| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Original Article |
School of Finance and Applied Statistics, Australian National University, Canberra, Australian Capital Territory 0200, Australia
Correspondence: Address correspondence and reprint requests to: Michael A. Martin, BSc, PhD; E-mail: Michael.Martin{at}anu.edu.au
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
Methods: Data were sourced from linked administrative data from the Western Australian Department of Health Record Linkage Unit. The data consisted of linked records containing the diagnosis, subsequent hospital admission, and death records of about 3000 women diagnosed with cancer in Western Australia between 1 January 1995 and 31 December 1999. Cox proportional hazards regression was used to investigate survival outcomes of breast-conserving surgery compared with that of mastectomy, adjusting for tumor size, nodal status, and subject age.
Results: The hazard of death is reduced by a factor of about one half for subjects whose treatment was breast-conserving surgery over treatment by mastectomy. Furthermore, the hazard of death increases substantially for subjects with nodal involvement over subjects for whom there has been no identified spread to regional lymph nodes. Hazard of death increases as both age and tumor size increase.
Conclusions: Western Australian breast cancer patients treated with breast-conserving surgery have improved survival outcomes over those treated with mastectomy, after allowing for tumor size, patient age, and lymph node involvement.
Key Words: Breast cancer • Breast-conserving surgery • Mastectomy • Record linkage • Survival analysis
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
A number of clinical trials have concluded that BCS and mastectomy have similar survival rates.1–3 This evidence is supported by a recent pooled analysis of six major clinical trials comparing mastectomy and BCS, which concluded that the two treatments have comparable effects on mortality, even after long-term followup.4 Other population-based studies have also shown that patients receiving mastectomy and BCS have similar survival rates.5,6
A recent Western Australian population-based study found that patients who underwent BCS were at a decreased risk of dying compared with patients who underwent mastectomy.7 However, the main aim of that study was to investigate how hospital, social, and other demographic factors were associated with survival outcomes for surgically treated breast cancer, and the investigation did not adjust for tumor size and nodal status, both of which are known to influence survival from breast cancer.8–10 We specifically adjust for tumor size and nodal status because they may be intrinsically associated with mastectomy being the treatment of choice. By explicitly adjusting for these factors, we can assess the relevance of treatment choice to survival rates without its effect being confounded by these other covariates. Our model also includes subject age as a covariate. Our modeling also estimates the functional form in which the continuous covariates age and tumor size enter the model for survival rate using a generalized additive Poisson regression model.11
Our study is retrospective and, consequently, our results need to be interpreted carefully in that light. Nevertheless, our findings are based on a comprehensive data source that is unique in the Australian context and that includes important covariate information through the linkage of separate cancer registry, hospital admission, and death registry databases. While a prospective clinical trial would be an optimal way to assess the difference in survival between the two treatments, it is unlikely, particularly in the Australian context, that such a trial would result in as large a number of subjects as described in our retrospective study.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
There were 2788 subjects who satisfied these criteria; one subject was discarded because her age was recorded incorrectly as 200; this left a sample of size 2787 for analysis. In this analysis, treatment was defined as the last surgical treatment from the date of diagnosis through to the end of the study or death, whichever occurred first. The date of diagnosis was defined for this study as the time at which the subject entered the study as recorded on the WACR because none of the databases linked by the Western Australian unit specifies an explicit diagnosis date. It is not uncommon for people who initially choose BCS to be readmitted for mastectomy because initial surgery reveals that the cancer is more extensive than initially suspected. Thus, if a subject had a lumpectomy but was later readmitted for a mastectomy before the end of the study, their treatment would be defined for this analysis as mastectomy. Within our data set, 12.25% of cases were of this nature. The treatment (surgery) variable was assigned as follows: A value of 0 was assigned to those subjects who underwent mastectomy as the final surgical treatment for breast cancer before the end of study as indicated on the WACR; a value of 1 was assigned to those subjects who underwent BCS as the final surgical treatment for the breast cancer as indicated on the WACR. The definitions of BCS and mastectomy were based on relevant ICD-9 and ICD-10 procedure codes. For the purposes of this study, procedures categorized as BCS included local excision, segmental resection, and partial mastectomy (ICD-10 3034200, 3034201, 3034600, 3034601, 3035000, and 3035001), while for mastectomy the included procedures were simple mastectomy, extended simple mastectomy, subcutaneous mastectomy, and radical mastectomy (ICD-10 3033800-803, 3035300–303, 3035600–603, and 3035904–907). Note that our definition of mastectomy includes subcutaneous mastectomy, a procedure that leaves more breast tissue than simple mastectomy, and that patients falling into this category may therefore experience an increased rate of local recurrence. While this choice may have some impact on our findings, we believe our classification of surgical treatments to be reasonable. The other covariates included in the analysis were the tumor size (in this study recorded as diameter measured in millimeters rather than volume, which is another commonly used measure of tumor size), age of the subject in years, and nodal status. Nodal status was a categorical variable with the categories "no nodal involvement," "1 to 3 nodes," and "4 or more nodes." The choice of categories follows Foo et al.12 and Jayasinghe et al.13
For our study, mortality from all causes was considered rather than only mortality directly attributed to breast cancer because we wish to assess overall survival outcomes for breast cancer patients based on our data. In any case, the clear majority of reported causes of death in our data set was breast cancer, with an even larger proportion resulting if other cancers are included whose origin may have been the initial breast tumor. While it is acknowledged that more often older patients die from causes unrelated to breast cancer, the small proportion of deaths from other causes in our data set leads us to conclude that this kind of competing-risks argument does not make a significant impact on our findings.
