A nomogram to predict the probability of axillary lymph node metastasis in early breast cancer patients with positive axillary ultrasound

Among patients with a preoperative positive axillary ultrasound, around 40% of them are pathologically proved to be free from axillary lymph node (ALN) metastasis. We aimed to develop and validate a model to predict the probability of ALN metastasis as a preoperative tool to support clinical decision-making. Clinicopathological features of 322 early breast cancer patients with positive axillary ultrasound findings were analyzed. Multivariate logistic regression analysis was performed to identify independent predictors of ALN metastasis. A model was created from the logistic regression analysis, comprising lymph node transverse diameter, cortex thickness, hilum status, clinical tumour size, histological grade and estrogen receptor, and it was subsequently validated in another 234 patients. Coefficient of determination (R2) and the area under the ROC curve (AUC) were calculated to be 0.9375 and 0.864, showing good calibration and discrimination of the model, respectively. The false-negative rates of the model were 0% and 5.3% for the predicted probability cut-off points of 7.1% and 13.8%, respectively. This means that omission of axillary surgery may be safe for patients with a predictive probability of less than 13.8%. After further validation in clinical practice, this model may support increasingly limited surgical approaches to the axilla in breast cancer.


Results
Clinicopathological characteristics of patients. The descriptive clinicopathological characteristics of patients in the modeling and validation group are provided in Table 1. No statistically significant differences were found between the two groups, except for P53, estrogen receptor (ER), and ALN variables including number of lymph nodes detected by ultrasound, transverse diameter, cortical thickness and absence of hilum.
197 patients (61.2%) positively expressed P53 in the modeling group, compared with 191 patients (81.6%) in the validation group. With regard to ER, the positive expression rate between the modeling group and validation group were quite similar, being 63% and 66.7%, respectively. However, the distribution of ER positive patients in the two groups were quite different, with a larger percentage of patients being ER 2+ in the modeling cohort (17.7%) compared to the validation cohort (9.0%), and with relatively smaller percentage in ER 1+ (6.8% versus 10.7%) and ER 3+ (38.5% versus 47.0%) in the modelling group. We considered that the difference might due to the relatively small number of patient population.
The absolute median number of transverse diameter and cortical thickness between the modeling group and the validation group matched closely, with 13 mm versus 15 mm, and both 4 mm, respectively. A total of 54.7% of patients lost their lymph node hilum in the validation group, compared with only 39.1% in the modeling group. More than one lymph node was found in 199 patients (61.8%) in the modeling group, compared with 166 patients (70.9%) in the validation group. Typical ultrasound findings of positive and negative lymph nodes are shown in Fig. 1.
Lymph node metastasis was detected in 163 patients (50.6%) in the modeling group, compared with 133 patients (56.8%) in the validation group. Results of comparison of lymph node metastasis by clinicopathological variables in the modeling group are shown in Table 2.
A nomogram was developed based on the result of multivariate logistic regression analysis (Fig. 2). The nomogram comprises nine rows and representation of each row is as follows. The mean predicted probability and mean actual rate of axillary metastasis in each decile were calculated. Calibration of the model was considered acceptable with R 2 of 0.9375 (Fig. 3). The Hosmer-Lemeshow Goodness-of-Fit test resulted in a p value of 0.18, indicating that the model fits well. The model was then validated with an additional set of 234 patients from the same institution. The performance of the model was good with an AUC of 0.864 (Fig. 4).
Sensitivity, specificity, accuracy and false-negative rate (FNR) of the model for several predicted probability cut-off values in the validation group are shown in Table 4. The FNRs of the model were 0% and 5.3% for the predicted probability cut-off points of 7.1% and 13.8%, respectively.   Discussion A precise noninvasive evaluation of ALN status preoperatively, although challenging, is very important for optimization of the treatment plan for patients with early breast cancer. Efforts have been made to construct a more precise and reliable tool for assessing the ALN status by using one, (e.g. ultrasound or digital mammography (MMG)), or a combination (e.g. magnetic resonance imaging (MRI), ultrasound and MMG) of current imaging modalities used in breast cancer patients, but the accuracy kept varying with less satisfactory. The FNRs of axillary ultrasound examination for determining lymph node metastasis were reported as 16.7% and 22.9% 15,16 . When combined with other modalities, such as MRI, physical examination, MMG, or positron emission tomography/computed tomography (PET/CT), the FNRs declined slightly, but were still as high as 14-16.9% 15,16 . Besides, the MRI and PET/CT examinations are too expensive to be routinely implemented in all patients. Diepstraten et al. 13 reported that the FNRs of ultrasound-guided ALN biopsy ranged from 15% to 40%, with a pooled estimate of 25%, which was much higher than the overall FNR of SLNB reported by the American Society of Clinical Oncology (ASCO) 17 . Moreover, the ultrasound-guided lymph node biopsy is an invasive procedure, which may increase risks like blood vessel damage. Development of a noninvasive, more practical, reliable and accurate clinical decision-making tool for predicting the probability of ALN metastasis is of great importance in clinical practice.
