Are Small Breast Cancers Good because They Are Small or Small because They Are Good?

The new england journal of medicine Special Report Are Small Breast Cancers Good because They Are Small or Small because They Are Good? Donald R. Lannin, M.D., and Shiyi Wang, M.D., Ph.D. The recent article by Welch et al. 1 in the Journal showed clearly that since the adoption of widespread screening mammography, small breast cancers have increased in incidence over three times more than large cancers have decreased. This implies that many small cancers are not destined to progress to large cancers; instead, their detection results in overdiagnosis. The purpose of our study was to characterize groups of tumors that are likely to contain a large portion of overdiagnosed cancers and to explain the mechanism that may have led to their overdiagnosis. Methods Invasive breast cancers that were diagnosed in 2001 2013 were identified from the Surveillance, Epidemiology, and End Results (SEER) database. 2 They were divided into three prognostic groups based on biologic factors: grade, estrogen-receptor (ER) status, and progesterone-receptor (PR) status. There were 12 combinations of these variables, each with a distinct prognosis. The four groups with the worst survival were grade 2 ER-negative and PR-negative, grade 3 ER-negative and PR-negative, grade 3 ER-positive and PRnegative, and grade 3 ER-negative and PR-positive; these were categorized as biologically unfavorable. The groups with the best survival were grade 1 ER-positive and PR-positive, grade 1 ER-positive and PR-negative, and grade 1 ERnegative and PR-positive; these were characterized as biologically favorable. All other groups were considered to be intermediate. The distribution of these three groups across tumor size was determined. We also examined the associations of tumor size and biologic features with breast cancer specific survival. Additional information regarding histologic features, lymph-node metastases, and distant metastases in these groups is provided in the Supplementary Appendix, available with the full text of this article at NEJM.org. We adopted the approach by Etzioni et al. 3 to evaluate the mean lead time (the length of time between when a cancer can be detected by screening and when it would have become clinically apparent without screening) across the three prognostic groups. We first applied the estimate of Welch et al. 1 that the overall rate of overdiagnosis for invasive tumors was 22%. We assumed that the favorable group had the highest rate of overdiagnosis and that the unfavorable group had the lowest rate of overdiagnosis. We varied these rates through plausible ranges. We then generated three virtual populations of women with age distributions as our three prognostic groups. For each group, we simulated life expectancies and lead times with a specified mean. The fraction of women with a life expectancy less than the lead time represented the percent of overdiagnosis. We assumed that the lead time followed an exponential distribution, and we used Weibull distributions with more and less extreme lead times as sensitivity analyses (see the Supplementary Appendix). We varied the mean lead time to find the value that yielded the percentage of overdiagnosis. Similar to Etzioni et al., 3 we also accounted for the observation that women who received screening were healthier than the general population, with a hazard ratio for death equal to 0.75. Results Figure 1 shows the distribution of tumors according to biologic category for each tumor size. Among women 40 years of age or older, tumors with favorable biologic features made up 38.2% 2286 n engl j med 376;23 nejm.org June 8, 2017

Special Report of the tumors that were 1 cm or less in the greatest dimension, and this steadily decreased to only 9.0% of the tumors greater than 5 cm, whereas tumors with unfavorable biologic features made up only 14.1% of the tumors 1 cm or less and increased to 35.8% of the tumors greater than 5 cm. For women younger than 40 years of age, the findings were similar, but the favorable tumors were only about half as common and the unfavorable tumors were much more common. A parallel analysis, shown in the Supplementary Appendix, validates these findings with the use of independent data from the National Cancer Database. Figure 2 shows the breast cancer specific survival among women 40 years of age or older for the favorable and unfavorable groups stratified according to T stage (i.e., T1 tumors from 0.1 to 2.0 cm and T2 tumors from 2.1 to 5.0 cm). Both tumor size and biologic features had a major influence on prognosis, but large tumors with favorable biologic features had a better prognosis than small tumors with unfavorable biologic features. The absolute difference in survival according to tumor size was smaller for biologically favorable tumors and greater for biologically unfavorable