Introduction

Approximately 20% of U.S. women diagnosed with breast cancer have ductal carcinoma in situ (DCIS) [1]. This percentage is even greater (over 27%) among women who regularly participate in mammography screening [2]. Approximately 85% of DCIS is diagnosed asymptomatically through screening mammography [3]. As of 2016, over 1 million women are estimated to have a personal history of a DCIS diagnosis [4]. Several lines of research suggest that a substantial fraction of DCIS—perhaps more than half—would never have caused the women harm if the DCIS had remained undetected [5,6,7,8]. Because a significant proportion will progress to invasive disease, treatment for DCIS typically is similar to approaches for locally limited invasive breast cancer including excision, radiation, and adjuvant endocrine therapy. Increasingly, women are choosing ipsilateral mastectomy, contralateral risk reduction mastectomy, and breast reconstruction which have substantial impacts relating to quality of life and financial costs [9].

In 2011, a new multigene assay for estimating the 10-year risk of local recurrence, the Oncotype DX® DCIS Score™, became available for women with DCIS treated by local excision [10]. Although use of the DCIS score assay is limited due to concerns of patient eligibility and lack of consideration of non-molecular predictive factors such as margin width and nuclear grade [11], independent studies confirm that this 12-gene assay predicts local recurrence after DCIS [12] and influences decisions regarding radiotherapy after breast-conserving surgery [13]. Additional approaches using features to identify the malignant potential of a DCIS lesion are actively under investigation, most often by combining different sources of information regarding histological and molecular markers as well as method of detection [14,15,16].

In anticipation of the development of more accurate methods of identifying non-progressive DCIS, we used the University of Wisconsin Breast Cancer Simulation Model to evaluate the impact of a hypothetical prognostic test for DCIS. Simulation modeling is a valuable tool to estimate the population impact of a new or hypothetical intervention as it allows for using multiple sources of input data without risk to human subjects. We started by assuming a perfect prognostic test to estimate, under ideal circumstances, the maximum impact on costs and quality of life if non-progressive DCIS lesions were not treated beyond diagnostic biopsy, sparing women further treatment. We then explored the impact of alternative assumptions of test performance, DCIS tumor growth, and screening mammography utilization.

Methods

University of Wisconsin Breast Cancer Simulation Model

The University of Wisconsin Breast Cancer Simulation Model is a discrete-event simulation model that uses multiple sources of U.S. cancer data to generate registry-like datasets of breast cancer incidence and mortality statistics according to patient demographics [17]. The model was developed as part of the Cancer Intervention and Surveillance Modeling Network (CISNET), a consortium funded by the National Cancer Institute [18, 19]. The simulation model is used to consider the benefits and harms of alternate screening and treatment programs for breast cancer [20,21,22,23,24]. The University of Wisconsin Health Sciences Human Subjects Committee determined that this study was exempt from review.

Briefly, the simulation model has four main components—natural history, detection, treatment, and non-breast cancer mortality; [25] additional information regarding the model is included in Fig. 1 and Online Appendix Table 1 [18, 23, 26,27,28,29,30,31,32,33,34,35,36,37,38,39]. The natural history component assumes that the probability of breast cancer onset is a function of risk based on trends in breast cancer according to single years of age at diagnosis, years of diagnosis, and years of birth [26, 27]. In the model, all invasive breast cancer tumors first progress through DCIS, so that a “tumor” can be either invasive or DCIS. Tumor growth is a function of a random initial growth parameter, with a fraction of tumors (initially set through calibration) having limited malignant potential, i.e., non-progressive growth. Thus, in the primary analysis, we did not assume a priori a specific fraction of non-progressive disease; rather, the fraction of non-progressive disease (30–50% with a “best” estimate of 42%) was determined by testing different fractions for the best fit to the observed historical patterns in stage-specific breast cancer incidence. In other words, not all DCIS or small localized invasive cancers progress to become a lethal event. After reaching a maximum size, also set through calibration, non-progressive tumors stop growing. If undetected, all non-progressive tumors become undetectable or disappear from the model after reaching their maximum size for a specified time set through calibration (2 years).

