Journal of Cancer Epidemiology

Volume 2019, Article ID 5072506, 14 pages

https://doi.org/10.1155/2019/5072506

## Use of Mastectomy for Overdiagnosed Breast Cancer in the United States: Analysis of the SEER 9 Cancer Registries

^{1}Data Scientist, Seattle, WA 98102, USA^{2}Department of Physics, Harvard University, Cambridge, MA 02138, USA^{3}Exergen Corp., Watertown, MA, USA^{4}Worldpay, Lowell, MA, USA

Correspondence should be addressed to C. Harding; ude.dravrah.tsop@gnidrahc

Received 25 August 2018; Revised 24 November 2018; Accepted 23 December 2018; Published 22 January 2019

Academic Editor: Eleanor Kane

Copyright © 2019 C. Harding et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

*Aim.* We investigated use of mastectomy as treatment for early breast cancer in the US and applied the resulting information to estimate the minimum and maximum rates at which mastectomy could plausibly be undergone by patients with overdiagnosed breast cancer. Little is currently known about overtreatments undergone by overdiagnosed patients.* Methods. *In the US, screening is often recommended at ages ≥40. The study population was women age ≥40 diagnosed with breast cancer in the US SEER 9 cancer registries during 2013 (n=26,017). We evaluated first-course surgical treatments and their associations with case characteristics. Additionally, a model was developed to estimate probability of mastectomy conditional on observed case characteristics. The model was then applied to evaluate possible rates of mastectomy in overdiagnosed patients. To obtain minimum and maximum plausible rates of this overtreatment, we respectively assumed the cases that were least and most likely to be treated by mastectomy had been overdiagnosed.* Results. *Of women diagnosed with breast cancer at age ≥40 in 2013, 33.8% received mastectomy. Mastectomy was common for most investigated breast cancer types, including for the early breast cancers among which overdiagnosis is thought to be most widespread: mastectomy was undergone in 26.4% of* in situ* and 28.0% of AJCC stage-I cases. These rates are substantively higher than in many European nations. The probability-based model indicated that between >0% and <18% of the study population could plausibly have undergone mastectomy for overdiagnosed cancer. This range reduced depending on the overdiagnosis rate, shrinking to >0% and <7% if 10% of breast cancers were overdiagnosed and >3% and <15% if 30% were overdiagnosed.* Conclusions.* Screening-associated overtreatment by mastectomy is considerably less common than overdiagnosis itself but should not be assumed to be negligible. Screening can prompt or prevent mastectomy, and the balance of this harm-benefit tradeoff is currently unclear.

#### 1. Introduction

Here, we studied the use of mastectomy for early and overdiagnosed breast cancers. We specifically sought to evaluate how often overdiagnosed breast cancers were treated by mastectomy, which is a form of overtreatment. Although many studies have investigated the overdiagnosis of early breast cancer [1] or the use of mastectomy for early breast cancer, few have investigated both [2–9]. If sufficiently common, use of mastectomy for overdiagnosed breast cancer could be one of largest inadvertent harms in cancer treatment. We therefore thought it was worth studying. We conducted this evaluation using data from women diagnosed with breast cancer in 2013 in the Surveillance, Epidemiology, and End Results 9 US cancer registries.

The relationship between screening and mastectomy is complicated because screening can prevent mastectomies from being needed in some cases and cause mastectomies to be performed “unnecessarily” in other cases: If screening allows a harmful breast cancer to be detected at an earlier stage than would otherwise be possible, then use of mastectomy may be averted. On the other hand, if screening leads to overdiagnosis, then mastectomy may be performed “unnecessarily.” Because of this complication and because the overall rate of overdiagnosis is currently unknown and controversial [10, 11], it is not possible to calculate the exact amount of overtreatment by mastectomy that occurs after overdiagnosis. Instead, we have the more modest goal of determining minimum and maximum rates at which mastectomy could plausibly be undergone by overdiagnosed patients. (In other words, we aim to find lower and upper bounds on the rate of mastectomy in overdiagnosed patients.)

There is often an expectation that overdiagnosed patients rarely undergo mastectomy, based on the assumption that mastectomy is usually performed for aggressive-appearing cancers that are unlikely to be overdiagnosed. To date, however, this expectation has not been tested in the US. In placing bounds on how often overdiagnosed patients were treated by mastectomy, we sought to determine whether this expectation is justified.

