The authors have declared that no competing interests exist.
Conceived and designed the experiments: MH HMV. Performed the experiments: NS. Analyzed the data: NS MH NBP HMV LWK. Contributed reagents/materials/analysis tools: CHY NBP NAT SCL SEL LML NS. Wrote the paper: NS MH CHY NBP NAT SCL SEL LML LWK HMV.
Lymph node ratio (LNR, i.e. the ratio of the number of positive nodes to the total number of nodes excised) is reported to be superior to the absolute number of nodes involved (pN stage) in classifying patients at high versus low risk of death following breast cancer. The added prognostic value of LNR over pN in addition to other prognostic factors has never been assessed.
All patients diagnosed with lymph node positive, non-metastatic invasive breast cancer at the National University Hospital (Singapore) and University of Malaya Medical Center (Kuala Lumpur) between 1990–2007 were included (n = 1589). Overall survival of the patients was estimated by the Kaplan Meier method for LNR [categorized as low (>0 and <0.2), intermediate (0.2–0.65) and high (>0.65–1)] and pN staging [pN1, pN2 and pN3]. Adjusted overall relative mortality risks associated with LNR and pN were calculated by Cox regression. The added prognostic value of LNR over pN was evaluated by comparing the discriminating capacity (as indicated by the c statistic) of two multivariate models, one including pN and one including LNR.
LNR was superior to pN in categorizing mortality risks for women ≥60 years, those with ER negative or grade 3 tumors. In combination with other factors (i.e. age, treatment, grade, tumor size and receptor status), substituting pN by LNR did not result in better discrimination of women at high versus low risk of death, neither for the entire cohort (c statistic 0.72 [0.70–0.75] and 0.73 [0.71–0.76] respectively for pN versus LNR), nor for the subgroups mentioned above.
In combination with other prognosticators, substitution of pN by LNR did not provide any added prognostic value for South East Asian breast cancer patients.
Axillary lymph node status is one of the most important prognostic factors for breast cancer
Increasing evidence suggests that the Lymph Node Ratio (LNR) (the ratio of the number of positive nodes to the total number of nodes excised), is a superior prognostic indicator compared to the absolute number of nodes involved
Prognostication, however, is a multivariable process, as the outcome of a disease is determined by a variety of (sometimes interacting) factors, and breast cancer is no exception. In addition to axillary lymph node status, prognosis is determined by a variety of factors, including, age, tumor size, grade, receptors status and treatment. Despite the large number of studies that have addressed LNR, not one has assessed the added prognostic value of LNR over pN in predicting overall survival after breast cancer. Via this study we aim to assess the added prognostic value of LNR over pN staging in the South East Asian setting by comparing the pN and LNR prediction models in terms of (1) predictive power, (2) discrimination and (3) net reclassification improvement of patient into appropriate risk categories of all cause mortality.
Data for this study were obtained from the Singapore Malaysia Hospital-based Breast Cancer Registry
The NUH breast cancer registry started in 1995 and contains information on 2,449 consecutive breast cancer patients diagnosed between 1990 and 2007. The UMMC breast cancer registry started in 1993 contains information on 3,320 patients diagnosed between 1993 and 2007. Details on both these registries are described elsewhere
We selected women diagnosed with non metastatic primary invasive breast cancer, with information on the number of excised and the number of positive axillary lymph nodes. Patients receiving neoadjuvant chemotherapy (N = 312), patients with a node negative (pN0) axilla (N = 2352), patients with missing information on exact number of lymph nodes involved (N = 664), with
Information recorded for each patient included age at diagnosis, ethnicity (Chinese, Malay, Indian or others), year of diagnosis, place of diagnosis (Singapore, Kuala Lumpur), date of death or date of last contact. Tumor characteristics included tumor size (<2 cm, 2–5 cm, >5 cm, unknown), estrogen (ER) and progesterone receptor (PR) status (positive i.e., ≥10% of epithelial tumor cells expressing receptors, negative and unknown), grade (good, moderate, poor, unknown). In terms of axillary dissection, we collected information on total number of axillary nodes examined and number of positive axilary nodes. LNR was categorized into three categories including, low (>0 and <0.2), intermediate (0.2 to 0.65) and high category (>0.65 to 1) groups as previously reported
Life table analysis was performed to calculate survival probabilities for the three pN categories and the three LNR categories. After testing for proportionality, we performed univariate Cox proportional hazard analysis to identify variables that were significantly associated with all cause mortality. Multivariate Cox proportional hazard analysis was applied 1) to calculate adjusted mortality risks and 2) to identify which combination of factors best predicted overall survival. For this we entered all variables univariately associated with overall survival with a p-value <0.2 into the model and used stepwise backward regression and maximum likelihood method to find the optimal fit. Internal validation of each model was done by bootstrap resampling.
