Nomogram预后模型：预测淋巴结阴性乳腺癌患者个体预后

A nomogram to predict individual prognosis in node-negative breast carcinoma

• C. Mazouni^a, b, , , 
• F. Spyratos^c, 
• S. Romain^a, 
• F. Fina^a, 
• P. Bonnier^d, 
• L.H. Ouafik^a, 
• P.M. Martin^a

• ^a Laboratoire de transfert d’oncologie biologique, Assistance Publique – Hôpitaux de Marseille, Faculté de Médecine Nord, Marseille, France
• ^b Département de chirurgie générale, Institut Gustave Roussy, 114 rue Edouard Vaillant, Villejuif 94805, France
• ^c Laboratoire d’Oncogénétique, Institut Curie – Hôpital René Huguenin, Paris, France
• ^d Institut de Chirurgie et d’Oncologie Gynécologique et Mammaire, Hôpital Beauregard, Marseille, France

• http://dx.doi.org/10.1016/j.ejca.2012.04.018, How to Cite or Link Using DOI

• Permissions & Reprints

---

Keywords

• Adjuvant!; 
• Breast cancer; 
• Nomogram; 
• PAI-1; 
• Prognosis score

---

1. Introduction

With the worldwide development of mammographic screening, the management of early breast cancer, and particularly node-negative breast cancer (NNBC) has increased. However, while the surgical procedures in these patients have tended to decrease with the development of the sentinel node biopsy (SNB), the decision regarding adjuvant systemic therapy in NNBC remains controversial because the extent of the chemotherapy benefit is low in NNBC.¹

The advent of translational and genomic research has improved our understanding of NNBC, and has shown that NNBC represent a heterogeneous group of tumours, influenced by a complex combination of clinical and biological factors. For instance, stratification by the oestrogen receptor (ER) status shows a different natural history as an assessed hazard of recurrence.^2 and 3 Thus in ER-positive NNBC, the administration of adjuvant treatment is under debate.⁴ In particular, the results of the long-term follow-up of National Surgical Adjuvant Breast and Bowel Project (NASBP) trials show a degree of time dependence in the treatment effect for hormone therapy alone or with chemotherapy compared to surgery alone.³ In ER-negative patients, chemotherapy had a substantial effect on reducing the hazard of recurrence in a regimen-dependent manner. Finally, the estimated benefit of adjuvant chemotherapy in early NNBC might appear modest and be offset by the associated toxicities.

Thus prognostic markers are needed to reduce overtreatment with pointless exposure to toxicity. Identifying prognostic markers that could help discriminate an individualised prognosis might help improve targeted treatment of NNBC.¹ During the last decade, numerous prognostic factors in primary BC have been described in the literature but few have reached a satisfactory level of evidence (LOE).^5 and 6 Recently, the urokinase plasminogen activator (uPA)–plasminogen activator inhibitor type 1 (PAI-1) complex was recognised as both a strong prognostic marker and a predictive marker, as demonstrated by the NNBC3 trial.^7 and 8 However, the use of uPA/PAI-1 in current practice for decision making does not appear to have been accepted by the panel audience in the 2011 St. Gallen Consensus conference.⁹ Recently, new tools including genomic signatures such as the Oncotype Dx, or Mammaprint have been proposed but clinical routine genomic profiling is not yet a standard. As a matter of fact, immunohistochemical testing of ER/PgR and HER2 still have a long-way to go in prognostic estimation. Simple tools developed from statistical analysis have also been proposed such as Adjuvant! Online based on multiparameter factors that might help estimate an individual prognosis in the absence or presence of adjuvant treatment. In a recent publication, Schmitt et al. proposed a risk prognosis score algorithm to improve the treatment strategy in NNBC.¹⁰

The purpose of the present study was twofold: (1) to analyse the relative contributions of tumour-related variables to survival in NNBC; and (2) to develop a nomogram to estimate the individual prognosis in NNBC.

2. Patients and methods

The present study was performed among a group of 305 consecutive patients with NNBC who had undergone primary surgery in the Department of Breast Surgery, in the Marseille Public Hospital System (n = 131) and at the Institut Curie – René Huguenin Hospital (n = 174), in France from January 1980 to December 2000. Clinical and histological characteristics of all patients were obtained from their medical records and entered prospectively into institutional clinical databases. To better estimate the prognostic variables, cases that had not received systemic therapy were collected to approach the natural outcome of NNBC. Patients who had received primary chemotherapy (CT) or hormone therapy (HT) according to local guidelines at the time, as well as patients with distant metastases at the time of the diagnosis were not selected for this study.

