Nonsentinel Axillary Lymph Node Status in Clinically Node-Negative Early Breast Cancer After Primary Systemic Therapy and Positive Sentinel Lymph Node: A Predictive Model Proposal

In clinically node-negative (cN0) early stage breast cancer (EBC) undergoing primary systemic treatment (PST), post-treatment positive sentinel lymph node (SLN+) directs axillary lymph node dissection (ALND), with uncertain impacts on outcomes and increased morbidities. We conducted an observational study on imaging-confirmed cN0 EBC, who underwent PST and breast surgery that resulted in SLN+ and underwent ALND. The association among baseline/postsurgical clinic–pathological factors and positive nonsentinel additional axillary lymph nodes (non-SLN+) was analyzed with logistic regression. LASSO regression (LR) identified variables to include in a predictive score of non-SLN+ (ALND-predict). The accuracy and calibration were assessed, an optimal cut-point was then identified, and in silico validation with bootstrap was undertaken. Non-SLN+ were detected in 22.2% cases after ALND. Only progesterone receptor (PR) levels and macrometastatic SLN+ were independently associated to non-SLN+. LR identified PR, Ki67, and type and number of SLN+ as the most efficient covariates. The ALND-predict score was built based on their LR coefficients, showing an area under the curve (AUC) of 0.83 and an optimal cut-off of 63, with a negative predictive value (NPV) of 0.925. Continuous and dichotomic scores had a good fit (p = 0.876 and p = 1.00, respectively) and were independently associated to non-SLN+ [adjusted odds ratio (aOR): 1.06, p = 0.002 and aOR: 23.77, p < 0.001, respectively]. After 5000 bootstrap-adjusted retesting, the estimated bias-corrected and accelerated 95%CI included the aOR. In cN0 EBC with post-PST SLN+, non-SLN+ at ALND are infrequent (~22%) and independently associated to PR levels and macrometastatic SLN. ALND-predict multiparametric score accurately predicted absence of non-SLN involvement, identifying most patients who could be safely spared unnecessary ALND. Prospective validation is required.

patients, also when upfront resectability is possible. 1 The main reasons lie in the higher rates of breast conserving surgery (BCS) in originally BCS-ineligible patients, the reduction in axillary lymph node dissection (ALND) rates, along with an early treatment of micrometastatic disease and the provision of tumor chemosensitivity information. [2][3][4][5][6][7][8][9] Additionally, several new adjuvant escalated or deescalated systemic therapeutic approaches are driven by the achievement (or not) of a pathologic complete response (pCR). 10,11 Likewise, surgical management of the axilla has undergone an important shift toward less radicality due to the important morbidities associated with ALND. The most relevant and frequent is lymphedema, occurring in * 5% of patients undergoing exclusive sentinel lymph node biopsy (SLNB), and 12-25% of patients undergoing ALND, [12][13][14] with incidence increasing up to six times when C 15 lymph-nodes are dissected. 15 Lymphedema is an important comorbidity, affecting different areas of physical and social functioning, bodily pain, and general and mental health. 16,17 Since PST can reduce the ALND rate, it might lower the occurrence of lymphedema. 6,18 At the same time, lymphedema rates after PST, especially neoadjuvant chemotherapy (NACT), have been found to be relatively high if ALND is ultimately performed, with an incidence at 3 years from surgery ranging from 25% to 58.4%, depending also on the diagnostic criterion (e.g., arm circumference increase, symptoms). 18,19 Moreover, a recent prospective study further confirmed that NACT per se is an independent risk factor for lymphedema. 20 For this reason, correctly identifying patients undergoing PST that should not receive postneoadjuvant ALND is an important unmet medical need.
In the case of upfront surgery, several randomized controlled trials (RCT) demonstrated that in patients with small primary tumors (cT1-2) cN0 at diagnosis, when sentinel lymph-node (SLN) are only affected by micrometastases or 1-2 macrometastases, ALND can be safely spared or replaced by radiotherapy (RT), with less comorbilities. [21][22][23][24][25][26] In the neoadjuvant setting, in baseline cN0 cases, SLNB is preferably carried out directly after PST, to obviate the need for two separate surgeries (i.e., SLNB pre/post systemic treatment), facilitate the final definition of axillary pCR, and reduce the sentinel lymph node (SLN)-positive (?) rates without affecting the detection rate. 7 Nevertheless, when metastases in SLN are detected in this context, evidence from the adjuvant setting cannot be easily extrapolated. Consequently, whether ALND could be safely spared in patients with initially cN0 tumors with more or less limited SLN? after PST is currently unknown. For this reason, we sought to identify main features associated with positive nonsentinel axillary lymph nodes (non-SLN?) in cN0 EBC patients who are diagnosed with postneoadjuvant SLN? and develop a multiparametric predictor of negative non-SLN, so to spare unnecessary ALND in this clinical setting.