Two models (A and B) were constructed. Each model contained the same baseline variables, i.e., age, radiotherapy, surgery type, grade and tumor size (base model). Model A contained pN stage in addition to the base model variables while Model B contained LNR in addition to the base model variables. From the final models, adjusted Hazard Ratio for pN and LNR were derived.
Base model : age, radiotherapy, surgery type, grade and tumor size.
Model A: age, radiotherapy, surgery type, grade, tumor size and pN stage.
Model B: age, radiotherapy, surgery type, grade, tumor size and LNR.
In order to ascertain the added prognostic value of LNR over pN, we compared the discriminative capacity of model A with model B. Discrimination indicates how well the model is able to distinguish between patients who will experience the outcome (death) and those who will not. Discrimination was assessed by the Concordance (c) statistic, the interpretation of which is equivalent to the area under the receiver operating characeristic (ROC) curve, that is, a c statistic of 0.5 indicates no discrimination above chance, whereas a c statistic of 1.0 indicates perfect discrimination. Comparison of c statistics between the model including pN Stage (Model A) with the one including LNR (Model B) tells whether one model is better in discriminating between poor and good survivors, and thus superior in predicting survival. Model calibration–the agreement between predicted risks and observed mortality risks–was assessed using the Hosmer Lemeshow test by comparing the predicted survival and the observed survival at 3-year follow-up.
Finally, the c statistic has been criticized for being insensitive in comparing models and for having little direct clinical relevance. Therefore, we calculated the Net Reclassification Improvement (NRI), which assesses the ability of a model including a new prognostic marker (LNR - model B) to more accurately reclassify individuals into higher or lower risk(of death) category compared to model A, i.e., to check whether model B was better at correctly reclassifying patients into high risk and low risk groups based on their predicted survival probability as compared to model A. The NRI is the difference in proportions of patients moving up and down risk categories (high, moderate and low risk of mortality) among patients with the event of interest (death) versus those without (in our case patients who died within 3 years of follow up versus those who survived).The NRI is similar to the percentage reclassified but distinguishes between movements in the correct direction (patients moving up the risk categories for event patients (deaths) and down for nonevent patients (survivors))
The NRI is calculated as follows:
Pup,event = number of events moving up/number of events.
Pdown,event = number of events moving down/number of events.
Pup,nonevent = number of nonevents moving up/number of non events.
Pdown,nonevent = number of nonevents moving down/number of non events.
NRI = (Pup,event - Pdown,event) – (Pup,nonevent - Pdown,nonevent).
Where “up” refers to the patients moving up in the risk categories based on the new model when being compared to the old model and “down” refers to the patients moving down in the risk categories based on the new model when being compared to the old model.
In order to estimate Pup,event, Pdown,event, Pup,nonevent, Pdown,nonevent, we first determined the the predicted survival probability for each patient based on models A and B. Based on this predicted survival probability patients were categorized into tertiles corresponding to low, intermediate and high risk of death at 3 years of follow up. The majority of the patients were correctly classified by both the models (as indicated by a high proportion of patients falling on the diagonals in the risk classification table).
After a recent publication suggested that LNR is particularly informative in subgroups of patients (i.e. patients with unfavorable tumor characteristics and younger patients) we performed subgroup analyses by age (<60 years and ≥60 years), receptor status (ER- vs ER+) and grade (1, 2 and 3)
All analyses were performed using STATA version 11.
