- © 2001 by American Society of Clinical Oncology
Omission of Histologic Grading From Clinical Decision Making May Result in Overuse of Adjuvant Therapies in Breast Cancer: Results From a Nationwide Study
- J. Lundin,
- M. Lundin,
- K. Holli,
- V. Kataja,
- L. Elomaa,
- L. Pylkkänen,
- T. Turpeenniemi-Hujanen and
- H. Joensuu
- From the Departments of Oncology, University of Helsinki, Helsinki; University of Tampere, Tampere; University of Kuopio, Kuopio; University of Turku, Turku; and University of Oulu, Oulu, Finland.
- Address reprint requests to Johan Lundin, MD, PhD, Helsinki University Central Hospital Clinical Research Institute, Haartmaninkatu 4, PO Box 105, FIN-00290 Helsinki, Finland; email johan.lundin{at}helsinki.fi
Abstract
PURPOSE: To investigate the influence of routinely performed histologic grading on breast cancer outcome prediction and patient selection for adjuvant therapy.
PATIENTS AND METHODS: The analysis is based on a cohort of 2,842 women diagnosed with breast cancer and comprising 91% of all breast cancers diagnosed in five defined geographical regions in Finland in 1991 through 1992. Data on clinicopathologic factors and follow-up were collected from hospital case records and national registries. Histologic grade assessed at diagnosis and other clinicopathologic data were available for 1,554 operable unilateral invasive carcinomas. The relative value of grade with respect to competing prognostic factors was estimated with the Cox proportional hazards model and logistic regression. Interactions and nonlinearity of factors were accounted for by using an artificial neural network.
RESULTS: Histologic grade was correlated strongly with survival in the entire series and in all subgroups studied. Women with well-differentiated node-negative cancer had a 97% 5-year distant disease-free survival rate as compared with 78% for women with poorly differentiated cancer. Grade was an independent prognostic factor in multivariate models and increased the predictive accuracy of a neural network model. Inclusion of grade data in a Cox multivariate model based on tumor size and hormone receptor status in node-negative cancer increased the proportion of patients with 5% or less risk for distant recurrence at 5 years from 15% to 54%.
CONCLUSION: Even when assessed by pathologists who have no special training in breast cancer pathology, histologic grade has substantial and independent prognostic value in breast cancer. Omission of grading from clinical decision making may result in considerable overuse of adjuvant therapies.
THE ASSOCIATION between the histopathologic features of breast cancer and survival has been evaluated in numerous studies since the beginning of the last century.1 There is currently a general consensus that patients with well-differentiated breast cancers have a better prognosis than those with low-grade cancers, but the clinical value of histologic grading in the identification of patients with low versus high risk of cancer recurrence is still disputed. Histologic grade may or may not be used as a criterion for selection of patients for adjuvant therapy in the clinical decision-making process. Histologic grading of breast cancer is performed by combining cell morphology (nuclear pleomorphism), tissue architecture (tubule formation), and assessment of the cell proliferation rate (mitotic count), but the prognostic value of the combination of these three features is still being discussed.2,3 Histologic grading has been considered as too subjective to be used clinically, and grading may be associated with lack of reproducibility even when performed by experienced pathologists.4-7 On the other hand, numerous studies have shown a significant association between histologic grade and survival in breast cancer, and there is no doubt that grading is simple, quick, and economical to perform.
The crucial question is whether grade adds independent prognostic information when used in combination with other prognostic variables in breast cancer. Histologic grade has been entered into multivariate models in numerous studies, but the conclusions have been divergent, and histologic grade has or has not been found to have independent influence on survival.3,8-13 The relative importance of histologic grade in multivariate analyses may depend on many factors, such as the selection process of breast cancers for the series being analyzed, the quality of assessment and subclassification of the competing factors included in the multivariate model, the end points chosen, and the personal skill of the pathologist responsible for histologic grading.