Patients
A prospectively maintained surgical database at the Hospital Clinic of Barcelona (HCB) including all breast cancer (BC) patients treated starting from January 2013 was consulted. We extracted data from all BC patients who respected all of the following inclusion criteria:

Study Objectives and Statistical Analyses
The objectives of this study were (1) to describe the main clinico-pathological features of patients with EBC with baseline cN0 undergoing PST and showing postneoadjuvant SLN?, to potentially identify characteristics associated to non-SLN? at subsequent ALND, and (2) to develop a predictive model of non-SLN? to avoid potentially unnecessary and harmful ALND.
Continuous variables were described by mean with standard deviation (SD), whilst categorical variables by proportions. Chi-squared and Student's t-tests were carried out to compare the distribution of main clinic-pathological variables in the groups of interest. Preliminary univariate and multivariate logistic regression analyses were performed to detect clinic-pathological factors associated with the presence of residual disease in non-SLN at ALND. Results were reported using odds ratios (OR) with 95% confidence intervals (CI). A logistic least absolute shrinkage and selection operator (LASSO) regression was then run with all clinic-pathological factors to identify variables with non-zero coefficients to include in a new multiparametric predictor score of non-SLN? at ALND (ALNDpredict). 29 Cases with missing values in at least one variable were removed. The minimum k penalizing factor was adopted for the purpose. The estimated coefficients of the selected variables were used to derive an unscaled score which was further scaled 0-100 to obtain the final ALNDpredict score. 30,31 A receiver operating characteristic (ROC) curve analysis was performed to evaluate the area under curve (AUC) with 95% CI of the model and the Youden Index was calculated to identify an optimal cutoff point. 30,31 The AUC gives an indication of the discriminatory performance of the model. An AUC of 0.5 indicates no discriminative performance, whereas an AUC of 1.0 indicates perfect discrimination. 31 A calibration of the ALND-predict continuous and dichotomic models was performed with the Hosmer-Lemershow test with g [ p ?
1. 32 Bootstrap analysis with calculation of bias-corrected and accelerated (BCa) 95% CI was undertaken to validate the predictors in artificial testing cohorts after 5000 resamplings. 33,34 The ALND-predict continuous and dichotomic models were compared using the Akaike Information Criterion (AIC). 35 The model with the lowest AIC (at least 2 AIC unit difference for significance) 35 was selected for further validation.
Significance was set at p \ 0.05 and Holms-Bonferroni corrections were performed to account for multiplicity, when appropriate. All data were analyzed with R vers. 3.4.1 and SPSS vers. 24 for Mac OSX.
The majority of patients (61.1%) received a PST with anthracycline ? taxane-based NACT, while 15.3% received a taxane-based NACT and 23.6% received NET with aromatase inhibitors (AI) or tamoxifen (median duration: 6 months; min-max range: 4-9 months). All patients with HER2? BC also received anti-HER2 agents. After PST, conservative breast surgery was undertaken in 40.3% patients. All patients within this study had a positive SLNB and the mean number of positive nodes was 1.5 (SD ± 0.7), with macrometastases observed in 48.6% of cases (Table 1). Of note, patients receiving NET were more frequently found with macrometastases (82.4% vs. 38.2%) and with no ITC (0% vs. 10.9%), compared with patients receiving NACT (p = 0.006) (Fig. 2). Regional nodal RT after surgery was then administered in 68.1% cases.
In this population, in-breast pCR, meaning absence of infiltrating BC in the surgical specimen, was achieved in 13 (18.1%) cases. After a median follow-up of 46.2 months (95% CI: 41.0-56.6 months), only three events in three different patients were detected in the form of bone metastasis (two in TNBC, one in HR?/HER2-BC) and only in one of these patients additional positive nodes after ALND had been detected.
A separate descriptive evaluation of the HR?/HER2subpopulation undergoing NET is reported in the Supplementary Materials.