According to the LNR classification, 758 (47.7%) patients were categorized as low category (>0 and <0.2), 574 (36.1%) as intermediate category (0.2 to 0.65) and 257 (16.2%) as high category (>0.65 to 1) LNR corresponding to low, intermediate and high risk of death respectively. For classic pN staging, 879 (55.2%) were pN1, 447 (28.1%) pN2 and 263 (16.7%) pN3 (
Variable | N (%) | Unadjusted HR (95% CI) | P value of unadjusted HR |
|
<0.001 | ||
Median (Range) | 50 (22 to 87) | ||
<40 years | 225 (14.2%) | 1 | |
40 to 49 years | 569 (35.8%) | 0.7 (0.5 to 0.9) | |
50 to 59 years | 470 (29.6%) | 1.0 (0.8 to 1.3) | |
≥60 years | 325 (20.5%) | 1.2 (0.9 to 1.6) | |
|
0.76 | ||
1990–2000 | 521 (32.8%) | 1 | |
2001–2007 | 1068 (67.2%) | 1.0 (0.8 to 1.2) | |
|
<0.001 | ||
Kuala Lumpur | 1015 (63.8) | 1 | |
Singapore | 574 (26.2%) | 0.4 (0.3 to 0.5) | |
|
0.005 | ||
Chinese | 1064 (67.0%) | 1 | |
Malay | 303 (19.1%) | 1.2 (1.0 to 1.5) | |
Indian | 176 (11.1%) | 1.5 (1.1 to 1.9) | |
Other | 46 (2.9%) | 0.8 (0.4 to 1.6) | |
|
<0.001 | ||
Negative | 662 (44.0%) | 1 | |
Positive | 844 (56.0%) | 0.5 (0.4 to 0.7) | |
Unknown | 83 | 0.8 (0.5 to 1.1) | |
|
<0.001 | ||
Negative | 596 (45.7%) | 1 | |
Positive | 706 (54.3%) | 0.4 (0.3 to 0.6) | |
Unknown | 287 | 0.8 (0.6 to 1.0) | |
|
<0.001 | ||
Low | 89 (6.2%) | 0.4 (0.2 to 0.7) | |
Moderate | 699 (49.1%) | 1 | |
High | 635 (44.6%) | 1.4 (1.1 to 1.6) | |
Unknown | 166 | 1.0 (0.7 to 1.4) | |
|
<0.001 | ||
≤2 cm | 381 (26.0%) | 0.5 (0.4 to 0.7) | |
2.1–5 cm | 868 (59.3%) | 1 | |
>5 cm | 214 (14.6%) | 1.6 (1.3 to 2.0) | |
Unknown | 126 | 0.8 (0.6 to 1.1) | |
|
<0.001 | ||
No | 430 (26.9%) | 1 | |
Yes | 1159 (72.9%) | 0.7 (0.5 to 0.8) | |
|
<0.001 | ||
No | 246 (15.5%) | 1 | |
Yes | 1343 (84.5%) | 0.5 (0.4 to 0.6) | |
|
<0.001 | ||
No | 560 (35.2%) | 1 | |
Yes | 1029 (64.8%) | 0.5 (0.4 to 0.6) | |
|
0.151 | ||
Median | 15 | ||
1–3 | 18 (1.1%) | 1.8 (0.9 to 3.3) | |
4–9 | 249 (15.7%) | 1.0 (0.8 to 1.2) | |
≥10 | 1322 (83.2%) | 1 | |
|
<0.001 | ||
Median | 3 | ||
1–3 | 879 (55.2%) | 1 | |
4–9 | 447 (28.1%) | 1.7 (1.4 to 2.1) | |
≥10 | 263 (16.7%) | 3.3 (2.6 to 4.1) | |
|
<0.001 | ||
Median | 0.22 | ||
0.01–0.2 | 758 (47.7%) | 1 | |
0.201–0.65 | 574 (36.1%) | 1.5 (1.2 to 1.8) | |
0.651–1 | 257 (16.2%) | 3.6 (2.9 tp 4.5) |
indicates valid proportions have been calculated (i.e., not considering unknown).
Five year survival probabilities for the patients categorized by LNR were 79%, 70% and 43% for low, intermediate and high category LNR groups respectively (
Variable | N (%) | N of deaths | 5 year Survival Probability (95% CI) | Unadjusted HR (95% CI) | Adjusted HR (95% CI) | c statistic (95% CI) |
|
0.72 (0.70 to 0.75) | |||||
pN1 | 879 (55.2%) | 256 | 79.0% (75.6% to 82.4%) | 1 | 1 | |
pN2 | 447 (28.1%) | 198 | 65.0% (59.0% to 71.0%) | 1.7 (1.4 to 2.1) | 1.9 (1.5 to 2.3) | |
pN3 | 263 (16.7%) | 151 | 48.0% (43.2% to 52.8%) | 3.3 (2.6 to 4.1) | 3.0 (2.4 to 3.7 | |
|
0.73 (0.71 to 0.76) | |||||
Low ≤0.20 | 758 (47.7%) | 213 | 79.0% (75.4% to 82.6%) | 1 | 1 | |
Intermediate >0.20 to ≤0.65 | 574 (36.1%) | 228 | 70.0% (65.2% to 74.8%) | 1.5 (1.2 to 1.8) | 1.5 (1.2 to 1.9) | |
High >0.65 | 257 (16.2%) | 164 | 43.0% (33.0% to 53.0%) | 3.6 (2.9 to 4.5) | 3.2 (2.6 to 4.0) |
Model A is adjusted for: age, radiotherapy, surgery type, grade and tumor size and pN stage and stratified by ER Status. Model B is adjusted for: age, radiotherapy, surgery type, grade and tumor size and LNR and stratified by ER Status. Both models were internally validated using bootstrap resampling.