tumors. Table 1 shows the results of various models to estimate lead times for the three biologic groups that will still result in an overall rate of overdiagnosis of 22% as proposed by Welch et al. A sensitivity analysis was included to estimate the lead times if the true rate of overdiagnosis were only three quarters or even half of that. Although the estimated lead times in the models varied rather widely, all the models showed that the lead time for the favorable tumors was at least an order of magnitude greater than that of the unfavorable tumors. Figure 3 shows the percentage of overdiagnosis according to age and tumor biologic features for model 2 in Table 1 that is, the model in which the distribution of overdiagnosis was 53% favorable, 44% intermediate, and 3% unfavorable biologic features. The percentage of overdiagnosis for the entire group was 22% as shown by Welch et al., but under the assumption that the lead time varies according to biologic group but not according to patient age, the model shows clearly that overdiagnosis is less common in young women and increases steadily with age. Tumor Biologic Group: A Patients <40 Yr of Age 100 Patients (%) 90 80 70 60 50 40 30 20 10 0 0.1 1.0 1.1 2.0 2.1 3.0 3.1 4.0 4.1 5.0 >5.0 Tumor Size (cm) No. of Patients 4030 9388 7586 3738 1974 3855 B Patients 40 Yr of Age 100 Patients (%) 90 80 70 60 50 40 30 20 10 Favorable Intermediate Unfavorable 0 0.1 1.0 1.1 2.0 2.1 3.0 3.1 4.0 4.1 5.0 >5.0 Tumor Size (cm) No. of Patients 150,751 207,379 106,827 42,803 21,895 38,487 Figure 1. Biologic Characteristics According to Tumor Size. Data are for invasive breast cancers diagnosed in 2001 2013 that were identified from the Surveillance, Epidemiology, and End Results database. The three biologic groups were defined as follows: favorable, grade 1 estrogen receptor (ER) positive and progesterone receptor (PR) positive, grade 1 ER-positive and PR-negative, and grade 1 ER-negative and PR-positive; intermediate, grade 1 ER-negative and PR-negative, grade 2 ER-positive and PRpositive, grade 2 ER-positive and PR-negative, grade 2 ER-negative and PRpositive, and grade 3 ER-positive and PR-positive; and unfavorable, grade 2 ER-negative and PR-negative, grade 3 ER-negative and PR-negative, grade 3 ER-positive and PR-negative, and grade 3 ER-negative and PR-positive. n engl j med 376;23 nejm.org June 8, 2017 2287

The new england journal of medicine Breast Cancer Specific Survival 1.0 0.9 0.8 0.7 0.6 0.0 0 12 24 36 48 60 72 84 96 108 120 Months Figure 2. Breast Cancer Specific Survival among Women 40 Years of Age or Older, According to Tumor Size and Biologic Features. T1 indicates a tumor size of 0.1 to 2.0 cm, and T2 a tumor size of 2.1 to 5.0 cm. The results for the other models were similar and are shown in Figures S2 and S3 in the Supplementary Appendix. Discussion T1 favorable T2 favorable T1 unfavorable T2 unfavorable Figure 1 shows a rather dramatic difference in the distribution of biologic category according to tumor size. Clearly, tumor size depends not just on when the cancer was detected, but also on tumor biologic features. To some extent, tumor size may just be a proxy for good or bad biologic features. If all tumors progressed, even at variable rates, we would expect a steady state to be reached in which there would be a similar distribution of tumor biologic features across size categories. Instead, these data provide fairly direct evidence that many small tumors with favorable biologic features do not progress to large tumors within the lifetime of the patient. In addition, the data imply that large tumors do not arise equally from all small tumors but develop preferentially from a distinct subpopulation of small tumors with unfavorable biologic features. The much larger incidence of favorable tumors in women 40 years of age or older suggests that these tumors are preferentially detected by mammography, because women younger than 40 years of age rarely undergo routine mammographic screening. The survival curves shown in Figure 2 allow us to derive some general principles about the interrelationship between tumor size and tumor biologic features. Both tumor size and biologic features influence prognosis, but frequently a large favorable tumor can have a better prognosis than a small unfavorable tumor. In addition, the difference in prognosis according to tumor size is smaller for favorable tumors and greater for unfavorable ones. Although grade and receptor status are rather crude predictors of biologic features as compared with molecular assays such as Oncotype DX or MammaPrint, it appears that the same relationships will apply. In the recently published Trial Assigning Individualized Options for Treatment (TAILORx), 4 the most favorable 16% of patients as assessed by the Oncotype DX assay had a similarly excellent prognosis regardless of whether their tumor was greater than or no more