A strong argument for the use of histologic grade in identification of patients with a high risk for breast cancer recurrence would be if grade could be shown to have independent influence on outcome in large and unselected series, where grading has been carried out by several pathologists and in the routine clinical setting. Somewhat surprisingly, unselected nationwide series in which the clinical value of routine histologic grading of breast cancer has been evaluated are lacking. In the present study we assessed the prognostic value of histologic grade in a comprehensive cohort of breast cancer patients diagnosed in Finland in 1991 and 1992. Histologic grading was performed by numerous pathologists with varying interest and experience in breast cancer pathology, and the results might thus be considered representative of the usefulness of histologic grade as a prognostic factor in the routine clinical practice. We also investigated the efficacy of histologic grading in the selection of patients for adjuvant therapy using 3%, 5%, and 10% risk levels for distant relapse in two different multivariate models and in a three-layer artificial neural network.
PATIENTS AND METHODS
Patients
Five well-defined geographical regions comprising approximately 50% of the Finnish population were selected for the study (Fig 1). Using the files of the Finnish Cancer Registry, we then identified women diagnosed with breast cancer within these regions in 1991 and 1992. Hospitals, practicing physicians, and pathologic and hematologic laboratories are requested to report to the Finnish Cancer Registry all cases of cancer that come to their attention. In addition, all death certificates in which cancer is mentioned are transferred from the files of the Statistics Finland to the Cancer Registry each year. A computer search of the files of the Finnish Cancer Registry identified a total of 2,930 breast cancer cases diagnosed in these regions in 1991 to 1992, which constitutes 53% of all 5,551 breast cancers diagnosed in Finland during this time period.
Fig 1. The proportion of patients included in the study of all patients diagnosed with breast cancer in 1991 through 1992 within the chosen regions, according to data from the Finnish Cancer Registry.
Using structured data collection forms, the individual clinical data of the patients were extracted from the hospital records, which were reviewed. An effort was made to obtain detailed clinical information on 50 characteristics, including the histologic type and grade of breast cancer, the number of metastatic and nonmetastatic nodes, primary tumor size, tumor estrogen receptor (ER) and progesterone receptor (PgR) content, treatment details, and follow-up data. In addition to the hospital case records, relapse and survival data were extracted from the files of the Finnish Cancer Registry and the hospital registries, when available. For the study inclusion we required the following minimum data to be available: date of breast cancer diagnosis; age at diagnosis; information on other malignancies than breast cancer in history, including data on bilateral breast cancer; postsurgical size of the primary tumor and axillary nodal status; follow-up data; and vital status data at the end of follow-up. This minimum information was available in 2,656 (91%) of the 2,930 eligible patients. The percentage of the patients with the minimum information available varied from 79% in southwestern to 97% in eastern Finland (Fig 1). In addition to the 2,656 patients with the minimum clinicopathologic information available, we included 186 further patients who fulfilled the minimum clinicopathologic information and were diagnosed with breast cancer within the same geographic regions during the same time interval, but who were not identified in the original computer search because the place of residence was outside the specified regions. Thus the total number of patients entered into the database was 2,842. For the present study, we excluded the following patients from this database: patients with lobular carcinoma-in-situ (n = 17), ductal carcinoma-in-situ (n = 187), those with distant metastasis at the time of diagnosis (M1, n = 133), women with synchronous or metachronous bilateral breast cancer (n = 281), patients with malignancy other than breast cancer in history (except for basal cell carcinoma or cervical carcinoma-in-situ, n = 201), and women who did not undergo breast surgery (n = 42) and those for whom histologic grading had not been carried out at the time of the diagnosis (n = 892), leaving 1,554 patients for the final analysis.