Tissue samples of invasive BC cases were obtained following Institutional Review Board approval and informed consent. Breast cancer tissues were prospectively collected. Tumour specimens were obtained from a pool of frozen specimens, and all samples had been stored in liquid nitrogen until use. Steroid receptors uPA1 and PAI1 had been assayed for all patients. Measurements had been obtained within 1 month of surgery. The processing and pathologic examination of the tumours had been performed as previously described.^11 and 12 Briefly, uPA and PAI-1 levels had been analysed in cytosolic tumour extracts by means of an enzyme-linked immunosorbent assay (Immunobinds ELISA Kit, American Diagnostic, Greenwich, CT, United States of America (USA)). Steroid hormone receptors (ER – PgR) had initially been determined biochemically in cytosol fractions and then expressed quantitatively [fmol/mg protein] (Abbott Laboratories, Diagnostic Division, Chicago, IL). Tumour grading was a∗∗ccording to the well-established Scarff–Bloom–Richardson criteria. Thymidine kinase (TK) activity had been measured in the cytosol using a radioenzymatic phosphorylation assay (TK-REA; Sangtec Medical, Bromma, Sweden) optimised to detect the foetal isoenzyme. The results are expressed in mU/mg protein. Quality control had been undertaken by frequent testing with internal controls according to European Organisation for Research and Treatment of Cancer (EORTC) standards. The results are expressed as pmol/mg protein.

3. Survival analysis

Survival rates were calculated using the Kaplan–Meier method and compared between groups with the log-rank test. Disease-free survival (DFS) was defined as the time from surgery to the date of disease recurrence or death from cancer if there was no earlier recurrence. Breast cancer-specific survival (BCSS) was computed from the date of surgery to the date of death from BC. Patients who had died of an unrelated cause were censored at the date of death.

Additionally, for each woman, Adjuvant! Online standard version 8 was used to generate 10-year predictions of BCSS and DFS. Such predictions were obtained by entering into the programme information on each patient’s age, tumour size, number of positive nodes, grade, ER status. The comorbidity assumption ‘average for age’ was used for the entire cohort as a default setting. To dichotomise the patients, we used the cut-off as proposed in the Microarray In Node-negative Disease may Avoid ChemoTherapy (MINDACT) trial¹³ and in the algorithm proposed by Schmidt et al. (2009). A 10-year breast cancer survival probability of at least 88% for ER-positive and of at least 92% for ER-negative carcinomas was selected as the cut-off point.

Univariate and multivariate Cox models were used to verify the independent prognostic power of each parameter. Nomogram development began by identifying patient characteristics predictive for overall survival in the multivariate Cox model, which used the Akaike’s information criterion (AIC) as a stopping rule. The nomogram was constructed as described by Kattan et al.¹⁴ The predictive accuracy of various Cox models was quantified by calculating the c-index, which is a probability of concordance between predicted and observed survival, equal to the area under the receiver operating characteristics curve for censored data.¹⁵ A c-index of 0.5 indicates that outcomes are completely random, whereas a c-index of 1 indicates that the model is a perfect predictor. A calibration curve, generated by plotting actuarial survival against predicted survival probabilities for patients stratified by predicted risk, assessed the prediction accuracy of the nomogram.

All p values are two sided. All analyses were performed in R, an open source statistical package (http://www.r-project.org/), using the Design library.¹⁶

4. Results

4.1. Patient characteristics

The demographics and clinico-pathologic characteristics of the 305 patients with primary BC are summarised in Table 1. More than 76% of patients had presented with ER-positive tumours. A total of 64.6% (N = 197) patients had undergone a breast-conserving lumpectomy with axillary lymph node dissection and 35.4% (N = 108) had undergone a modified radical mastectomy with auxiliary lymph node dissection.

4.2. 10-year breast cancer-specific survival and disease-free survival

Median follow-up was 241 months. A total of 23.6% (N = 72) of patients had relapsed and 22.9% (N = 70) had died of BC.

According to the univariate Cox proportional hazard model for BCSS (Table 2), of the eight variables tested, age (p < 0.001), uPA (p = 0.003) and PAI-1 (p < 0.001) were predictors of BCSS (Table 2). A simplified prognostic index (PI) was developed using the six prognostic factors (Table 3). The index score is based on the sum total of factors with points allotted for each of the following variables: age, PgR status, pT, and PAI levels.

On the basis of the results of the multivariate analysis, risk score (RS) points were assigned to the four risk factors for BCSS as shown in Table 3. Each independent predictor for BCSS was assigned risk score (RS) points of 1–3. The PI was obtained by summing all RS for each patient. The Prognostic Index ranges from 3–8 points and the risk is assigned as follows: index score 1–3, low; index score 4–5, intermediate; and index score 6 or greater, high risk (Table 3).