#### 2. Materials and Methods

##### 2.1. Data Source

Data on 27,389 women diagnosed with* in situ* or invasive breast cancer in the year 2013 were obtained from Surveillance, Epidemiology, and End Results Cancer Registries (SEER) grouping 9, which includes the following regions: San Francisco-Oakland, Connecticut, the Detroit Metropolitan Area, Hawaii, Iowa, New Mexico, the Seattle Puget Sound Area, Utah, and the Atlanta Metropolitan Area. Of the 27,389 women, we excluded 1137 (4.2%) who were diagnosed with breast cancer at ages younger than 40, as well as an additional 235 (0.9%) for whom surgical treatment information was unavailable. The remaining 26,017 were included in our analyses, which amounts to about 9% of all women diagnosed with breast cancers in the US in 2013 [12].

The ages of included patients were limited to ≥40 because rates of mammography screening in the US are low before 40 and high afterward. For example, in the year 2010 Behavioral Risk Factor Surveillance System survey of the US population, receipt of at least 1 mammogram in the past 2 years was reported by approximately 8%, 9%, 46%, 77%, 81%, 83%, and 76% of women age 20, 30, 40, 50, 60, 70, and 80, respectively [13]. Screening participation increases suddenly at age 40 because several prominent US medical organizations recommend this as the preferred age to begin mammography screening [14]. Other studies of the US also report high mammography screening rates for ages ≥40, including older ages [15, 16].

The current study includes both screened women who were diagnosed with breast cancer and unscreened women who were diagnosed with breast cancer. This is because the SEER 9 registries do not record information on screening participation for individual patients. Nonetheless, the rate of screening participation is very high in the SEER 9 registry population as a whole. For example, data for 2008-2010 indicate that, of all women age ≥40 in the SEER 9 population, approximately 73% received at least 1 mammogram in the past 2 years [17]. We believe this high rate of screening participation makes the study population suitable for studying overdiagnosis and overtreatment, especially because the rate of screening participation was similar or lower in the screening arms of several of the randomized trials of mammography screening. For example, 74%, 68%, and 65% of the women assigned to screening arms of the Malmo I, UK Age, and New York HIP trials actually received their first screenings [18].

Some patients in our dataset had records for more than 1 breast cancer diagnosed in 2013 (*n*=1,049; 4.0%). For these patients, our analyses are of the surgical treatment and case characteristics in the registry records associated with the first of their year 2013 diagnoses. Before making this decision, we checked that only a negligible number of patients had different surgical treatments in the registry records associated with their first and later year 2013 diagnoses (*n*=68; 0.3%).

We analyzed SEER data on surgical treatments that were performed as part of first-course therapy [19, 20]. When the available case documentation did not provide enough information to determine whether therapy was first or later course, it was recorded in SEER as first course if given in the first year after diagnosis, and was considered to be later course if given in the second or later years after diagnosis [19].

All data used for this study are deidentified and publicly available from SEER using SEERStat software.

##### 2.2. Definitions

Total mastectomy was defined as simple mastectomy or modified radical mastectomy. Breast-conserving surgery was defined as lumpectomy, excisional biopsy, segmental/subtotal mastectomy, quadrantectomy, tylectomy, wedge resection, nipple resection, or partial mastectomy, not otherwise specified. SEER records the most extensive surgical procedure that was performed. In the overall cohort, mastectomy, breast-conserving surgery, other surgical therapies (including subcutaneous mastectomy), and no surgery of the primary site were performed in 33.8%, 56.4%, 1.3%, and 8.4% of included cases, respectively.

Breast cancer cases are defined as overdiagnosed if they were diagnosed because of screening, but if the cancer would not have been noticed or caused harm in the patient’s lifetime in the absence of screening. Since overdiagnosed cancers do not require treatment, any treatment provided for them is regarded as overtreatment.

##### 2.3. Estimation Approach

We sought to place bounds on how often mastectomy could plausibly be performed for overdiagnosed breast cancer. To obtain the bounds, three pieces of information were used: (A) a set of criteria that were used to rule out overdiagnosis in some cases, (B) an estimate of the proportion of breast cancer cases that are overdiagnosed, and (C) estimates of the probability of treatment by mastectomy for each case.

###### 2.3.1. Information A: Criteria Used to Rule out Overdiagnosis

We ruled out breast cancer cases from being overdiagnosed if they had any of the characteristics listed in Table 2. Because the characteristics reflect a behavior that is aggressive, advanced, and/or would quickly become clinically evident in the absence of screening, these presence of these characteristics indicates the breast cancer is highly unlikely to have been overdiagnosed.