In the present study, we developed a predictive model to estimate the probability of ALN metastasis in early breast cancer patients with positive axillary ultrasound findings. The model was well calibrated in the modeling group and showed good performance for evaluation of nodal metastasis in the validation group with an AUC of 0.864. Some clinicopathological variables showed statistically significant differences between the modeling group and the validation group, which might be due to the relatively small sample size. Despite significant difference existed, the absolute median number of transverse diameter and cortical thickness between two groups matched closely. That explains why the model still performed well in the validation group, although transverse diameter and cortical thickness served as independent risk factors for node metastasis. Our result emphasizes the stability of the model when applied to a different set of patients. Since stability serves as one of the most important quality parameters for a well-developed model, the stability of our model showcased in this study provides us confidence that the model is generally applicable to other Asian patient populations, although further validation in independent patient cohorts is desirable.
According to NSABP B-32 18 , micrometastastic disease was an independent prognostic factor for patients with initially negative SLN. Therefore, in our study, IHC staining was routinely performed to detect micrometastasis when no tumour cells were identified on H&E staining. In a multivariate analysis, six variables emerged as independent predictors for ALN metastasis and were included in the final model. They were clinical tumour size, histological grade, ER, transverse diameter, cortical thickness and absence of hilum of the lymph node as detected by ultrasound. Tumour size and histological grade have been reported to be risk factors for ALN metastasis in many other studies 19   metastasis was found in tumours with negative expression of either ER 25 or PR 23,26 . In this study, we found that ER overexpression was associated with higher probability of ALN metastasis. This finding may seem counterintuitive, but is similar to the findings from Bevilacqua et al. 20 . Although we do not know the actual reason of this phenomenon, we hypothesize that ER negative tumours may prefer hematogenous metastasis rather than lymphatic metastasis. To the best of our knowledge, this is the first ALN metastasis predictive model mainly incorporating ultrasound parameters reported in the English literature. Preoperative axillary ultrasound is a simple test that is routinely used for assessing the clinical ALN status in breast cancer patients. According to previous reports, thickening of cortex and disappearance of hilum were associated with lymph node metastasis 14,16 . The result of our study is in line with those findings. In our final nomogram, transverse diameter, cortical thickness and hilum status were included as independent predictors of ALN metastasis. For an expert radiologist, axillary ultrasound can almost always identify lymph nodes and access the node morphology [27][28][29] . In that case, the three ultrasound parameters incorporated in our model could be easily obtained in almost all breast cancer patients, which will ensure the application of the model. Around 15-40% of patients with positive axillary ultrasound may have tumour-negative nodes on pathological evaluation 14,[30][31][32] . Obviously, it is of great value to identify those patients preoperatively, for whom SLNB or even omitting surgical intervention may be a suitable option. We created the predictive model based on data from patients with positive axillary ultrasound findings in order to select those patients with a low risk of node metastasis.  All patients were grouped into deciles according to their predicted probabilities. The mean predicted probability of each decile was plotted against the actual probability of axillary lymph node (ALN) metastasis. The reference line represents perfect equality of the predicted probability and the actual incidence of ALN metastasis. The coefficient of determination (R 2 ) reflects the calibration of the model. The calibration of the model was considered acceptable with R 2 of 0.9375.
In our study, patients with a predictive probability of less than 7.1% presented with no apparent lymph node metastasis. When the predicted probability cut-off point was set at 13.8%, the FNR of the model was only 5.3%, much lower than the FNR of SLNB reported by ASCO (overall 8.4%, range 0-29%) 17 . In that condition, omission of SLNB appears to be acceptable after thorough discussion with patients, especially in elderly patients with comorbid conditions. The pathological ALN status is not the only determinant to decide whether a patient should receive systemic adjuvant therapy or not. Histological grade, molecular subtype, tumour type, tumour size, and age of the patient should also be taken into consideration. In our study, only two patients, with an estimated predictive value of less than 13.8%, did not receive systemic therapy according to local guidelines. That is to say, patients would not be undertreated in terms of systemic therapy although they did not receive axillary surgery. Our model could provide doctors with a tool to estimate the probability of ALN metastasis and help them weigh the risks and benefits of SLNB more appropriately. Ultimately, this may prevent selected patients from undergoing unnecessary axillary surgery and associated comorbidity.
To our knowledge, several models have been developed to predict the probability of ALN metastasis, with reported AUC ranging from 0.702 to 0.79 20,21,33,34 . Among these models, the model developed by Bevilacqua et al. from the Memorial Sloan-Kettering Cancer Center (MSKCC) in 2007 is the most widely used. The MSKCC model produced an AUC of 0.754 when validated in another patient cohort from the same center 20 . Subsequently, the model was validated in a German breast cancer patient cohort 24 and two Chinese breast cancer patient cohorts 25,33 , with an AUC of 0.78, 0.71 and 0.73, respectively, all showing a reasonable ability in distinguishing patients with positive lymph nodes from those with negative nodes. The MSKCC model incorporates nine parameters (tumour type, pathological tumour size, lymphovascular invasion (LVI), tumour location, age, multifocality, nuclear grade, ER and PR), some of which are available only post-operatively, such as pathological tumour size, LVI, and multifocality. This may limit the clinical implication of the model, and thus a second surgical procedure might be needed. Two more predictive models reported by Takada et al. 34 and Chen et al. 33 also require parameters that are not available preoperatively. We believe that, when compared with those models, our model shows some advantages. First, our model contains only six variables that are all routinely available prior to surgery, e.g. histological grade and hormone receptor status can be obtained after core needle biopsy of the primary tumour. This is expected to facilitate the clinical utility of the model in different settings, irrespective of infrastructure. Second, our model showed an excellent discrimination, yielding an AUC of 0.864. By calculating the individual   probability of ALN metastasis using our user-friendly web-based calculator, doctors could inform their patients regarding their chances of having a positive ALN preoperatively. This knowledge will allow patients to better participate in discussing the treatment strategy on axillary nodes and make an informed decision.