Fig. 1
figure 1

Schema representing breast cancer natural history, detection, and treatment as represented in the simulation model: a in the absence of a DCIS prognostic test and b in the presence of a perfect DCIS prognostic test. If the DCIS prognostic test is positive, indicating that the DCIS is non-progressive, then the woman receives no treatment and eventually dies from a cause other than breast cancer. If the DCIS prognostic test is negative, the DCIS is progressive, the woman receives breast cancer treatment, and survival depends on whether the cancer is cured (determined stochastically) and on the efficacy of treatment. See Table 1 for the description of the implementation of the DCIS prognostic test and cancer detection in the simulation model

Full size image

The detection component assumes that breast cancer can be detected either by screening mammography or through clinical surfacing by clinical examination or self-detection [27]. The probability that any undetected tumor will be symptomatically detected is a function of tumor diameter, so that larger tumors are more likely to be clinically detected at any time relative to smaller tumors. Screen detection of a tumor during the preclinical detectable period could result in the identification and treatment of earlier stage or smaller tumors and lead to breast cancer mortality reduction. Sensitivity of digital mammography depends on a woman’s age (< 50, 50–69, and ≥ 70), breast density, and also on tumor diameter and calendar year [25]. In addition, the probability that a woman has a digital mammogram is a function of calendar year and age, so that it is more likely that a woman will have a mammogram with increasing age and time [30]. Screening mammography rates are based on observed dissemination patterns in the U.S. estimated from the Breast Cancer Surveillance Consortium [30, 40].

Breast cancer treatment is modeled using a “cured”/“not cured” approach. Uncured tumors continue to grow until they reach the distant disease stage and, eventually, cause death from breast cancer unless the woman dies from other causes or reaches age 100 (the end of the simulation). The probability of cure depends on the treatment given, tumor stage at detection, and age [25]. Surgical treatment assignment depends on the year when the tumor was detected based on patterns of care observed in the SEER Program and the National Cancer Data Base [9, 23]; adjuvant therapy is assigned based on stage of disease at detection and ER/HER2 status [41]. Survival rates for specific treatments including surgery, radiation, chemotherapy, endocrine therapy, and targeted therapy (e.g., trastuzumab) are based on recent clinical trial results [23, 32]. The non-breast cancer mortality component assumes that death from other causes will occur independently of a woman’s breast cancer status, and the probability of death from other causes depends on age [38, 42].

The simulated population included women born in 1970. This birth cohort was chosen since these women experience modern conditions (e.g., digital mammography performance, treatment effectiveness, competing mortality, etc.) and for consistency with recent collaborative modeling reports [18, 22]. Current digital mammography performance and treatment effectiveness were used for future calendar years. In each simulation, subgroups of women were followed from age 20 until death or age 100, whichever occurs first.

Outcomes

In all simulated scenarios, primary outcomes were quality-adjusted life years (QALYs) and costs. We used age-specific quality of life weights for general health based on utilities from noninstitutionalized female civilians in the U.S. (Online Appendix Table 1) [33]. Adjustments were made for stage-specific breast cancer treatment. Costs were divided into annual surveillance mammography costs among diagnosed DCIS cases who were not treated beyond biopsy, stage-specific treatment costs (including subsequent surveillance mammography among treated cases) [19, 36], and prognostic test costs, reported in 2015 U.S. dollars (Table 1). Cost of the hypothetical DCIS prognostic test was assumed to be $3419.42, based on Medicare reimbursement for a currently available diagnostic test (Oncotype DX® Breast DCIS Score™, HCPSC 81519). Life years, QALYs, and costs were discounted at 3% per year.

Table 1 Values for costs and quality of life decrements used in the simulation model
Full size table

Secondary outcomes included breast cancer deaths, life years, the numbers of prognostic tests, and the number of DCIS cases treated (or not treated).

Comparison scenario: current practice for DCIS management

As our primary comparison scenario, we assumed current mammography dissemination and modern breast cancer treatment. The hypothetical DCIS prognostic test was not implemented. All women diagnosed with breast cancer were given treatment (100% adherence) according to guidelines.

Alternate scenarios: DCIS prognostic test

One primary alternate scenario was modeled by modifying the simulation model to include application of a perfect DCIS prognostic test at the time of diagnosis (Fig. 1, panel B). If the test was “positive,” indicating non-progressive disease, then the case was not treated beyond diagnostic biopsy, thus also avoiding a decrement to quality of life associated with treatment. DCIS cases with positive prognostic tests were assigned annual mammography (“surveillance” mammography after diagnosis) until death or a diagnosis with invasive breast cancer. (Under this scenario with the perfect test, no “non-progressive” DCIS progressed to invasive breast cancer; this only occurred with false-positive prognostic tests in the sensitivity analysis, described below.) If the prognostic test was “negative,” indicating that the tumor was progressive, then the case received standard treatment.