than 2 cm in the greatest dimension. The results of this trial for tumors with unfavorable biologic features have not yet been reported, but we predict that the outcome for unfavorable tumors will differ significantly according to tumor size. These observations regarding tumor biologic features highlight the need for modeling efforts to incorporate multiple lead times. Etzioni et al. 3 provide a functional definition of overdiagnosis as that which occurs when the time to death from a cause other than cancer is less than the lead time. In other words, it does not matter if the tumor would eventually progress, but whether it would progress within the lifetime of the patient. Therefore, if life expectancy and either lead time or the rate of overdiagnosis is known, the other can be calculated. Etzioni et al. used an estimated mean lead time of 40 months for invasive breast cancers, which was derived from previous studies, and calculated that the rate of overdiagnosis was approximately 7%. However, using a single mean lead time may overlook the effects of tumor biologic features on lead time and overdiagnosis. We used the same relationship and a known rate of overdiagnosis to estimate the lead time for each of the three biologic groups. There is no evidence that unfavorable tumors do not progress, so the only overdiagnosis for that group is due to death in the short term from unrelated causes (only 1 or 2% of deaths). This results in short lead times and is compatible with the clinical observation that high-grade and triple- 2288 n engl j med 376;23 nejm.org June 8, 2017

Special Report Table 1. Estimated Mean Lead Time among Three Tumor Biologic Groups for Various Amounts and Distributions of Overdiagnosis.* Amount and Distribution of Overdiagnosis Favorable (N = 131,896) Tumor Biologic Group Intermediate (N = 318,325) Unfavorable (N = 135,388) 22%, or 128,834 cases Model 1 No. overdiagnosed 52,822 70,859 5153 Rate of overdiagnosis (%) 40.0 22.3 3.8 Mean lead time (yr) 19.9 10.6 2.0 Model 2 No. overdiagnosed 68,282 56,687 3865 Rate of overdiagnosis (%) 51.8 17.8 2.9 Mean lead time (yr) 29.6 8.5 1.4 Model 3 No. overdiagnosed 83,742 42,515 2577 Rate of overdiagnosis (%) 63.5 13.4 2.1 Mean lead time (yr) 44.9 6.4 0.9 16.5%, or 96,625 cases Model 1 No. overdiagnosed 39,616 53,144 3865 Rate of overdiagnosis (%) 30.0 16.7 2.9 Mean lead time (yr) 13.8 8.1 1.4 Model 2 No. overdiagnosed 51,212 42,515 2899 Rate of overdiagnosis (%) 38.8 13.4 2.1 Mean lead time (yr) 19.0 6.4 0.9 Model 3 No. overdiagnosed 62,807 31,886 1933 Rate of overdiagnosis (%) 47.6 10.0 1.4 Mean lead time (yr) 25.6 4.9 0.3 11%, or 64,417 cases Model 1 No. overdiagnosed 26,411 35,429 2577 Rate of overdiagnosis (%) 20.0 11.1 1.9 Mean lead time (yr) 8.9 5.5 0.8 Model 2 No. overdiagnosed 34,141 28,343 1933 Rate of overdiagnosis (%) 25.9 8.9 1.4 Mean lead time (yr) 11.6 4.4 0.3 Model 3 No. overdiagnosed 41,871 21,258 1288 Rate of overdiagnosis (%) 31.7 6.7 1.0 Mean lead time (yr) 14.7 3.3 <0.1 * The three biologic groups were defined as follows: favorable, grade 1 estrogen receptor (ER) positive and progesterone receptor (PR) positive, grade 1 ER-positive and PR-negative, and grade 1 ER-negative and PR-positive; intermediate, grade 1 ER-negative and PR-negative, grade 2 ER-positive and PR-positive, grade 2 ER-positive and PR-negative, grade 2 ER-negative and PR-positive, and grade 3 ER-positive and PR-positive; and unfavorable, grade 2 ER-negative and PRnegative, grade 3 ER-negative and PR-negative, grade 3 ER-positive and PR-negative, and grade 3 ER-negative and PRpositive. The lead time was assumed to follow an exponential distribution. The results of sensitivity analyses that used Weibull distributions with more and less extreme lead times are available in Table S3 in the Supplementary Appendix. We varied the three distributions of overdiagnosis by favorability among plausible ranges. The distributions were as follows: model 1, 41% favorable, 55% intermediate, and 4% unfavorable; model 2, 53% favorable, 44% intermediate, and 3% unfavorable; and model 3, 65% favorable, 33% intermediate, and 2% unfavorable. n engl j med 376;23 nejm.org June 8, 2017 2289