Cox proportional hazards regression14 was used to investigate whether method of treatment (BCS or mastectomy) had differential effects on the survival rates of women with breast cancer, after adjusting for age, tumor size, and nodal status. The statistical software S-Plus15 was used to implement the Cox proportional hazard regression models. The issue of the functional form in which continuous covariates should appropriately enter the Cox model to obtain an adequate fit to the data is a potentially difficult one. This choice can be made using the Martingale residuals from a Cox model fit. Therneau and Grambsch16 showed that if the appropriate functional form for covariate j is f, then under certain assumptions a scatterplot smooth of the Martingale residuals from a null Cox model plotted against j will display the form of f. Unfortunately, this straightforward approach can fail when the covariates are correlated, as is likely in this circumstance. This problem can be addressed by using an approach described by Grambsch et al.,17 which models the expected hazard rate as a Poisson regression model and uses the generalized additive model (GAM) approach of Hastie and Tibshirani11 to fit this model nonparametrically, with the resultant functional form of f acting as a guide as to the way in which the associated covariate should enter the Cox model.
An initial Cox proportional hazards model was fitted to the data relating survival times to the variables of age, tumor size, surgery (0/1), and nodal involvement (3 categories) and all relevant two-way interactions. All interaction terms were statistically insignificant. Plots of scaled Schoenfeld residuals arising from this initial fit suggested that the proportional hazards assumption underlying the Cox model was not satisfied with respect to the age and tumor size variables, so the question of the functional form in which the continuous variables age and tumor size should appear in the model needed to be addressed. The Poisson regression approach of Grambsch et al.17 was used to decide the appropriate functional forms. Figure 1
displays the functional forms suggested by the Poisson regression GAM algorithm for the age and tumor size variables. For the purposes of modeling, simple parametric equivalents for these functional forms were the inclusion of quadratic terms in age and in tumor size into the Cox model. Fitting the suggested model and assessing the coefficients of squared age and squared tumor size for significance revealed that the quadratic terms in age and size were highly significant (P < 0.0001 and 0.005, respectively), suggesting that a Cox model that includes quadratic terms in age and tumor size was reasonable (see Table 5). A model including all two-way interactions was also used, with all interaction terms proving insignificant. Plots of scaled Schoenfeld residuals from this model versus time suggested no problems with the proportional hazards assumption with respect to the included covariates.
|
|
for the chosen model, is the inclusion of piecewise linear terms in age and tumor size within the Cox model. A model that included piecewise linear rather than quadratic terms in both age and tumor size was also fit as an alternative model to that chosen. All terms in the model were significant, with the results for surgery and nodal involvement not materially different from those reported for the chosen model. As part of our investigations we also considered a competing-risks Cox survival model that took into account causes of death other than breast cancer as a competing risk. For this model and our data, the competing risks were death from breast cancer and any other cause of death such as other cancers, heart disease, stroke, and diseases of the digestive system. It should be noted that for this data set there were only a very small number of deaths from causes other than breast cancer, so the resultant model fit needs to be interpreted with particular caution.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
gives a breakdown of survival outcome by surgical treatment, and we note in particular that the overall mortality rate of BCS patients was 4% compared with an overall mortality rate of 14% for mastectomy patients. Table 2 shows a comparison of tumor size and age characteristics for subjects broken down by survival outcome and the median followup time in days for each group. Table 3 shows a similar summary of tumor size and age characteristics broken down by surgical treatment. Table 4 shows a breakdown of survival outcomes by type of surgical treatment for the 394 patients who remained in the study for longer than five years. Our analysis considers overall mortality, but of the deaths in our study, the overwhelming majority, about 87% , were attributable to breast cancer, while an additional 3% were from other cancers which may have originated from the initial breast cancer.