Building the ALND-Predict Multiparametric Score
We carried out a LASSO regression and extracted the non-zero coefficients of the variables ultimately identified (i.e., PR levels, dichotomized Ki67, number of SLN?, and type of SLN? involvement) (Fig. 3), with the exception of ''suspicious unconfirmed nodal involvement,'' which was excluded to mitigate nonreproducibility issues [the suspicion mostly relied on ultrasonographic (US) evaluation]. Then, we calculated an unscaled predictive score, according to the following formula: (0.00202 9 PR%) ? (-0.127 9 Ki67 as dichotomic) ? (0.0517 9 number of positive SLN) ? (0.165 9 SLN? type as dichotomic). We further rescaled the score (original values range: -0.075, ?0.4817) to reflect a range from 0 to 100 points and named it the ALND-predict score. The AUC for predicting non-SLN? was 0.83 (95% CI: 0.71-0.96). We also built an alternative model with all LASSO-identified variables. The AUC difference with the ALND-predict was 0.009 (95% CI: -0.031, 0.049) and nonsignificant for De Long p = 0.666. Thus, we retained the four-variable ALND-predict. An optimal cutoff based on the Youden Index was found to be 63, with a sensitivity of 0.75 (95% CI: 0.48-0.93) and a specificity of 0.88 (0.76-0.95) (Fig. 3). The positive predictive value (PPV) was 0.63 (0.44-0.88) and the negative predictive value (NPV) was 0.92 (95% CI: 0.79 0.97). Both ALND-predict as continuous score and dichotomic   Table 2). The variables included in the predictor were excluded to avoid multicollinearity. However, the AUC of the model outperformed the AUC of all its single variables (Supplementary Fig. 1).
On the 5000 bootstrap-adjusted retesting of the multivariate models, estimated BCa 95% CI for the continuous (1.01-1.11) and dichotomic (1.61-319.58) predictor included the adjusted OR in both cases.
In addition, we reapplied the ALND-predict only to the subset of NACT-treated patients to evaluate its performance in this context, as NACT is more frequently carried out than NET. The predictor identified three false positives

DISCUSSION
In the neoadjuvant setting of EBC, defining the optimal axillary management in SLN? patients with cN0 EBC at baseline is a relevant unmet need. Noteworthy, studies involving baseline cN0 with SLN? after postneoadjuvant SLNB have usually included baseline cN? and/or cN0 postneoadjuvant SLN-negative cases, jeopardizing the evidence accumulated so far. 36,37 Furthermore, additional axillary lymph node involvement after NACT has been reported in up to 60% cases. 38,39 Consequently, ALND is usually recommended by most guidelines in this scenario. 1,14,40,41 In this perspective, we investigated non-SLN? rates after ALND in a pure cohort of non-inflammatory cT1-4 cN0 EBC patients with SLN? following PST (either NACT or NET). The mean number of positive nodes was 1.5, with approximately half ITC/micrometastases and half macrometastases. NACT was generally more effective than NET in reducing the rate of macrometastatic SLN. Only increasing PR levels and macrometastatic SLN? were significantly and independently associated with non-SLN? after ALND. Notably, we also observed that SLN with macrometastatic dissemination and a higher number of lymph nodes involved, were more frequently associated to non-SLN? in HR?/HER2-and TNBC. Unexpectedly, the opposite association was observed within the HER2? subgroup. Whether this was related to However, metastatic non-SLN after ALND were found in only 22.2% of patients and only three distant recurrences were detected. Moreover, additional non-SLN? at ALND had only been detected in one case, suggesting that ALND might be spared in a relevant proportion of initially cN0 EBC treated with PST. Notably, tumor relapses in our cohort seemed to be more related to BC IHC-defined molecular subtypes rather than axillary node status (two out of three distant recurrences were in TNBC patients). Nevertheless, the limited cohort prevents us from drawing clear conclusions regarding this aspect.