A total of 1,041 (67%) of the patients were treated with mastectomy and axillary node dissection, 446 (29%) with lumpectomy and axillary node dissection, and 54 (4%) with simple mastectomy or lumpectomy. Nine hundred thirty-three (61%) women received postoperative radiotherapy. The majority of node-positive patients (n = 505, 92%) received adjuvant endocrine therapy or cancer chemotherapy, and 83 (9%) of the node-negative patients were treated with adjuvant therapy. Adjuvant therapy usually consisted of antiestrogen therapy (tamoxifen 20 mg or toremifene 40 mg daily) in postmenopausal women and combination chemotherapy in premenopausal women. The most common chemotherapy regimen used was intravenous cyclophosphamide, methotrexate, and fluorouracil for six to nine cycles. The median follow-up time for the unrelapsed patients was 6.8 years (range, 5.1 to 7.8 years).
Prognostic Factors
Histologic typing and evaluation of grade components (mitotic count, nuclear pleomorphism, and tubule formation) were performed at the time of diagnosis by more than 50 pathologists according to the World Health Organization classification.14 For the analyses, the tumors were classified into three histologic types: ductal carcinoma (not otherwise specified, includes apocrine, mixed mucinous, and atypical medullary types), lobular carcinoma (infiltrating lobular carcinoma with variants), and the special histologic types (includes tubular, medullary, cribriform, papillary, and pure mucinous carcinomas). The proportion of well (grade 1), moderately (grade 2), and poorly (grade 3) differentiated cancers varied somewhat between the five study regions (grade 1 range, 18% to 33%; grade 2 range, 40% to 55%; grade 3 range, 19% to 36%).
The longest diameter of the tumor was extracted from the surgery or pathology report. ER and PgR status was determined in the majority of cases by immunohistochemistry and was classified either as positive or negative.
Statistical Analysis
The χ2 test was used to test for associations between factors. Life-tables were calculated according to the Kaplan-Meier method. Distant disease-free survival (DDFS) was calculated from the date of the diagnosis to the occurrence of metastases outside the locoregional area or death from breast cancer, whichever came first. Breast cancer-specific overall survival (OS) was calculated from the date of the diagnosis to death from breast cancer, and patients who died from intercurrent causes were censored. Survival curves were compared with the log-rank test. Multivariate survival analyses were performed with the Cox proportional hazards model and logistic regression, respectively, entering the following covariates: grade (well differentiated, 1; moderately differentiated, 2; poorly differentiated, 3), tumor size in centimeters, the number of metastatic lymph nodes, ER and PgR status (positive or borderline, 0; negative, 1), histologic type (lobular or special, 0; ductal, 1), age at diagnosis in quartiles (23 to 49 years, 1; 50 to 58 years, 2; 59 to 70 years, 3; 71 to 96 years, 4). A P value of .05 was adopted as the limit for inclusion of a covariate. Individual prognostic scores were calculated using the coefficients derived from the Cox and logistic regression models. The area under the receiver operating characteristic (ROC) curve (AUC) was used as a measure of accuracy of the prognostic scores in separating disease-free survivors from those who experienced relapse.15 The statistical comparison of the areas under two ROC curves was performed according to Hanley and McNeil.16
Artificial Neural Network Analysis
A three-layer artificial neural network model was constructed as earlier described17 with a commercially available computer program using a modified cascade method together with an adaptive gradient learning rule (NeuralWorks Predict/NeuralSIM, NeuralWare, Inc, Pittsburgh, PA). The cascade mode of construction entails adding hidden nodes, one or more than one at a time, and always connecting all of the previous nodes to the current node.18 Direct connections between input and output nodes were also allowed. Network output ranged from 0 to 1. The data set used in the neural network was the same as in the Cox and logistic regression models and only patients with complete data on the selected variables were included in these analyses. From the entire data set, 30% of the patients were randomly chosen for the test set. During training the model was scored on this test set to choose between different candidate hidden nodes and to avoid overtraining. Overtraining of the network was also reduced using a weight decay method. Several candidate networks were trained, and the network with the highest accuracy on the test set at 5 years was chosen for final calculation of predictive accuracy.
