The newly developed SimDCIS model was successfully validated, with an excellent match of simulated screen-detected DCIS stratified by age and grade to observed data from Dutch and UK screening. Model uncertainty was highest for a change in DCIS onset probability, with a maximum variation in screen-detection rate of 12%. DCIS estimates for regression, progression to IBC, clinical detection, and screen-detection in the Dutch screening setting were 7% (with a range of 0–14%), 19% (15–24%), 7% (0–14%), and 63% (58–68%), respectively. Screen-detected grade distribution was 20%, 38%, and 42% for grades 1, 2, and 3, respectively.
The SimDCIS model was successfully constructed and validated with accurate simulations of the Dutch and UK screening settings for screen-detected DCIS by age and grade. Internal, external, and cross-validation showed that simulated data matched observed data. Only a slight overestimation of women aged 60–64 years in Dutch screening and an overestimation of women aged 50–54 years and an underestimation of 70–74 years in UK screening were observed. Observed Dutch data showed a decrease for this specific age group, which was not present in simulated data. This sudden decrease could be explained by the very small number of DCIS cases, especially within a 5-year age group (e.g., approximately 250 DCIS in 200,000 women aged 60–64), which can result in sudden data fluctuations. Additionally, the number of mammograms per age group in the model was based on 2011, as this was the most recent data found per age group, which may have resulted in a slightly different distribution of mammograms per age. Overestimation of 50–54 years of age and underestimation of 70–74 years in the UK setting could be explained by deviations in age at screening in the UK setting. DCIS was detected at screening in women aged 45–49 and 70 + years in observed data as a part of the AgeX trial active in the UK since 2009 to assess the effect of inviting women age 47–49 and 71–73 years (30). This caused some DCIS to be detected before 50 and after 71 years, which was not possible in the model. Apart from these subtle differences, overall validation was very good. Therefore, SimDCIS was shown to accurately simulate the natural history of DCIS and deemed suitable to provide accurate estimates of DCIS.
New DCIS estimates from SimDCIS were in line with previous literature. Regression of DCIS was previously estimated at 1–10% (6, 20). SimDCIS showed 7% regression, with a broader range of 0–14%. This range was only broader when the regression input parameter was varied to its boundaries. For all other parameter variations, regression was estimated at 7–9% (Appendix A.4.2). SimDCIS showed 15–24% progression of DCIS to IBC suggesting that only approximately 1 in 5 DCIS will progress to IBC. Previous estimates of DCIS progression to IBC ranged from 20–91% (6), but recent literature shows evidence for the lower end of the range. Non-progressive screen-detected breast cancers were estimated at 15.7% in Norwegian screening, with 18% DCIS of total breast cancers as probable largest contributor, which suggests a similar estimate (31). In addition, a recent review suggested that only 1 in 6 of detected DCIS will progress to IBC (32). Screen-detected grade distribution was estimated at 20%, 38%, and 42% for grades 1, 2, and 3, which matched Dutch data of 17%, 38%, and 45%, respectively (5). Previous literature reported 18–20%, 31–32%, and 48–51% for grades 1, 2, and 3 in the Netherlands up to 2015 (7, 33), whereas pathology studies have found 14%, 42%, and 43%, respectively (3). This might indicate a relatively lower detection of grades 2 and 3 compared to grade 1 in the current Dutch screening setting.
There is consensus that DCIS contributes considerably to overdiagnosis, while on the other hand, increased screen-detection rate of DCIS is associated with lower IBC rates (3, 34). The extent to which overdiagnosis occurs is difficult to establish given the unobservable natural history of DCIS due to the standard of immediate treatment upon detection (3). Improved estimation of DCIS natural history parameters and DCIS estimates in screening setting will improve estimations of DCIS overdiagnosis, which currently range from 20–91% (3, 6, 9). Improved overdiagnosis estimates could provide insight into which women could be offered active surveillance instead of treatment, as women with low-risk DCIS offered the choice between conventional treatment and active surveillance showed a preference for the latter (35).
This study has several limitations. Firstly, the model is limited by the availability and quality of existing data. Most data used were recent data from the Netherlands and UK, which are both high-quality databases, but for number of mammograms per age, the most recent data available were from 2011. Also, the UK Frequency Trial took place from 1989 to 1996. As this was the only comparable incidence data used by a DCIS model, it was suitable for cross validation with adjustment of mortality, screening sensitivity, and participation, as DCIS natural history is not expected to be different over time. Secondly, the model only provides a simplification for the general population. Simulations of high-risk women were not included, which suggests a possible underestimation of present DCIS and clinical detection, although the number of DCIS was very small, and model uncertainty is expected to be significantly larger. Additionally, the focus of this study was on screen-detected DCIS. Most diagnoses that contribute to overdiagnosis are the screen-detected tumours, and clinical detection is rare. Thirdly, this study focused on DCIS, although approximately 80% of breast cancers are invasive. For the aim of this study, the total number of IBC was not relevant and considered outside the scope. For future research on total breast cancer, the well-validated SiMRiSc model with focus on IBC could be combined with SimDCIS (9, 36, 37).
Also, several strengths should be considered. SimDCIS includes age- and grade-dependent input parameters, to provide more accurate estimates based on these two well-established risk factors for DCIS onset and progression (6, 7). Previous DCIS models have been constructed, but showed a large range of conclusions (6, 7, 11–16), were based on a large range of assumptions (4, 6), or did not include differences in progression to IBC for varying ages and grades (6). Furthermore, modelling studies should provide a full validation and sensitivity analysis to ensure that results are valid and reliable and that assumptions are reported to enhance transparency (17, 18). In this study, internal, external, and cross-validation and univariate and probabilistic sensitivity analyses were performed, which showed valid and reliable results. In addition, SimDCIS only needed ~ 35 seconds to run 1 scenario of 10 iterations with 100,000 women. Therefore, SimDCIS contributes to breast cancer research because it accurately simulates DCIS natural history, provides fast and accurate estimates, and assesses benefits and harms of different scenarios. Also, DCIS estimates were made in many scenarios, and still narrowed existing estimates. Therefore, the estimates from this study can inform on existing gaps in knowledge and contribute to potentially more accurate estimates of overdiagnosis.