|
|
|
|
shows that all the coefficients in the final fitted Cox proportional hazards model were statistically significant. The "exp(coef)" column of Table 5 details for the categorical covariates the ratio of the hazard for a subject with each characteristic (coded as 1) to the hazard for a subject without the respective characteristic (coded as 0), holding all other characteristics fixed, with values for nodal status to be compared with the base state "no nodal involvement." Notably, the hazard of death was reduced by a factor of about one half (55.88% ) for subjects whose treatment was BCS over those whose treatment was mastectomy, all other factors held constant. Furthermore, the hazard of death is roughly twice as high for subjects with 1–3 positive nodes than for subjects for whom there had been no identified spread to regional lymph nodes, and about 4.5 times greater for subjects with 4 or more positive nodes than for those without nodal involvement. As might be anticipated, hazard of death generally increases as both age and tumor size increase. The fitted relationship with age shows a shallow decline in hazard until about age 45 whereupon the hazard begins to rise. The slightly increased hazard for younger patients has been commented on in the literature7 and may be a result of breast cancer in younger women sometimes presenting as more aggressive than that in older patients. Nevertheless, the shape of the fitted hazard with respect to age is fairly shallow until about age 45, then it grows slowly until it increases more rapidly for older patients. Of course, older women have the additional risk of death from diseases unrelated to breast cancer, which may partly account for the increased hazard in older patients. The fitted relationship with tumor size shows a rapidly increasing hazard as size grows to about 50 mm, followed by a fairly stable hazard (holding other variables fixed) at higher tumor sizes. While this result may seem counterintuitive, it must be remembered that tumor size is positively correlated with the extent of nodal involvement, so its effect on survival is difficult to interpret on its own without considering nodal involvement as well. Age, by comparison, has very little correlation with either tumor size or nodal involvement, so its effect is more easily assessed in isolation.
As a means of visualizing our results, survival curves (based on the fitted Cox proportional hazards model) for subjects of median age and tumor size but with varying levels of the categorical factors for treatment option and nodal involvement are plotted in Figure 2. The survival curves illustrate better outcomes for such a subject treated via BCS regardless of nodal involvement, although clearly patients with no nodal involvement have better survival outcomes in general than those having nodal involvement. Average age/tumor size subjects treated via BCS who have either no nodal involvement or 1–3 positive nodes experienced favorable survival probability over all such patients treated via mastectomy, with mastectomy patients with 1–4 positive nodes having a relatively poor survival experience.
|
. With respect to causes of death other than breast cancer, an even further reduced relative hazard was enjoyed by BCS over mastectomy at around 15% . However, it should be noted that this latter result needs to be interpreted particularly cautiously because of the very small number of patients in the study whose cause of death was not breast cancer (only around 13% ). The further reduction in hazard for BCS patients with respect to other causes of death may be a result of the tendency for older patients to be those who are both more likely to be treated via mastectomy and to die from causes other than breast cancer. Unfortunately, the small number of such patients involved in this study makes it very difficult to separate these effects satisfactorily.
Also of interest is consideration of the effect of "salvage mastectomy" cases (i.e., patients who initially have BCS as their surgical treatment but who subsequently undergo mastectomy) on the analysis. For the data studied here, 12.25% of cases (342 of 2787) fell into this category. The presence of such cases has the potential to bias our results because these patients may have significantly poorer prognoses than other patients. As a sensitivity check on our initial results which included all reported cases, we refit our original Cox survival model to a reduced data set that included only initial BCS or mastectomy cases (i.e., excluding salvage mastectomy cases). The results of this model fit on the reduced data are presented in Table 6. These results reveal that when salvage mastectomy patients are excluded, BCS is predicted to reduce patients’ hazard slightly more than in the full-data case (to 40% of the baseline hazard compared with 44% for the full data), as might be anticipated. Nevertheless, the extent of this difference does not appear to be of practical significance and the results of the study do not change qualitatively.
