975 trial was 0.57, reflecting a 43% reduction in the risk for disease-specific death due to screening and treatment advances during this time.
Long-term (25-year) results were similar to the 10-year findings with respect to the relative effect of the modeled systemic therapies, but ARRs were substantially higher, as expected with the longer follow-up (Appendix Table 2).
Assuming Weibull distributions, as fitted to SEER survival data, yielded a reduced effect of screening on disease-specific mortality, with MRRs ranging from 0.93 to 0.95 across trials (Appendix Table 3). Nevertheless, MRRs comparing screening and control groups were similar across trials, indicating only a minor effect of treatment advances on screening efficacy.
When we allowed treatment advances to benefit the stages differently, we found a greater range of MRRs across trials. Across 16 virtual trials with varying efficacies of early- and advanced-stage treatment (4 levels of efficacy per stage), screening reduced mortality by 2% to 13% (Appendix Figure). The most modest MRRs occurred when advanced treatment was highly effective but early treatment was not, and the lowest MRRs occurred when advanced treatment was ineffective and early treatment was highly beneficial. This is intuitively reasonable because under a given stage shift, screening provides the most benefit when patients with advanced-stage disease fare poorly, but they are amenable to highly curative therapy when they are diagnosed early.
When the reduction in advanced-stage incidence with screening increased to 25%, the projected mortality reductions in each trial increased but still remained similar across trials. Within-trial MRRs over 10 years ranged from 0.83 to 0.85, and ARRs ranged from 5 to 7 deaths per 10 000 women (Appendix Table 4). With a 50% reduction in advanced-stage incidence with screening, within-trial MRRs ranged from 0.67 to 0.70 and ARRs ranged from 9 to 14 deaths per 10 000 women (Appendix Table 5).
Adding a within-stage survival benefit associated with screening with an associated HR of 0.70 (Appendix Table 6) improved the within-trial MRRs to 0.85. In this setting, we still projected almost identical within-trial MRRs for the 3 virtual trials, indicating the robustness of our main result to the screening benefit mechanism.
In this study, we developed a novel modeling framework to investigate whether screening benefit assessed in historical trials is likely to be preserved in a changing therapeutic landscape. Our modeling framework decouple the effects of screening on disease stage from the effects of primary systemic treatment. We obtained similar MRRs across the 3 virtual trials, indicating that systemic treatment changes in the United States do not substantively change the relative mortality effect of mammography screening. This result was robust to different assumptions for the extent to which screening affects advanced-stage incidence and suggests that speculation that treatment changes were responsible for the negative results of the Canadian trial and other mammography studies is probably incorrect. However, we did observe differences in the within-trial ARRs, particularly over the long term. The results of our 25-year analysis show that the absolute number of lives saved by screening tends to decrease when more efficacious treatments are available. This is a natural consequence of the decrease in absolute mortality that occurs with improved systemic therapies. A consequence of this effect is that in the presence of effective systemic therapies, the number needed to screen is likely to increase even as the MRR associated with screening remains unchanged.
Our models projected a reduction of about 10% in the risk for breast cancer death when we assumed that screening reduces advanced-stage disease at diagnosis by 15%, which was the median reduction reported across the major breast screening trials. This mortality reduction is consistent with the estimate from the recent Cochrane review (9), which was restricted to trials deemed to have adequate randomization (summary MRR, 0.90 [95% CI, 0.79 to 0.92]), and lies within the range provided by a U.S. modeling study (19). It is more modest than the summary reduction provided by the meta-analyses of the Independent U.K. Panel on Breast Cancer Screening (12) and the Canadian Task Force on Preventive Health Care (34), as well as the overall Cochrane review estimate (9), all of which reflect a 20% reduction in mortality associated with screening. Our more modest results may be a consequence of our assumption that the effect of screening is to reduce the incidence of advanced-stage tumors at diagnosis. When we modeled a within–early-stage benefit on top of the benefit from the reduction in advanced-stage disease, we projected a result that was closer to that obtained in trial meta-analyses.