For our bounds to be valid, we had to be especially careful that our criteria did not misclassify overdiagnoses as nonoverdiagnoses. As a consequence, some of the criteria in Table 2 may appear overly conservative. For example, because 2.0-3.9 cm tumors could be overdiagnoses in rare cases, we did not exclude them. Had they been excluded, our bounds might have been rendered invalid, especially since mastectomy becomes more common at larger sizes. Similarly, we did not rule out cases with 1 positive lymph node because they could occasionally be overdiagnoses with a false-positive lymph node, and mastectomy might be especially common for these cases. (False-positive lymph node biopsy findings have been reported [21–23], though the false-positive rate appears to be unknown.)

We have tried to be suitably conservative when selecting these criteria, but we realize that some will debate our choices. To address this, we conducted supplementary analyses in which we tried alternative criteria and examined how estimates of overtreatment by mastectomy were affected. For example, we tried ruling out overdiagnosis for cases with ≥1 positive lymph node and/or tumor sizes of ≥3.0 cm, and found that our bounds on use of mastectomy for overdiagnosed cancer changed by only a couple of percentage points. Accordingly, our judgments of how many lymph nodes and what tumor sizes fully rule out overdiagnosis did not have large consequences for our results. More details can be found in the supporting information (Table S2 and Figure S1).

###### 2.3.2. Information B: Estimates of the Proportion of Breast Cancers That Are Overdiagnosed

The amount of overdiagnosis that is occurring is not clear and, in the prior literature, estimates of overdiagnosis rates have ranged widely from <1% to >50%, changing greatly depending on study designs, settings, and measures of overdiagnosis [10, 24–28]. To account for this variation, we performed our analyses several times, using different estimated values for the proportion of breast cancers in the study population that were overdiagnosed. The range of investigated values was 0% to 37%. We chose this range based on the following considerations: In the SEER 9 cancer registries, mammography screening was rare during and before 1980. Since then, both screening rates and breast cancer incidence have increased [29]. Assuming that the incidence of nonoverdiagnosed breast cancer incidence has either been constant or increasing over 1980-2013, and that mammography screening is responsible for almost all overdiagnoses of breast cancer, then the rate of nonoverdiagnosed breast cancer cannot be substantively lower than the incidence rate observed in 1980, and the rate of overdiagnosis cannot be substantively higher than the overall increase in breast cancer incidence from 1980 to 2013. So, whatever it is, the true amount of overdiagnosis lies between these two values. Among women age ≥40 in the SEER 9 cancer registries, the age-standardized incidence of breast cancer was 230.1 per 100,000 in 1980 and 364.6 per 100,000 in 2013. Therefore, under the noted assumptions, at least 0% and at most 37% of breast cancers in the study population could be overdiagnosed (37% = 1 − 230.1/364.6).

###### 2.3.3. Information C: Estimates of the Probability of Treatment by Mastectomy

We used a regression analysis to estimate the probability of treatment with mastectomy according to the recorded characteristics of the cases in the study population at diagnosis. Thirty-three characteristics were included in our analysis, including various patient, disease, and regional attributes (Table S1).

If we had used only a couple of characteristics—say stage and grade—then determining the probability of mastectomy would not require regression. Instead, we would simply calculate the proportion of cases treated by mastectomy for each unique combination of stage and grade. (In other words, we would create a cross-table.) However, as the number of characteristic increases, the number of unique combinations that need to be considered becomes huge, making estimates of the proportion of cases treated by mastectomy unstable. To address this sparse-data problem, we used regression modeling to estimate the probabilities of treatment by mastectomy, instead of calculating these values directly in cross-tables. We performed the regression using a random forest model. This is a common, basic method from the machine learning literature that was selected because it offers reliable performance, is resilient to the curse of dimensionality, and does not generally overfit [30–32].

Using the randomforestSRC package [31, 33], a random forest model was trained with 2500 trees, the square root of the total number of variables as the number of variables tried per node split, Gini index splitting, a leaf size of 1, and a maximum of 25 random splits for multivalue variables. These hyperparameters were not tuned. The random forest was fit to cases diagnosed in 2013 (training set) and tested on cases diagnosed in 2012 (test set). For the year 2013 probabilities of mastectomy analyzed in this article, we used out-of-bag estimates to avoid overfitting. The calibration of the random forest was good for both the training and test set (Figure S2). In regard to accuracy and discriminative performance, Breir score values were 0.176 for 2012 and 0.175 for 2013, and c-statistic values (areas under the receiver operating characteristic curves) were 0.745 for 2012 and 0.742 for 2013. Because the bounds obtained in our analysis are dependent on the discriminative performance of the fitted model, we also performed sensitivity analyses in which investigated whether performance was substantially changed by fitting the model on half (random sample of 2013) and twice (years 2012 and 2013 together) as many records, and by using half and twice as many trees. Calibrations curves, Breier scores, and c-statistics values were similar to those reported above, as were the lower and upper bounds on the frequency of mastectomy for overdiagnosed cancer. These and all other statistical analyses were conducted in R (The R Foundation for Statistical Computing; Vienna, Austria).