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
The goal of our analysis has been the assessment of survival outcomes associated with treatment choice, adjusting as far as possible for indicators of disease progression such as tumor size and lymph node involvement and for patient age. Our results suggest that survival prospects for subjects treated with BCS are generally favorable compared with those for subjects treated with mastectomy. Of course, it is not possible in such a study to separate out the effects of the factors relevant for predicting survival outcomes beause those factors are invariably correlated with one another and with the survival response. For example, tumor size is correlated with the treatment choice intrinsically, so the effects of those variables on survival are confounded in a model that includes both variables. Generally, BCS is a suitable treatment for smaller tumors and these data suggest that it offers a better chance of survival. Patients who have larger tumors are more likely to have a mastectomy. They are also more likely to have more developed forms of cancer and a worse prognosis than patients with smaller, less developed breast cancers. As a result, the effects of tumor size and surgery are not readily separable because BCS is more commonly used for smaller tumors (median tumor size for BCS patients was 15 mm in our data compared with 25 mm for mastectomy patients). Nevertheless, our modeling suggests the usefulness of treatment option as a covariate in predicting survival despite the presence in the model of other variables related to disease progression. Thus, to an extent our model does suggest more favorable survival outcomes for patients treated by BCS than by mastectomy. In our modeling, tumor size, age, and nodal involvement are all found significant, along with the surgical treatment option. Older women, women with positive lymph nodes, and women with large tumors experience a higher rate of mortality than younger women, women with negative lymph nodes, and women with smaller tumors, respectively. We note, however, that women who choose BCS generally have less developed forms of breast cancer and this characteristic on its own is associated with a better chance of survival. We also note that while information on tumor size, or a prior indication thereof, may be relevant in assisting patients decide which surgical treatment to use, nodal status is not useful because the decision to undergo mastectomy after local excision is not made on the basis of whether nodes are involved.
There are several plausible explanations for the differential survival prognoses observed between BCS and mastectomy patients in our study. For example, the common use of radiation therapy following BCS is likely to result in improved prognosis over procedures that are not followed by such adjuvant therapy. Unfortunately, our data did not contain sufficient relevant information about the prevalence of radiation therapy in the two treatment groups to allow us to include such information in our modeling. Also, the added trauma of mastectomy over BCS is likely to have a negative impact on immune function with a subsequent effect on survival outcomes.
Post-treatment outcomes other than survival are also relevant issues for patients considering breast cancer treatment options. Breast-conserving treatment, while associated with improved survival outcomes in this study, can come at the cost of increased long-term morbidity associated with pain, cosmetic issues such as distortion, and psychological issues related to the potential for disease recurrence.18 Of course, the psychological impact of mastectomy, e.g., relating to body image, is another feature that may impact patient choice. Quality of life issues, along with survival, remain important determinants driving treatment choices for breast cancer patients.
Limitations of the present study must also be acknowledged. First, while our study involved a large number of patients and considered followup over a reasonable timeframe, the WACR database extends considerably far back into the past but does not include information on tumor size, a variable crucial to our analysis, before 1995. Although the length of followup dictated by our available data was reasonable, according to the National Surgical Adjuvant Breast and Bowel Project, a substantial proportion of events occur even after five years, so our results need to be interpreted accordingly. Also, our study did not include a number of covariates also related to survival outcomes such as the social and demographic variables included in a previous study. A larger study that incorporates a larger set of variables, both clinical and socioeconomic, might be useful in assessing the combined impact of these types of factors in assessing survival outcomes. Also, while the Western Australian database is unique in the Australian context in linking critical data from several sources and is an extraordinarily rich and comprehensive resource, it still lacks information on certain variables (e.g., tumor grade, more information on the use of adjuvant treatments such as radiation therapy) that could be usefully considered when modeling breast cancer survival.
Despite these limitations, our study has quantified the relationship between survival and various characteristics of breast cancer patients, including tumor size, age, nodal status, and the type of surgical treatment used. Our results suggest that the type of surgical treatment has an impact on survival outcomes, even after taking into account other factors such as tumor size, with BCS patients faring better than mastectomy patients within the context of our study.
Received for publication July 8, 2006. Accepted for publication July 13, 2006.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
 |
|
 |
|
  B. Thompson, P. Baade, M. Coory, P. Carriere, and L. Fritschi Patterns of Surgical Treatment for Women Diagnosed with Early Breast Cancer in Queensland Ann. Surg. Oncol., February 1, 2008; 15(2): 443 - 451. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K.-D. Yu, G.-H. Di, J. Wu, J.-S. Lu, K.-W. Shen, Z.-Z. Shen, and Z.-M. Shao Development and trends of surgical modalities for breast cancer in china: a review of 16-year data. Ann. Surg. Oncol., September 1, 2007; 14(9): 2502 - 2509. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