The new england journal of medicine Overdiagnosis (%) Tumor Biologic Group: Favorable Intermediate Unfavorable 100 90 80 70 60 50 40 30 20 10 0 40s 50s 60s 70s 80s Age Group (yr) Figure 3. Percentage of Overdiagnosis According to Age and Tumor Biologic Group for Model 2 with Respect to Distribution of Overdiagnosis. This analysis assumes that the tumor lead time varies according to biologic features but not according to age group. The percent overdiagnosis for the entire group combined was 22% as found by Welch et al. 1 In model 2, the distribution of overdiagnosis was 53% favorable, 44% intermediate, and 3% unfavorable biologic features. For similar data for the other models, see Figure S3 in the Supplementary Appendix. negative tumors frequently present as interval cancers with a lead time of less than 1 year. 5 However, all the models shown in Table 1 that use the 22% rate of overdiagnosis that was suggested by Welch et al. indicate that the lead time for the favorable cancers is greater than 19 years. A sensitivity analysis shows that even if the rate of overdiagnosis were only half of that found by Welch et al., the mean lead time is still at least 8.9 years. Previous estimates of breast-cancer lead times, primarily in patients from the 1960s and 1970s, used tumors that eventually became palpable and therefore were more representative of our unfavorable or intermediate group. 6,7 The relationships shown in Figure 1 and Figure 2 explain both why screening mammography causes considerable overdiagnosis and also why its effectiveness is limited. Because of the long lead times, mammography is very good at detecting tumors with favorable biologic features, and therefore these tumors are overrepresented among small tumors. However, many of them do not progress within the patient s lifetime; therefore, they contribute substantially to overdiagnosis. Furthermore, the ones that do progress still have an excellent prognosis even when they become large, so there is little benefit to detecting them early. In contrast, the prognosis for the tumors with unfavorable biologic features is considerably better if they can be diagnosed when under 2 cm in size. Unfortunately, because of the short lead times, they are rarely diagnosed early and therefore are substantially underrepresented among small tumors. There are several caveats that should be mentioned regarding this analysis. First, we considered only invasive cancers. Noninvasive cancers are not just small cancers; their biologic features are very different, and the benefits and harms of detection are different as well. They warrant a separate analysis. Second, neither this analysis nor the study by Welch et al. had data regarding which tumors were actually diagnosed by screening mammography; this was just inferred. However, Hayse et al. 8 used an institutional database at Yale University that included the method of detection and showed directly that diagnosis by mammography was associated with low-grade cancers with a favorable molecular profile. Similarly, a study involving patients enrolled in the MINDACT (Microarray in Node-Negative and 1 to 3 Positive Lymph Node Disease May Avoid Chemotherapy) trial showed that mammography resulted in tumors diagnosed with favorable biologic features as assessed by MammaPrint. 9 Third, our analysis assumes that the favorable low-grade, hormone receptor positive tumors do not dedifferentiate over time to high-grade or receptornegative tumors. This seems consistent with current findings. Although prospective studies do show a discordance in receptor status of 10 to 15% between the primary tumor and subsequent distant metastases, these changes are usually the result of selection pressures on the tumor because of treatment. 10 It is thought that in the primary tumor, low-grade and high-grade tumors arise by different molecular mechanisms, 11 and it is very rare for a low-grade tumor to dedifferentiate into a high-grade tumor. 12 Fourth, it is likely that tumor biologic features represent a continuous variable, and our analysis, similar to those with Oncotype DX recurrence scores, involves dividing the patients into prognostic groups with the use of somewhat arbitrary cut points. We believe that it is useful to appreciate that the mean lead time for our favorable group is 15 to 20 years. However, we certainly could not rule out that there could be subpopulations within our favorable group with shorter lead times and others that do not have disease progression at all or that even have disease regression. Fifth, we did not include human epidermal growth factor receptor 2 (HER2) as one of the 2290 n engl j med 376;23 nejm.org June 8, 2017

Special Report biologic factors because it has been available in the SEER data set only since 2010 and therefore the available follow-up is short. In the future, prognostic groups might be better defined by incorporating HER2, lymphovascular invasion, histologic features, and molecular assays such as Oncotype DX and MammaPrint. When survival improves for a group of patients matched according to tumor stage, it is usually ascribed to improved treatment. 