###### 2.3.4. Estimating Overtreatment by Mastectomy

We obtained bounds on the frequency at which mastectomy is performed for overdiagnosed cancer by applying Information A, B, and C. The following steps were used: First, we excluded all cases that had characteristics ruling out overdiagnosis (applying Information A). Second, we considered that each remaining case belonged to one of two groups, the overdiagnosed group or the nonoverdiagnosed group, but that the membership of these groups was not observable. We assumed that the overdiagnosed group had a specific size (applying Information B). Third, we analyzed how the probability of treatment by mastectomy varied according to characteristics at diagnosis (applying Information C). To obtain a minimum plausible estimate (lower bound) of how often mastectomy was performed for overdiagnosed cancer, we filled up the overdiagnosed group with the cases that had the least probabilities of treatment by mastectomy. On the other hand, to obtain a maximum plausible estimate (upper bound), we filled up the overdiagnosed group with the cases that had the greatest probabilities of treatment by mastectomy.

For example, suppose that we rule out the cases that cannot be overdiagnoses and are left with 75% of the original study cohort. Suppose also that 30% of the entire cohort are overdiagnoses. Then, simple calculation shows that 40% of the remaining cases are overdiagnoses [40% = 30% / 75%]. We do not know which breast cancer cases belong to the 40% that are overdiagnoses, and this prevents us from calculating exactly how common it is for overdiagnosed cancers to be treated by mastectomy. However, we can still make progress based on a key observation: No one is able to identify overdiagnosed cases; therefore, the probability of treatment by mastectomy is the same for overdiagnosed and nonoverdiagnosed cases that share the same observed characteristics. So, we reason that the actual frequency of mastectomy-treated overdiagnoses cannot reasonably be less than it would be if the overdiagnosed cases were the cases that had characteristics associated with the lowest probability of treatment by mastectomy. Similarly, the actual frequency of mastectomy-treated overdiagnoses cannot reasonably be greater than it would be if the overdiagnosed cases were the cases that had characteristics associated with the greatest probability of treatment by mastectomy. In this way, we obtain minimum and maximum plausible estimates of the frequency at which mastectomy is performed for overdiagnosed breast cancer.

In more statistical detail, our approach is as follows: After excluding the cases that cannot be overdiagnoses (Information A), we are left with* n* cases, some of which are overdiagnosed and others of which are nonoverdiagnosed. Denote by* X* the 33 characteristics included in our regression analyses (Information C), and let the values of these characteristics for case* i* be* x*_{i}. Further, let* M *denote that mastectomy was performed and* V* denote that overdiagnosis occurred. We explain the method of obtaining bounds in the large-sample limit, which is a good approximation for the analysis in this paper because of the very large-sample size.

We are interested in estimating the proportion of the* n* cases in which mastectomy was performed for overdiagnosed cancer. This is,Currently, no one can identify cases that have been overdiagnosed. (Indeed, if overdiagnosed cases could be identified, they would not be treated, and there would be no need for our study.) For this reason, we make our key assumption: Conditional on the observed characteristics of a case at diagnosis, the probability of mastectomy would not be different if the case was overdiagnosed cancer or if it was nonoverdiagnosed cancer. The overdiagnosed and nonoverdiagnosed cases are then exchangeable conditional on observed characteristics.Plugging these results into our expression for the proportion of cases with mastectomy after overdiagnosis (Expr. (1)), we have In this expression, is estimated using a regression model (Information C), while is estimated by the proportion of all cases in the study cohort with . Only , the probability of overdiagnosis conditional on the observed characteristics, is unknown.

If we assume that the proportion of the* n *cases that are overdiagnosed takes a known value, say* q* (Information B), then this restricts the values that can take. By distributing the allowed values of in such a way that maximizes the value of Expr. (3), we obtain an upper bound on the frequency of mastectomy-treated overdiagnoses. Similarly, by distributing the allowed values of to minimize Expr. (3), we obtain a lower bound.