The stage-shift assumption or mechanism underlying screening benefit is well-established in models of cancer screening. In the U.S. modeling study (19), each of the 7 models of breast cancer screening and treatment that were presented used a version of the stage-shift mechanism. Across the models, screening accounted for a decrease of 7% to 22% (median, 15%) in U.S. breast cancer mortality in 2000, whereas adjuvant systemic therapies accounted for a decrease of 14% to 21% (median, 19%). In our 1999 virtual trial, screening accounted for a decrease in breast cancer mortality of approximately 9% (within-trial MRR, 0.91), whereas systemic therapies accounted for a decrease of approximately 22% (Table 2) (10-year across-trial MRR, 0.78 [when the 1999 and 1975 control groups were compared]). Our finding of a relatively higher contribution of treatment to reductions in mortality could be because we modeled the effect of systemic therapies at all stages of disease rather than focusing on systemic therapies used only in the adjuvant setting for local and regional stages, as was done in Berry and colleagues’ study (19). When we set the treatment HR to 1 for advanced-stage disease (Appendix Table 7), we found that screening accounted for more of the reduction in mortality in our 1999 virtual trial than treatment.
Other changes in the management and diagnosis of breast cancer have occurred since the trials. Examples include the advent of digital mammography and, more recently, tomosynthesis, which may confer improved sensitivity to detect early tumors. In addition, improvements in focal therapies, such as radiation, could affect localized cases of disease. Improvements in supportive therapies might have enhanced the ability of patients to complete systemic treatment regimens and improved the efficacy of these regimens. Our sensitivity analyses were designed to investigate how such changes might affect our main results. We found that with a greater shift from advanced- to early-stage breast cancer at diagnosis, screening efficacy increased but remained relatively unaffected by treatment changes. However, screening efficacy was enhanced when treatment benefit was increased among patients with early-stage disease.
Our study focused on screening benefit in terms of mortality reductions, but personal and policy decisions about mammography screening must also take potential harms into account. In particular, the issue of overdiagnosis has received much attention in recent years (35–37). Because we focused on disease cases that would have been diagnosed only in the absence of screening, we did not address the problem of overdiagnosis. Changes in treatment patterns will not affect overdiagnosis but can affect overtreatment. Changes in screening technologies could affect overdiagnosis, but assessment of overdiagnosis is complex (38) and calls for data and analytics that are beyond the scope of our study.
Because our framework was highly simplified, examining the limitations of the model in light of their potential effect on our main conclusions is important. The most influential simplification was the assumption that the stage-shift mechanism captured the essence of the benefit of early detection. In addition, recent advances have identified features of tumor biology that are highly significant to disease prognosis and therapeutic response (39). Genomic tests have emerged that can effectively classify disease prognosis (40). “Triple-negative” patients with breast cancer seem to have particularly poor survival, and optimal treatment remains undefined (41). Certain tumor types may be more easily detected with existing screening technologies (42). Our tumor subgroups were those for which we could source reliable treatment efficacies; we recognize that the field is evolving and that recent research has identified tumor subgroups within standard classifications, such as by ER status, for which chemotherapy has variable performance (43).
In conclusion, our findings indicate that results from historical breast screening trials are likely to provide reasonable approximations to the relative mortality reductions that would have resulted with contemporary systemic treatments. Our findings may help set realistic expectations about what questions can be answered by the initiation of new screening trials and may inform policymakers and parties responsible for creating screening guidelines.
Grant Support: In part by a Comparative Effectiveness Research dissertation award from the University of Washington Pharmaceutical Outcomes Research and Policy Program and by the National Cancer Institute (grant 1RC4CA155806).
Disclaimer: The results and conclusions are those of the authors only and do not reflect the opinions of the University of Washington or the National Cancer Institute.
Disclosures: Authors have disclosed no conflicts of interest. Forms can be viewed at www.acponline.org/authors/icmje/ConflictOfInterestForms.do?msNum=M15-0754.
Reproducible Research Statement: Study protocol: Available from Dr. Etzioni (gro.hctuhderf@inoizter). Statistical code: Available at https://github.com/netterie/screentreat. Data set: All data are from publicly available sources (see Appendix Table 1). The user interface for the model is available at https://emarkowitz.shinyapps.io/screening_trial_updater.
Author Contributions: Conception and design: J. Birnbaum, V.K. Gadi, E. Markowitz, R. Etzioni.Analysis and interpretation of the data: J. Birnbaum, V.K. Gadi, E. Markowitz, R. Etzioni.
Drafting of the article: J. Birnbaum, V.K. Gadi, R. Etzioni.
Critical revision of the article for important intellectual content: V.K. Gadi, R. Etzioni.
Final approval of the article: J. Birnbaum, V.K. Gadi, R. Etzioni.
Statistical expertise: J. Birnbaum, R. Etzioni.
Obtaining of funding: J. Birnbaum, R. Etzioni.
Collection and assembly of data: J. Birnbaum, E. Markowitz, R. Etzioni.