RESULTS
There were 406 well-differentiated (26%; grade 1), 734 moderately (47%; grade 2), and 414 poorly (27%; grade 3) differentiated cancers in the patient series (Table 1). A poor grade of differentiation was significantly associated with a high number of metastatic axillary nodes, a large tumor size, detection of cancer outside a screening program, ductal histologic type, negative ER and PgR status, and the youngest and oldest age quartiles (data not shown). There was a highly significant difference in both DDFS and breast cancer–specific OS between patients according to histologic grade (Figs 2A and 2B). The 5-year DDFS of patients with grade 1 cancer was 95% as compared with 77% and 66% among those with grade 2 and grade 3 cancer, respectively (P < .0001). Grade was significantly associated with DDFS in all subgroups studied: in node-negative and node-positive breast cancer (Figs 3A and 3B, P < .0001 and P < .0001, respectively), in cancers with a primary tumor ≤ 2 cm (P < .0001), and in those more than 2 cm in diameter (P < .0001), in ductal (P < .0001) and lobular carcinoma (P = .0047), and in all age quartiles (P = .0004 for 23 to 49 years, P < .0001 for 50 to 58 years, P < .0001 for 59 to 70 years, and P < .0001 for 71 to 96 years at diagnosis), and it was significantly associated with survival in all five geographical regions studied (P < .0001, P = .0001, P = .0038, P < .0001, and P < .0001) (Table 2).
5-Year DDFS According to Clinicopathologic Characteristics in 1,554 Patients With Breast Cancer
Fig 2. DDFS (upper panel) and OS (lower panel) of 1,554 patients with breast cancer according to histologic grade. —, well-differentiated cancers (n = 406); – –, moderately differentiated cancers (n = 734); ---, poorly differentiated cancers, (n = 414). P < .0001 for both analyses (log-rank test for trend).
Fig 3. DDFS of 950 women with node-negative (upper panel) and 553 with node-positive (lower panel) breast cancer according to histologic grade. —, well-differentiated cancers; – –, moderately differentiated cancers; ---, poorly differentiated cancers. P < .0001 for both analyses (log-rank test for trend).
5-Year DDFS According to Grade in Different Subgroups of Patients With Breast Cancer
In a Cox multivariate model, the number of positive axillary lymph nodes, tumor size, histologic grade, and the PgR status were independent prognostic factors for DDFS, whereas age, histologic type, and the ER status were not retained in the model (Table 3a). Inclusion of the patients with missing data on grade in a multivariable model resulted only in small changes in the relative hazards of the other variables. Neither the set of the other variables that were found to be independent prognostic factors nor their order of significance was changed. If patients with missing data on hormone-receptor status were included in the model, the value of grade is unlikely to decrease, because grade is a highly significant prognostic factor also in this subgroup (P < .0001).
Cox Proportional Hazards Regression of Independent Prognostic Factors for DDFS
Similarly, grade was an independent prognostic factor for DDFS when tested in a Cox multivariate model among node-negative breast cancer patients who had not been treated with adjuvant therapy. In this subgroup, too, the primary tumor size, the PgR status and grade were independent prognostic factors, whereas age, histologic type, and ER status did not add significant prognostic information (Table 3b).
A prognostic score was calculated for each patient using the coefficients obtained from the Cox and logistic regression multivariate models. The predictive accuracy for distant relapse, measured using the AUC at 5-years of follow-up, was significantly higher for both models when histologic grade was included in the analyses as compared with the same analyses without grade (Table 4). The result remained similar regardless of whether all patients or only those with node-negative cancer were included in the analyses. In the calculations listed in Table 4, an arbitrarily chosen cutoff value of 5% or less risk for distant relapse at 5 years of follow-up was used to define a low-risk group of patients who might be spared of adjuvant therapy. When a combination of tumor size and the PgR status (the ER status was less effective) was entered into the Cox multivariate model and prognostic scores calculated for patients with node-negative breast cancer, only 16% of the patients were defined as having a low risk for cancer recurrence. When histologic grade was added to the same model, the proportion of low-risk patients increased 3.7-fold up to 54%, demonstrating the substantial increase in the predictive accuracy of the model after inclusion of the histologic grade as prognostic parameter. The results remained essentially similar when logistic regression was used instead of Cox multivariate analysis (Table 4).