In practice, the upper bound is obtained simply by assigning a value of 1 to for the proportion* q *of cases for which is largest, and assigning a value of 0 otherwise. Similarly, lower bound is obtained by assigning 1 to for the proportion* q *of cases for which is smallest, and assigning 0 otherwise. In this way, we obtain bounds on the proportion of breast cancers cases in the study population that were overdiagnosed and overtreated by mastectomy.

Appendix S1 provides additional detail, including discussion of the key independence/exchangeability assumption and explanation of how our approach relates to other methods, such as propensity scores and regression standardization.

##### 2.4. Sensitivity Analysis for Omitted Variables

Though the main analysis of this study includes adjustment for 33 variables, several variables relevant to use of mastectomy were not recorded in our data source and therefore could not be adjusted for. For example, the data source did not record most cancer symptoms, breast cancer-related mutations (e.g.,* BRCA* mutations), family histories, screening histories, or whether mastectomy became necessary following breast-conserving surgery (e.g., due to recurrence or incomplete resection). Additionally, the data source often had missing values, and missingness could be informative of surgical choices in some cases. For these reasons, we performed an additional analysis in which we investigated the sensitivity of our results to any omitted variables and missing values that are relevant to use of mastectomy. A full explanation of the method is given in Appendix S1. In brief, the sensitivity analysis assumes that the predictions of mastectomy use are not systematically biased, but that omitted variables and missing values could increase their variance. The analysis is governed by a sensitivity parameter, which is the largest odds ratio (OR) by which omitted variables and missing data can change the probabilities of mastectomy from their estimated values. We used ORs that ranged from 1 to 25 to evaluate the maximum extent to which our results could be changed by omitted variables and missing data.

##### 2.5. Separate Analysis of* Ductal Carcinoma In Situ*

In a supplementary analysis, we repeated our evaluation of overtreatment by mastectomy for women diagnosed with ductal carcinoma* in situ* (DCIS), specifically. The supplemental analysis proceeded identically to the main analysis, with two exceptions: First, instead of using the rule-out criteria shown in Table 2, we ruled out all cases that were not DCIS. DCIS was defined as* in situ* breast cancer with ICD-O-3 code 8201, 8230, 8500-8507, or 8523 [34]. Second, the range of possible overdiagnosis rates was changed from 0-37% to 0-90% for DCIS cases, with the maximum of this range chosen based on the observation that DCIS incidence increased from 6.5 to 66.2 per 100,000 from 1980 to 2013 among women age ≥40 in SEER 9 (90% = 1 – 6.5/66.2; the calculation is analogous to that reported for overall breast cancer in Section 2.3).

##### 2.6. Interpretation of Bounds

Our bounds estimate the minimum and maximum plausible percentages of the study population who underwent mastectomy for overdiagnosed breast cancer. The study population is all women in the SEER 9 registries who were diagnosed with breast cancer (screen-detected or clinically detected) at age ≥40 in 2013.

When interpreting the bounds given in our results and figures, it is important to remember that they do not provide any information about the location of the true value within the bounds. They merely show the values that are plausible. For example, if our methods show 3%-15% of cases are overdiagnoses treated by mastectomy, then this does not provide any information about whether the true value is near the middle of this range, 9%, or nearer the edges. Further, the bounds do not tell us about the rate of overtreatment in years or areas other than those included in the study population. For example, our findings are for 2013, and the rate of mastectomy-treated overdiagnoses is likely somewhat different today. Finally, although the ranges tell us about the rate of mastectomy-treated overdiagnoses in the study population as a whole, they do not provide any information about the probability of overtreatment by mastectomy for individual patients. If we are considering an individual patient, then the probability that she received mastectomy for overdiagnosed cancer can be lower or higher than the range, depending on the characteristics of her case.

When reporting bounds, we rounded the percentages outwards to be conservative. For example, a bound of 5.6%-12.4% was rounded to 5%-13%.

#### 3. Results

##### 3.1. Overall Use of Mastectomy

Table 1 summarizes the characteristics of the study population: women diagnosed with breast cancer at age ≥40 in 2013 in the SEER 9 cancer registries. Overall, 33.8% of the 26,017 included patients received treatment by mastectomy. Larger tumor sizes and younger patient ages were associated with progressively higher rates of mastectomy (p < 0.0001 for each trend; *χ*^{2} test for trend). However, inspecting the percentage values shows that mastectomy was common for all categories investigated in the Table, including for all tumor sizes and all ages.