On the basis of the PI, patients were assigned to one of three risk categories with 61 (37%) of the patients in the low-risk group, the remaining cases were categorised as intermediate (N = 161) or high risk (N = 83) (Fig. 1). According to Adjuvant! Online, one hundred and seventy-seven (58%) of the patients were categorised as low-risk and 128 were assigned to the high-risk group (42%).

In the univariate analysis, the PI was significantly correlated with BCSS (hazard ratio (HR): 4.1 for the intermediate-risk group, p = 0.02, HR: 8.8, p < 0.001 for the high-risk group), while the risk classification according to Adjuvant! Online did not predict BC survival in our cohort (HR: 1.4, p = 0.14). Kaplan–Meier estimates showed that after 10 years, 100% of the patients in the low-risk group were alive compared with 97.1% in the intermediate-risk group and only 88.8% in the high risk-group for the PI score, p < 0.001, (Fig. 2A). According to Adjuvant! Online, 10-year survival was 99.3% in the low-risk group versus 90.6% in the high-risk group (p = 0.14) (Fig. 2B). The percentages of 10-year survival probabilities are provided in Table 4.

Concerning DFS, in the univariate analysis, only age (p = 0.03) and PAI levels (p < 0.001) were significant predictors while the prognostic index score was of borderline significance (p = 0.05). Conversely, Adjuvant! Online did not reach statistical significance (p = 0.46). Kaplan–Meier estimates showed that after 10 years, 96.4% of the patients in the low-risk group were disease free compared with 90.6% in the intermediate-risk group and only 82.3% in the high-risk group for the PI score (p = 0.13). According to the Adjuvant! Online score, 10-year DFS was 93.8% in the low-risk group versus 84% in the high-risk group (p = 0.46). The respective DFS curves for the PI and the Adjuvant! Online score are shown in Fig. 3A and B.

4.3. Prognosis nomogram

A nomogram (Fig. 4) was developed to predict BCSS using the six covariates identified in the multivariate model (Table 2). The selection of variables to be introduced into the multivariate analysis was based on fast backward elimination. The multivariate Cox model including six variables (age, uPA, PAI-1, PgR, pT, and ER) is shown in Table 2. Age (p < 0.001), PgR status (p = 0.02) and the PAI-1 level (p < 0.001) were independent predictors of BCSS in the multivariate Cox regression model (Table 2). Using the backward selection procedure and AIC, age, pT (p = 0.12), the PgR status and PAI-1 level were retained to construct the nomogram.

The nomogram is used by totalling the points identified at the top of the scale for each independent covariate. This total point score is then identified on the total points scale to determine the probability of 10-year BCSS. The predictive accuracy of this nomogram relative to our data, using Harrell’s c-index, was 0.71 before calibration and 0.70 after calibration. The slight early survival advantage in our series appeared to emanate predominantly from the patients with higher nomogram scores corresponding to a poorer prognosis. The calibration curve (Fig. 5) illustrates how the predictions from the nomogram compare with the actual outcomes for the 305 patients.

5. Discussion

The nomogram we developed is an interesting new tool to approach breast cancer-specific survival in NNBC and therefore anticipates the need for additional adjuvant treatment. From the different nomograms or statistical models proposed in BC,^{17, 18 and 19} few have been of interest for the prognosis.^{20, 21 and 22} The most popular is Adjuvant! Online, which provides the 10-year survival probability based on patient age, tumour size and grade, ER-status, nodal status, and co-morbidities. The estimates were derived from the results of a large number of trials including thousands of patients.^22 and 23 However, all of these models were based on clinical or classic histological factors and none of them included emerging markers. The interest of our study is to have provided the first innovative tool with clinical, histological and the stromal-related protease PAI-I recently validated as a level of evidence I prognostic marker.^5 and 6

Currently, there is growing interest in the application of the nomogram in oncology to improve decision making. The rationale for using multiple factors to predict the prognosis is based on the fact that a disease is associated with multiple risk factors. Incorporating multiple variables into cancer prediction models provides more accurate predictions than simple risk classification.²⁴ Nomograms proposed have frequently been used to identify prognostic groups^25 and 26 and have the advantage of being visually intuitive and capable of pinpointing potential interactions among predictors. In the present analysis, we identified independent predictors for survival in NNBC that included four recipient risk factors, i.e. age, pT, PgR status and PAI-1 level. An advantage of the nomogram is that it is a weighted model, combining independent prognostic factors and enabling an appraisal of the extent of the impact of each of the factors on the probability of survival. Thus, in our model, two major heavily-weighted factors were age and PAI-1. Unsurprisingly, PAI-1 was highly significant in our study and has reproducibly been correlated with survival among patients in several large studies.^{8, 27 and 28} Although pT was not a statistically significant predictor of the prognosis, it was also included in the risk predictor model and in the risk score estimation as this could increase the predictive power.²⁶