In univariate Cox regression analysis, age at diagnosis, place of diagnosis, year of diagnosis, ethnicity, receptor status (ER and PR), treatment, grade, stage, tumor size, pN staging were independently and significantly associated with all cause mortality (
Both models A (base model plus pN) and B (base model plus LNR) were well calibrated (p-value Hosmer Lemeshow test 0.67 and 0.83 respectively). In terms of discriminating ability, both models performed equally well, as shown by the c statistic for model A of 0.72 (95% CI 0.70 to 0.75) and c statistic for the model B of 0.73 (95% CI 0.71 to 0.76). The substantial overlap between the two 95% confidence intervals indicated that LNR did not provide any added prognostic value when compared to pN staging in predicting all cause mortality.
Based on individual predicted survival probabilities (from both pN staging and LNR models), when patients were categorized into tertiles of low, intermediate and high risk of death, the LNR model reclassified an additional 8.0% (n = 49) of patients with the event (death) into high risk groups and incorrectly reclassified 4.5% (n = 29) of the patients with the event into low risk groups. Among the patients without the event (alive), an additional 5.6% (n = 52) of patients were reclassified into low risk groups while 5.7% (n = 53) of the patients without the event were reclassified into high risk (
As per model A (with pN) | |||||
Low risk of death | Intermediate riskof death | High risk of death | Total | ||
For patients withthe event (Dead) | Low risk of death | 127 | 24 | 151 | |
Intermediate risk of death | 23 | 335 | 25 | 383 | |
High risk of death | 6 | 65 | 21 | ||
Total | 150 | 365 | 90 | 605 | |
|
|||||
For patients without theevent (alive) | Low risk of death | 405 | 45 | 450 | |
Intermediate risk of death | 48 | 396 | 8 | 452 | |
High risk of death | 4 | 16 | 20 | ||
Total | 453 | 445 | 24 | 922 |
Net Reclassification Index (NRI) = 3.2% (p value 0.08). Patients are categorized into risk categories of death based on their individual survival probabilities obtained from models A and B such that a patient with a high survival probability is categorized into the ‘low risk of death’ group and so on.
Subgroup analysis showed that LNR was superior to pN staging in categorizing patients’ risk of death for patients aged 60 years and above, patients with ER negative tumors and patients with high grade tumors, as in, for these subgroups, 95% confidence intervals (CIs) for intermediate and high risk LNR groups did not overlap while they did for the pN2 and pN3 categories. However, in terms of discriminating ability, models for all subgroup analyses including LNR performed as well as the models including pN respectively, as attested by the c statistics and largely overlapping 95% CIs (
Patients ≥60 years of age at diagnosis (N = 325) | |||||
N (%) | N Death (%) | Unadj HR (95% CI) | Adj HR |
C statistic (95% CI) | |
|
0.75 (0.70 to 0.81) | ||||
pN1 | 175 (53.8%) | 53 (36.3%) | 1 | 1 | |
pN2 | 89 (27.4%) | 51 (34.9%) | 2.8 (1.8 to 4.1) | 2.7 (1.8 to 4.1) | |
pN3 | 61 (18.8%) | 42 (28.8%) | 4.2 (2.7 to 6.3) | 4.2 (2.6 to 6.7) | |
|
0.76 (0.71 to 0.80) | ||||
Low ≤0.20 | 147 (45.2%) | 44 (30.1%) | 1 | 1 | |
Intermediate >0.20 to ≤0.65 | 112 (34.5%) | 51 (34.9%) | 1.6 (1.0 to 2.4) | 1.8 (1.1 to 2.7) | |
High >0.65 | 66 (20.3) | 51 (34.9%) | 5.2 (3.4 to 7.8) | 4.5 (2.8 to 7.0) | |
|
|||||
|
|
|
|
||
|
0.84 (0.80 to 0.87) | ||||
pN1 | 339 (51.2%) | 100 (36.0%) | 1 | 1 | |
pN2 | 206 (31.1%) | 106 (38.1%) | 2.0 (1.5 to 2.6) | 2.0 (1.5 to 2.7) | |
pN3 | 117 (17.7%) | 72 (25.9%) | 3.1 (2.3 to 4.3) | 3.0 (2.1 to 4.1) | |
|
0.85 (0.81 to 0.88) | ||||
Low ≤0.20 | 304 (45.9%) | 93 (33.6%) | 1 | 1 | |
Intermediate >0.20 to ≤0.65 | 233 (35.2%) | 95 (33.9%) | 1.4 (1.0 to 1.9) | 1.5 (1.1 to 2.0) | |
High >0.65 | 125 (18.9%) | 90 (32.5%) | 3.7 (2.7 to 4.9 | 3.5 (2.5 to 4.8) | |
|
|||||
|
|
|
|
||
|
0.76 (0.72 to 0.80) | ||||
pN1 | 320 (50.4%) | 109 (40.1%) | 1 | 1 | |
pN2 | 180 (28.3%) | 84 (30.9%) | 1.6 (1.2 to 2.1) | 1.7 (1.2 to 2.3) | |