Accuracy of Different Prognostic Models in Predicting 5-Year DDFS in Patients With Breast Cancer
Because some factors, such as age at diagnosis, did not have a linear influence on prognosis (Table 1), we also used an artificial neural network analysis to investigate whether histologic grade could add significant prognostic information when nonlinearities and interactions were accounted for (Table 4). A trained artificial neural network using data on age at diagnosis, tumor size, PgR status, and histologic type was clearly more effective in defining women with a low risk of recurrence than Cox proportional hazard model and logistic regression using data only on tumor size and PgR status (Table 4). Similarly to multivariate Cox proportional hazard and logistic regression models, adding data on histologic grade clearly improved the predictive accuracy of the neural network. For example, in the subset of women who had not been treated with any kind of adjuvant therapy and who had node-negative breast cancer, adding data on histologic grade to the artificial neural network model increased the proportion of women who were predicted to have 5% or lower risk for recurrence from 41% to 64%.
Because the 5% risk for cancer recurrence at 5 years of follow-up is an arbitrary cutoff value, we compared the patients’ individual risks for recurrence also at 3% and 10% cut-offs in a Cox proportional hazards model, logistic regression model, and in artificial neural network with and without data on histologic grade. The results are shown graphically in Fig 4. Regardless of the mathematical approach used and the cutoff level chosen, addition of data regarding histologic grade increased the number of patients with node-negative breast cancer who could be identified as having a low risk of cancer recurrence at 5 years of follow-up.
Fig 4. Addition of data on histologic grade to three different prognostic models (—) increased the number of patients identified as having a ≤ 3%, 5%, or 10% risk for breast cancer recurrence compared with the respective models without grade (– –).
DISCUSSION
The present data provide strong evidence that histologic grade is a valuable parameter in predicting outcome of breast cancer and in identification of patients who have a low risk for breast cancer recurrence and might thus be spared from receiving adjuvant therapy. Although several large studies have emphasized the prognostic value of grade,9,11,19-21 other investigators disagree.10,22 Criticism has, however, mainly focused on problems with interobserver reproducibility.5-7 In the present nationwide series, histologic grading was performed by dozens of pathologists as a part of their routine daily work, and yet, grade correlated well with survival in the entire series, in all subgroups investigated, and within all geographical districts when analyzed separately. Moreover, histologic grade was significantly associated with survival whether all node- negative patients or only those who had received no adjuvant therapy were included in the analyses. Grade added independent prognostic value to commonly used prognostic factors in the two multivariate models used (Cox proportional hazards model and logistic regression), and it increased the predictive accuracy of a trained artificial neural network.
The predictive efficacy of routine histologic grade in outcome prediction was surprisingly good. In the entire series, women with well-differentiated breast cancer had as high as 95% 5-year DDFS as compared with only 66% among those with poorly differentiated breast cancer, and in the subgroup of node-negative breast cancer, the corresponding figures were 97% and 78%, respectively. The influence of adding data on histologic grade to the patient selection for adjuvant therapy was perhaps even more striking. In node-negative breast cancer, selection of patients for adjuvant therapy may be performed according to tumor size alone or by a combination of two factors, such as tumor size and the hormone-receptor status. In the present series when individual risks for breast cancer recurrence at 5 years of follow-up were calculated based on the Cox multivariate proportional hazards model, and when tumor size and the PgR status were used as the selection parameters, only 97 (15%) of a cohort of 653 patients with node-negative breast cancer were identified as low-risk patients who might be spared from adjuvant therapy, but when data pertaining to histologic grade were added to the same model, the number of low-risk patients increased to 354 (54%). Hence, as many as 257 additional patients (39%) were identified as having 5% or less risk for breast cancer recurrence. Histologic grading is associated with minimal costs, and, therefore, its use in clinical decision making may result in substantial savings and a greatly increased proportion of patients who can be spared treatment toxicity.