In addition, in our study a derived prognostic index is proposed based on these factors. The present model adopted slightly liberal criteria (p < 0.10) to avoid missing well established risk factors that affect survival, namely age in our case. Each predictor was assigned a point which is summed to stratify patients according to one of three risk categories. This prognostic model allowed risk stratification of NNBC that are associated with different outcomes and performed better than Adjuvant! Online for prognosis prediction. Interestingly, a lower proportion of the patients were classified in the low-risk group (37%) according to the PI score, as compared to the Adjuvant! Online classification (58%). Consequently, on the basis of the PI score, fewer patients would have received CT. However, in the PI, an additional intermediate-risk group is proposed with a more pronounced different outcome for DFS.

We acknowledge that our study had some limitations. For instance, human epidermal growth factor receptor 2 (HER2) was not included in our model as it was not measured at the time of treatment. However, this model was appropriate for the comparison with the current version of Adjuvant! Online. A further limitation of this study is the absence of an external validation of our proposed model to test its reproducibility and performance. This is a pilot model pending further work on a larger sample size from different institutions. The observed area under curve (AUC) of 0.711 still represents a high level of predictive accuracy. Furthermore, the current nomogram shows fairly good predicting performance although it might be optimistic regarding the high probabilities of survival as shown by the calibration. Additionally, the current limitation in the use of uPA/PAI-1 is the assay method currently available. According to the ASCO guidelines of 2007,⁵only uPA/PAI-1 levels measured by ELISA are validated. Currently, ELISA is the only method for measurement of uPA and PAI-1 levels, and required fresh extract tissues to measure proteins’ levels.²⁹ This might explained the non-wide use of this assay since with the development of mammograhic screening programmes have lead to the increase of small tumour detection, and the concurrence with other immunohistochemistry or genomic assays have limit the availability of tissues. Some authors have explored the value of uPA/PAI-1 assays on preoperative core biopsy,^30 and 31 and showed some variability in the level of expression of uPA/PAI-1, related to time from biopsy to surgery. However this can be overcome by choosing samples that are not in close proximity to the biopsy channel. Some alternate methods have been investigated for quantification of uPA/PAI-1 in FFPE tissues. However, it appeared that immunohistochemistry (IHC) is less quantitative and more subjective than the ELISA; ELISA values overlapped the IHC scoring classes. Other authors proposed RT-PCR-based assays to determine uPA and PAI-1 mRNA levels but with conflicting results.^32 and 33

In our series, patient’s age was related to prognosis, and older patient’s presented with adverse prognosis. In the risk score we developed, age <50 years was scored one point, the lower influence on prognosis. This point might be discussed since classically young patients developing breast cancer present lower survival. However, in our series, we are evaluated node negative breast cancer with have higher survival and these tumours did not receive adjuvant CT which presume a good biologic profile of the tumour. Moreover, the prognosis of BC in young (<50 years) women is varying according to the time from occurrence of the disease. The lowest excess of mortality is observed within the first 5 years after diagnosis.³⁴ Moreover, the prognosis of patients diagnosed with BC before age 50 years has considerably improved during the past decades. Nevertheless, these patients continue to have increased mortality throughout at least 40 years after diagnosis.³⁵

We believe that our nomogram could be a simple and easy tool for both the physician and patients for estimating the disease outcome (natural history of the disease) in the absence of treatment in NNBC, and could contribute to decision making regarding adjuvant chemotherapy.^1 and 7 Thus, for example, a patient aged 55 years (38 points), with a negative PgR tumour (30 points), measuring 20 mm (37 points) and a 5 ng/mg PAI level (10 points) would score 115 total points that converts to a 10-year survival probability of 87%. Based on this estimate of the natural history of the patient’s tumour, the possibility of administering CT might be discussed. The accuracy of our model could probably be improved by integrating additional predictor variables or existing biomarkers. Future work will focus on validating this model, both externally and in a prospective manner. We are also currently investigating incorporating into this type of prognostic modelling newer prognostic factors such as the 70-gene prognostic signature. In addition, survival modelling is being developed for patients after adjuvant treatment and for other end-points such as time to treatment, time to progression, time to treatment failure.

6. Conclusion

We have developed a nomogram that helps to estimate BCSS in NNBC. The addition of stromal-related protease PAI-1 might improve the decision strategy as compared to Adjuvant! Online. A PI score derived from the nomogram might improve the selection of the appropriate treatment strategy based on risk assessment for each individual patient.

Conflict of interest statement

None declared.