1 However, it could also be due to changes in tumor biologic features. Our most favorable group, which quite likely contains a large portion of the overdiagnosed cancers (grade 1 hormone receptor positive tumors <2 cm), had a 10-year breast cancer specific survival of 97%. If our estimates are correct, approximately 50% of the cancers in the favorable group in TAILORx may be overdiagnosed. No wonder the survival in this group was so good. The conclusion that these patients do not need chemotherapy represents an important advance to prevent unnecessary treatment. The data shown in Figure 3 illustrate that overdiagnosis is much more prevalent in the elderly than in younger persons. This does not mean, however, that all the excess cancers found by Welch et al. were in the elderly. Because of the long lead times, many cancers that were destined to be diagnosed in women in their 70s are now diagnosed in women in their 50s or 60s, but they are not considered overdiagnosed because they would have been diagnosed later. On the other hand, a large proportion of new screening-detected cancers in women in their 70s would not be diagnosed within the patient s lifetime if it were not for the screening. This may have a bearing on recommendations about when to stop mammographic screening. The good news is that there may be new opportunities to help clinicians manage the problem of overdiagnosis. The eighth edition of the Cancer Staging Manual of the American Joint Committee on Cancer 13 that has just been released has added formalized prognostic groups that include grade and receptor status very similar to those used in this analysis, and it even incorporates some elements of assays such as Oncotype DX and MammaPrint. This means that clinicians will be able to prospectively identify patients with stage, age, and biologic profiles that suggest possible overdiagnosis. Of course, in an individual case it is not yet possible to say with certainty that a cancer is overdiagnosed, so treatment cannot be withheld. However, as outcome data accumulate, trials can be designed to provide less treatment to more favorable groups. Two of the major harms of overdiagnosis are the overtreatment that results and the anxiety and fear that a cancer diagnosis engenders. Educating physicians, patients, and the public that some cancers are indolent and individualizing treatment algorithms to provide personalized medicine will aid in addressing this problem. Disclosure forms provided by the authors are available with the full text of this article at NEJM.org. From the Department of Surgery, Yale University School of Medicine (D.R.L.), Yale University School of Public Health (S.W.), and Yale Comprehensive Cancer Center (D.R.L., S.W.), New Haven, CT. Address reprint requests to Dr. Lannin at the Department of Surgery, Yale University School of Medicine, P.O. Box 208062, New Haven, CT 06520, or at donald.lannin@yale.edu. 1. Welch HG, Prorok PC, O Malley AJ, Kramer BS. Breast-cancer tumor size, overdiagnosis, and mammography screening effectiveness. N Engl J Med 2016; 375: 1438-47. 2. Surveillance, Epidemiology, and End Results (SEER) Program. Research data (1973 2013), National Cancer Institute, DCCPS, Surveillance Research Program, Surveillance Systems Branch, released April 2016, based on the November 2015 submission (https:/ / seer.cancer.gov/ ). 3. Etzioni R, Xia J, Hubbard R, Weiss NS, Gulati R. A reality check for overdiagnosis estimates associated with breast cancer screening. J Natl Cancer Inst 2014; 106: 106. 4. Sparano JA, Gray RJ, Makower DF, et al. Prospective validation of a 21-gene expression assay in breast cancer. N Engl J Med 2015; 373: 2005-14. 5. Collett K, Stefansson IM, Eide J, et al. A basal epithelial phenotype is more frequent in interval breast cancers compared with screen detected tumors. Cancer Epidemiol Biomarkers Prev 2005; 14: 1108-12. 6. Shen Y, Zelen M. Screening sensitivity and sojourn time from breast cancer early detection clinical trials: mammograms and physical examinations. J Clin Oncol 2001; 19: 3490-9. 7. Duffy SW, Chen HH, Tabar L, Day NE. Estimation of mean sojourn time in breast cancer screening using a Markov chain model of both entry to and exit from the preclinical detectable phase. Stat Med 1995; 14: 1531-43. 8. Hayse B, Hooley RJ, Killelea BK, Horowitz NR, Chagpar AB, Lannin DR. Breast cancer biology varies by method of detection and may contribute to overdiagnosis. Surgery 2016; 160: 454-62. 9. Drukker CA, Schmidt MK, Rutgers EJ, et al. Mammographic screening detects low-risk tumor biology breast cancers. Breast Cancer Res Treat 2014; 144: 103-11. 10. Thompson AM, Jordan LB, Quinlan P, et al. Prospective comparison of switches in biomarker status between primary and recurrent breast cancer: the Breast Recurrence In Tissues Study (BRITS). Breast Cancer Res 2010; 12: R92. 11. Reis-Filho JS, Simpson PT, Gale T, Lakhani SR. The molecular genetics of breast cancer: the contribution of comparative genomic hybridization. Pathol Res Pract 2005; 201: 713-25. 12. Schymik B, Buerger H, Krämer A, et al. Is there progression through grade in ductal invasive breast cancer? Breast Cancer Res Treat 2012; 135: 693-703. 13. Amin MB, Edge SB, Greene FL, et al., eds. AJCC cancer staging manual. 8th ed. New York: Springer, 2017. DOI: 10.1056/NEJMsr1613680 Copyright 2017 Massachusetts Medical Society. n engl j med 376;23 nejm.org June 8, 2017 2291