Although our model provides a promising predictive value in early breast cancer patients, it still has some limitations. First, the model was only validated in one small internal validation cohort (n = 234). It needs to be further validated in external validation groups to evaluate its predictive ability and generalizability. Second, risk factors like clinical tumour size, cortical thickness and transverse diameter of lymph node may differ when measured by different doctors due to inter-observer variability. Taking the above-mentioned into account, our model cannot be considered an alternative to SLNB, even in patients with a low predicted probability of ALN metastasis. Validation studies using independent datasets to judge our model on its merits are encouraged. After further validation in clinical practice, this model may support increasingly limited surgical approaches to the axilla in breast cancer. All patients underwent a preoperative axillary ultrasound assessment using the IU22 (PHILIPS, The Netherlands) and ACUSON S2000 (SIEMENS, Germany) with a high-frequency transducer (12 to 15 MHz). Characteristics of lymph nodes including number of nodes, transverse diameter, longitudinal diameter, longitudinal-to-transverse axis ratio, cortical thickness, absence of medulla and absence of hilum were recorded. The ultrasound findings were reviewed and re-assessed by an experienced radiologist (C.C.) with more than 10 years of experience in breast ultrasound imaging. Surgical treatment for the primary disease included breast-conserving surgery or mastectomy with or without breast reconstruction. Systemic and radiation therapy were performed according to the local clinical practice guideline.

Methods
We utilized each of the following variables: age at diagnosis, menopausal status, clinical tumour size, tumour location, histological type, histological grade, ER, PR, human epidermal growth factor receptor 2 (Her-2), Ki-67, P53, vascular endothelial growth factor C (VEGF-C), as well as molecular subtype, and all above-mentioned variables of lymph node on ultrasonography. If more than one lymph node was detected, the variables were assessed on the lymph node most suspicious for metastasis.
Tumour tissues were obtained from paraffin embedded specimens for hematoxylin and eosin (H&E), immunohistochemistry (IHC), and fluorescence in situ hybridization (FISH) staining. The technique of histopathological analysis has been described previously 37 . Grading of invasive tumour was scored according to the Nottingham (Elston-Ellis) modification of the Scarf-Bloom-Richardson grading system. All nodes were examined postoperatively with serial section H&E staining. IHC staining was performed to determine whether micrometastasis (0.2-2 mm cancer foci) existed or not when no cancer cells were identified on H&E staining. ER 37 , PR 37 , P53 37 , and VEGF-C 38 were considered positive if immunostaining was positive in more than 10% of tumour cells. Her-2 positivity was defined as a score of 3+ on IHC 37 or amplification on FISH 39 . Ki-67 was considered positive if nuclear staining of tumour cells was > 14% and negative if ≤ 14% 19 .
Statistical analysis. Differences of continuous variables between groups were analyzed using the Mann-Whitney U test. The chi-square test was performed to compare the rates between different groups. In the modeling group, a univariate analysis was performed to assess risk factors for ALN metastasis. Variables with a p-value less than 0.05 in the univariate analysis were included in a binary logistic regression analysis using a backward selection procedure in order to discover the independent risk factors of ALN metastasis. Variables with a p-value less than 0.05 in the multivariate analysis, as an independent risk factor, were included in an equation for predicting the probability of ALN metastasis. Coefficients for each variable and the constant in the equation were generated based on multivariate analysis. A nomogram was developed to be a graphic representation of the model. To calibrate the model, the patients in the modeling group were grouped into deciles with respect to their predicted probabilities of lymph node metastasis calculated by the equation. For each decile, the mean predicted metastatic probability was compared with the mean actual metastatic rate. The calibration was assessed graphically and by calculation of the coefficient of determination (R 2 ). The predictive model was then validated with an additional set of 234 Chinese patients in the validation group. The receiver operating characteristic (ROC) curve was drawn, and the area under the curve (AUC) was used to assess the predictive accuracy of the model. Fit of the model was evaluated by the Hosmer-Lemeshow Goodness-of-Fit test. Two tailed p-values of less than 0.05 were considered statistically significant. Statistical analysis was performed by using the statistical software SPSS (version 19, SPSS Inc., Chicago, IL) and "R" (version 3.1.0). Ethical approval. This study has been approved by the Ethics Committee of the Cancer Hospital of Shantou University Medical College, and was performed in accordance with the ethical standards laid down in the 1964