pN3 | 135 (21.3%) | 79 (29.0%) | 2.6 (1.9 to 3.5) | 2.6 (1.9 to 3.5) | |
|
0.76 (0.72 to 0.81) | ||||
Low ≤0.20 | 286 (45.0%) | 100 (36.9%) | 1 | 1 | |
Intermediate >0.20 to ≤0.65 | 229 (36.1%) | 94 (34.7%) | 1.3 (1.0 to 1.7) | 1.4 (1.1 to 1.8) | |
High >0.65 | 120 (18.9%) | 77 (28.4%) | 2.9 (2.1 to 3.1) | 2.7 (2.0 to 3.7) |
Model adjusted for age at diagnosis, chemotherapy, radiotherapy, surgery type, grade and tumor size and stratified by ER status.
Model adjusted for age at diagnosis, chemotherapy, surgery type and tumor size.
Model adjusted for age at diagnosis, chemotherapy, radiotherapy, surgery type and tumor size and stratified by ER status. All models were internally validated using bootstrap resampling.
Although a majority of the patients (∼83%) did have at least ten lymph nodes examined, about 17% of the patients had less than 10 nodes removed during axillary dissection. We performed a subgroup analysis to assess the added prognostic value of LNR for patients with less than 10 nodes retrieved but even for this subset of patients, both pN staging and LNR predicted all cause mortality equally well (data not shown). Different cut offs for LNR were tested for the entire dataset but no new cut offs of LNR for South East Asian patients were established.
This study shows that pN staging as well as the LNR are comparable in predicting overall survival of women with breast cancer, except for patients aged 60 or more, patients with ER negative tumors and patients with high grade tumors. Here, LNR was superior in categorizing patients into intermediate and high risk strata as compared to pN stage. However, in combination with other prognostic factors, LNR did not provide any additional prognostic information over pN staging, neither for the entire cohort, nor for the subgroups of older women and those with ER negative of grade 3 disease. The fact that LNR was not superior to the pN staging was seen in other Asian studies as well
There are several independent but interrelated prognostic factors that predict for recurrence and survival of breast cancer patients. These include amongst others, tumor size, axillary nodal status, histopathology, steroid receptors, HER 2 status, proliferative rate, ploidy, and oncogene amplification
The number of lymph nodes retrieved and examined is highly dependent on surgical expertise, the institution’s protocol and the pathologists’ experience
Results from our study showed that LNR and pN were equally good at predicting all cause mortality overall but within certain subgroups (ER negative patients, patients aged 60 years or more and patients with high grade tumors), LNR was better at categorizing patients into risk categories. The intermediate category LNR was truly intermediate for these subgroups, i.e., the 95% Confidence Interval (CI) of the Hazard Ratio overlapped neither the low nor the high category LNRs, whereas the pN2 and pN3 CIs overlapped (
Recent studies have indicated that full axillary clearance following a positive sentinel node biopsy does not affect survival in certain (low risk) categories of breast cancer patients
We acknowledge that our study suffers from several shortcomings, including a relatively short follow up time. In addition, we assessed all cause mortality as our end point as no data on cause of death was available. This could have led to a mixing of effects as this analysis allowed for competing risks of death. Also, additional information on HER2/NEU receptor status, socioecomonic status and comorbidity could have allowed for a deeper understanding of the association.
Among South East Asian breast cancer patients, both the Lymph Node Ratio and the pN staging system seem to be equally good at predicting all cause mortality based on the cut offs used for LNR in this study. LNR may be better than pN in dividing tumors into high vs low risk for certain subgroup of patients, but LNR has no added prognostic value over pN staging in addition to other prognosticators.
(DOCX)