A strength of the present study is that all prognostic parameters were determined at the time of the diagnosis and without knowledge of the final outcome. The study covers more than one half of all breast cancer patients diagnosed in a single country, and the coverage within the five geographical study regions chosen was on average more than 90%, making the risk of a severe selection bias minimal as compared with single-center studies or studies that are based on patients selected for a clinical trial. Another strength of the study is that the treatment of the patients was relatively homogeneous, and only relatively few patients with node-negative disease received adjuvant therapy. Future studies may have to deal with patient series in which also node-negative patients have been treated with various adjuvant therapies.
A weakness of the study is that a considerable number (n = 892) of cancers in the series of 2,842 patients were not graded. This was mostly due to differences in the willingness of individual pathologists to grade breast cancer, and some pathologists preferred the use of other analyses considered by them as more objective, such as determination of the S-phase fraction by DNA flow cytometry or Ki-67 immunostaining. We made no attempt to perform histologic grading in retrospect for those cases for which information on grade was missing, because now clinical follow-up data are available, and grading would no longer be performed as a part of routine pathologic assessment by a number of pathologists. However, we believe that the present series is representative and suitable for analysis of the prognostic value of histologic grade in a routine clinical setting. The results on the competing prognostic factors remained essentially similar irrespective of whether the patients with missing data on histologic grade were included in the analysis.
We included age and histologic type in the artificial neural network analysis, although they were not independent prognostic factors in the Cox proportional hazards model or in the logistic regression. This was because we wanted to allow for possible interactions and nonlinear effects of these variables. It is clear from Table 1 that the effect of age at diagnosis on prognosis is not linear, and the youngest and the oldest quartiles are associated with the worst outcome. The conventional linear multivariate models cannot account for such nonlinearity without complex transformations and may underestimate the effect of nonlinear covariates. Even when the nonlinearity of competing prognostic factors was accounted for, histologic grade retained its independent prognostic value in prediction of 5-year DDFS. It is worth noting than when histologic grade was entered into a neural network model together with tumor size, PgR status, age at diagnosis, and histologic type, as many as 64% of node-negative patients who had received no adjuvant therapy were identified as having 5% or less 5-year risk for distant relapse. This suggests that by using a combination of prognostic factors and a trained artificial neural network, the majority of node-negative breast cancer patients may be spared of adjuvant therapy, provided that the 5% or less risk level for distant relapse within 5 years after the diagnosis can be accepted.
It is important to evaluate the prognostic value of histologic grade in relation to novel molecular-genetic markers in future studies and to validate its independent prognostic value in multivariate models and in large unselected series together with such factors. For this purpose we are currently in the process of collecting archival tumor specimens from the respective pathology laboratories that participated in the present study to be analyzed for a panel of proposed biologic prognostic factors. It remains to be seen whether some of the novel biologic factors can replace histologic grade as a simple and powerful prognostic variable and produce superior results when performed in the routine clinical setting. A recent study indicates that grade may yet be a more powerful predictor than intensively studied molecular-genetic factors such as p53 and c-erbB-2.23
We conclude that histologic grade, when performed in the routine clinical setting and by a number of pathologists with varying interests and experience in breast cancer pathology, is a valuable and independent prognostic factor in breast cancer. Estimation of breast cancer recurrence risk and selection of node-negative breast cancer patients for adjuvant therapy on the basis of tumor size and hormone receptor status alone without data on histologic grade leads to substantial overuse of adjuvant therapies with associated toxicities and costs. Histologic grade should not be omitted from the clinical decision-making process unless it can be replaced by more effective factors for outcome prediction that work reliably in the routine clinical setting.
Acknowledgments
This study was supported by grants from Finska Läkaresällskapet, the Käthe och Olof Rundberg foundation, and K. Albin Johansson foundation.
- Received March 17, 2000.
- Accepted July 25, 2000.