Analysis

Two million women were simulated for each scenario. Outcomes were compared between the scenarios with and without a perfect hypothetical DCIS prognostic test (100% sensitivity, 100% specificity). We compared additional scenarios to examine the effects of variation in prognostic test performance characteristics, quality of life decrements, and model parameters governing disease progression. Prognostic test performance was varied such that sensitivity of the prognostic test to identify non-progressive disease ranged from 50 to 95%, while specificity of the test to identify progressive disease ranged from 70 to 95%. We varied decrements in quality of life associated with breast cancer treatment; specifically, we increased the disutility of treatment for DCIS and localized invasive breast cancer (from 0.1 to 0.2) since stage-specific treatment disutilities were initially small for early-stage disease (Table 1) [19, 35]. We varied model parameters regarding non-progressive tumors by setting the percent of DCIS tumors designated as non-progressive as 30 or 50%, instead of the “best” value set through calibration (42%), to reflect the range generally reported in the literature. Finally, we examined alternative schedules for screening mammography (for women without a personal history of breast cancer) including those recommended by the U.S. Preventive Services Task Force (biennial ages 50–74 year), the American Cancer Society (annual 45–54 year, biennial 55–79 year), and the American College of Radiology (annual 40–74 year).

Results

In the absence of a DCIS prognostic test, the model predicted that 13.9% of women were diagnosed with breast cancer in their lifetimes. Slightly over 20% of breast cancers were DCIS (20.6%); overall, 8% of all breast cancer cases and 3% of DCIS cases died of breast cancer. Costs were estimated to equal $13,321 per DCIS case.

In the perfect test scenario, the DCIS prognostic test was applied to 2.8% of all women (2844 per 100,000), with approximately 60% of tests being positive for a non-progressive lesion (1696 per 100,000) (Table 2). Although 42% of all breast cancer tumors were non-progressive at onset, the percent of non-progressive tumors decreases with increasing stage at diagnosis. Thus, 60% of DCIS cases were spared treatment. Among women diagnosed with DCIS, the gain in QALYs was modest (0.044 per case, or 16 days of perfect health) and mortality was unchanged (3.0%); however, the reduction in costs was substantial. All DCIS cases experienced the cost of the prognostic test with an estimated total expected discounted cost per DCIS case of $1598 (assuming a 3% annual discounting factor for future costs). DCIS cases with positive prognostic tests incurred surveillance mammography costs ($705 per case), whereas DCIS cases with negative tests incurred both surveillance mammography and treatment costs. Total lifetime costs were reduced by 40% averaged across all DCIS cases.

Table 2 Outcomes comparing scenarios with and without a perfect DCIS prognostic test
Full size table

In scenarios where specificity remained perfect (no progressive disease was misidentified as non-progressive) but sensitivity was decreased as low as 20% (as little as 20% of non-progressive disease tested positive), breast cancer mortality among cases diagnosed with DCIS remained unchanged and the effect on QALYs was small, but costs increased from $8334 to $13,521 per DCIS case (Table 3). In scenarios with perfect sensitivity but reduced specificity (some progressive disease was misidentified as non-progressive), the effect on QALYs and costs was modest and breast cancer deaths among women diagnosed with DCIS increased by 0.1–0.4% (2–11 per 100,000 women). Changes in sensitivity had a greater effect on the number of DCIS cases that avoided treatment than changes in specificity (Fig. 2).

Table 3 Sensitivity analyses of outcomes based on a DCIS prognostic test with a range of test sensitivity and specificity settings
Full size table
Fig. 2
figure 2

Reduction in the number of DCIS cases treated after diagnosis according to the sensitivity and specificity of the hypothetical DCIS prognostic test. Positive DCIS prognostic test indicates non-progressive disease, and a negative DCIS test indicates progressive disease

Full size image

Changing the decrement in quality of life associated with DCIS and localized invasive treatment from 0.1 to 0.2 for 2 years doubled the gain in QALYs for DCIS cases in the presence of a perfect test from 16 to 32 days (Table 2). When we varied assumptions by increasing the prevalence of non-progressive breast tumors, the benefits of a DCIS prognostic test were greater, with a savings of almost $1000 per DCIS case if 50% were non-progressive instead of 42%.

Varying screening schedules did not substantially impact outcomes (Online Appendix Table 2). Screening strategies with more frequent mammography generally lead to greater numbers of detected DCIS with slightly fewer cases avoiding treatment and slightly greater costs in the presence of a DCIS prognostic test, but these differences were small.