Our ALND-predict multiparametric score based on PR, Ki67, and SLN? status, was significantly and independently associated, either as continuous (value range 0-100) or dichotomic variable (cutoff of 63), with non-SLN status. More specifically, in patients with SLN? after PST (NACT or NET) and tumor surgery, the odds of presenting non-SLN? at ALND were 6% higher for each unitary increase in the ALND-predict score. Furthermore, a value C 63 was able to identify the majority of non-SLN? patients at ALND and, importantly, almost all negative cases. The sensitivity was 75%, the specificity was 87.5%, the PPV was 63.2%, and the NPV was 92.5%. This means that in case of SLN? EBC patients after PST, the predictor would be able to effectively identify almost all patients with truly negative non-SLN, with only a 7.5% chance of missing patients with non-SLN? if ALND were performed. This is remarkable, also considering the low proportion of non-SLN? observed after subsequent ALND, the high risk of lymphedema and other morbidities, as well as the uncertainties regarding the true impact on local/distant relapse risk reduction, which seems to be relatively low, at least in our cohort. Moreover a 7.5% chance of undetected true negative would be lower than the false negative rate currently considered acceptable for SLN (\ 10%). 42  outperformed all the single variables that it integrated. If results for ALND-predict are further confirmed, its subsequent implementation in the clinic might pave the way toward an ALND-sparing approach in baseline non-inflammatory cT1-4 cN0 EBC treated with PST, and resulting SLN? after tumor surgery and SLNB. Our study presents some limitations. First, it is debated whether non-SLN? is truly prognostically unfavorable in neoadjuvant-pretreated EBC, and we could not properly test this hypothesis in our cohort. Second, SLN involvement was assessed in most cases ([75%) with the One-Step Nucleic Acid (OSNA) technique. 44 Hence, the presence of extracapsular nodal involvement was not uniformly reported and could not be included as a variable in our regression models. Third, the study is retrospective in its nature. Nevertheless, all patients had been consecutively enrolled in a surgical prospective observational database, hence all patients complying with inclusion criteria were included in our study. In addition, despite a reduced number of patients, the population of interest is extremely selected for its characteristics, with no available and detailed data published so far. Only 24-32% of patients with cN0 EBC are SLN? after PST, and results from other studies are mixed with BC patients with other features, such as cN? at diagnosis before receiving PST or negative SLN after PST. 36,37 However, positive or negative SLN after PST may have different prognostic implications, also depending on whether patients were initially classified as cN0 or cN?. At present, the most solid evidence on ALND-sparing approaches are relegated to EBC with cT1-2 cN0 BC undergoing primary surgery and subsequent adjuvant treatments. 21,22,25,26,45,46 Importantly, axillary management in cN0 patients at diagnosis with low SLN tumor burden (ITC/micrometastasis) after PST was subject to controversy in the latest St. Gallen International Consensus. Many panelists felt axillary RT could be an alternative to axillary dissection in such situations. Other panelists urged caution, noting persistent risks of residual axillary nodal involvement, and recommended awaiting the results of ongoing phase III trials. 1,47 A major limitation of ALND-predict is that it was built on a cohort that included HR?/HER2-BC patients treated with NET. This PST approach is not as widespread as NACT and the role of ALND in case of SLN? after NET is even less clear in this scenario. Hence, we exploratorily rebuilt the predictor model by removing NET-receiving patients at the time of performing the LASSO regression for the variables selection, so to obtain an alternative predictor. Nevertheless, we preferred to ultimately retain the original ALND-predict because (1) it was associated with potentially less ALND in a pure NACT-treated population maintaining the false negative rate below the currently accepted cut-off of 10%; 42,43 (2) it took into account different neoadjuvant therapeutic approaches (NACT versus NET) when constructing the model, making it more generalizable to institutions performing NET; and (3) it was built on a higher number of patients. To note, we also reapplied the original ALND-predict only to patients undergoing NACT, confirming a false negative rate of \ 10%.
Finally, we had no validation cohort for the ALNDpredict. However, we derived the multiparametric model from centrally reviewed parameters (all assessed at the HCB), with an optimal nodal staging to define cN0 at diagnosis (clinical, US-, and MRI-based), with a robust statistical methodology, and provided an internal validation with a 5000 bootstrap-adjusted retesting of both the continuous and dichotomic predictors, which is a widely accepted methodology for validating logistic regressionbased predictive models in the absence of external cohorts. 33,49,50 While this does not necessarily imply the possibility of a direct clinical implementation of the predictor, it highlights a remarkably stable accuracy in detecting patients to whom ALND should not be offered. Furthermore, the model is easily appliable to almost every clinical practice scenario, being cheap and based on parameters routinely available in the setting of interest. Therefore, ALND-predict merits further prospective validation.  OPEN ACCESS This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons. org/licenses/by/4.0/.