Discussion

This modeling study demonstrates the maximum potential impact of a hypothetical prognostic test for non-progressive DCIS. Results showed that, under optimal conditions, treatment cost savings would be substantial in the presence of an accurate prognostic test. Given that about 63,000 women are diagnosed with DCIS in the U.S. each year [43], this corresponds to an annual savings of over $334 million. The availability of a perfect DCIS prognostic test would improve quality of life, on average, about 2–4 weeks of perfect health across all women diagnosed with DCIS. The estimated utility of the perfect prognostic (0.044) is slightly greater than the quality of life impact of a screening mammogram (0.01–0.03) but smaller than the impact of breast-conserving surgery (0.07–0.09) [44,45,46].

Our primary analysis assumed that the hypothetical prognostic test was perfectly accurate. While this assumption is unrealistic, these results demonstrate the optimal impact of a prognostic test that can be anticipated. Our scenario analyses demonstrate how the projected costs and the number of DCIS cases that avoided treatment are dependent on test sensitivity; conversely, breast cancer mortality depends on test specificity. The cost of the prognostic test was based on current Medicare reimbursement for the Oncotype DX test which is greater than the costs of surveillance mammography or breast MRI but less than the costs of breast cancer treatment. We estimated costs of surveillance mammography in women with a positive diagnostic test, but it is likely that a variety of surveillance approaches would be utilized in the setting of a prognostic test (e.g., imaging with MRI or more frequent mammography), so that our analysis did not fully estimate the variety or magnitude of surveillance costs.

The breast cancer simulation model has been calibrated to match breast cancer incidence and mortality as observed in the US, and incorporates inputs on modern screening performance and treatment benefit. Other model inputs are unobservable or difficult to quantify, so that the model requires certain assumptions. We explored the extent that assumptions regarding tumor growth, quality of life utilities, and the frequency of mammography screening had on the results. In particular, the percent of tumors designated at biological onset as non-progressive (42%) was initially set through model calibration [17]. We conducted a series of secondary analyses to explore the impact of this parameter and found that the results were sensitive to the fraction of tumors that were non-progressive; the value of a prognostic test increases along with the fraction of tumors that are non-progressive. The proportion of DCIS lesions that are non-progressive is debated, with studies finding a very wide range of potential values, from 10 to 80% [5,6,7,8]. We also assumed that positive results of the hypothetical prognostic test meant that women avoided all treatments including excision, whereas potentially more likely clinical scenarios for such a test may include breast-conserving surgery or endocrine therapy and avoidance of some but not all treatment approaches. Thus, the cost savings resulting from our modeling experiments describe the most optimistic scenario in terms of cost savings. Treatment costs used in this study arise from studies that provided values for the phase of care (initial, interim, end of life) rather than for the type of treatment (surgery, radiation, endocrine therapy) [36], thus preventing the estimation of cost savings when specific treatments are omitted. Our model was also limited since women were at risk of only a single breast cancer diagnosis; future studies should consider risk of multiple primaries since over 10% of breast cancer survivors experience a second primary diagnosis [47]. In addition, future analysis should account for up-staging that occurs among approximately 24% of women who are diagnosed with DCIS at biopsy but have invasive breast cancer detected at the time of surgery [48]. Finally, we did not estimate emotional costs such as anxiety that might be associated with DCIS testing and forgoing treatment. Randomized trials investigating these issues will shed light on these qualitative components of the clinical cascade, which will then be available for modeling in the future.

Efforts to identify indolent DCIS and improve decision-making surrounding treatment for early-stage breast cancer are ongoing, and recent developments in molecular-based assays will likely play an important role in future DCIS management. Alternative approaches to immediate surgery, including active surveillance using advanced imaging, are also under evaluation [49]. While randomized trials will ultimately provide evidence regarding the benefits of actual prognostic tests in the future, simulation modeling is a valuable tool to explore unobservable phenomenon like tumor growth in asymptomatic women. Modeling also allows incorporation of multiple data sources to inform the analysis based on modern breast cancer screening and treatment. This study provides estimates of the potential impact of implementing a tailored approach to DCIS therapy, where women with greater risk of recurrence or a second breast cancer diagnosis receive more extensive treatment. Our findings suggest that a DCIS prognostic test that could identify non-progressive DCIS would substantially reduce costs associated with DCIS management.