Artificial neural network models to predict nodal status in clinically node-negative breast cancer

Sentinel lymph node biopsy (SLNB) is the standard axillary staging procedure for patients with clinically node-negative primary breast cancer. In the majority of patients, SLNB will prove negative and no nodal metastasis is diagnosed [1]. Moreover, in approximately half of the SLNB-positive cases, no further metastatic lymph nodes will be harvested during routine completion axillary lymph node dissection (ALND) [2]. While a prospective randomized trial [3] questioned the value of axillary surgical staging in selected low-risk patients, the American College of Surgeons Oncology Group (ACOSOG) Z0011 trial suggested that patients with 1–2 sentinel node metastasis were eligible for minimalistic axillary surgical interventions without completion ALND, and reported no negative consequences for survival or locoregional recurrence after 10 years of follow-up [4, 5]. However, patients with heavy-burden axillary disease (stage N2) could benefit from preoperative selection for neoadjuvant therapy or direct ALND. An improvement in breast cancer management has been to seek an individualized surgical approach to the axilla, and an accurate prediction of the axillary status preoperatively would facilitate individualized surgical decisions.

However, validation of prediction models for nodal status have shown diverse accuracies in estimating nodal involvement [6, 7] and mirror the complexity of factors related to axillary metastasis, and the paucity of models to analyze nonlinear dynamics between relevant variables. Artificial neural networks (ANNs) are nonlinear machine learning methods proposed as supplements to standard statistical models for predicting multifaceted biological events [8, 9], and help in the exploration of underlying nonlinear interactions of interconnected predictors [10]. ANNs have gained utility in various clinical settings, and are being used as diagnostic and prognostic tools in cancer [11, 12], and for prediction of surgical outcomes in various disease conditions [13, 14].

The primary aim of this study was to utilize commonly available patient-related and clinicopathological characteristics in ANN modeling to predict nodal axillary status. The end-points were chosen to reflect the extent of nodal metastatic burden, with an aim to designate no lymph node metastasis (N0), metastases in 1–3 lymph nodes (N1) and metastases in ≥ 4 lymph nodes (N2), respectively. A secondary aim was to assess possible clinical benefit in detecting disease-free axilla (N0). Patient stratification preoperatively using the ANN model applying nodal predictive variables would help identify patients least likely to benefit from SLNB, consequently reducing the rate of unbeneficial surgery. In the clinical setting, the models may be useful tools for risk-benefit analysis of axillary treatment and contribute to improved patient stratification for surgical axillary interventions.

The current study presents ANN-based models for the prediction of nodal status based on routinely available clinicopathological characteristics, in a cohort of primary breast cancer patients consecutively and prospectively included in a pathology database. Internally validated performances displayed AUCs ranging from 0.705–0.747, with good calibration. These models highlighted the utility of nonlinear assessments of clinical characteristics and histological variables for prediction of axillary nodal status, especially in distinguishing disease-free axilla (N0) from high-burden disease (N2).

ANN models have been useful in detecting nodal metastasis on histopathological slides [22], and in evaluating risk of non-sentinel node involvement in breast malignancies [23]. Previous studies have proposed ANN-based algorithms for predicting nodal metastasis [24‐26], though these studies were either based on small sample sizes or were conducted in selected patient cohorts. To the best of our knowledge, this is the largest study to present ANN-based algorithms predicting the extent of nodal metastatic burden in a population-based, contemporary, breast cancer cohort.

To predict nodal metastasis, our model integrates a complex set of input variables which reflect the multifactorial nature of the axillary metastatic process [27]. Variables in ANN-based models should not be taken for independent since the cause and effect reflect a dynamic process. Nevertheless, an attempt was made to better comprehend the importance of each variable by sensitivity analysis. While mean odds ratios were used for simplicity, the corresponding percentiles emphasized the dynamic nature of the input variables.

The present results reinforced tumor size [28] and LVI [29] as the most significant predictors of axillary metastasis. Age was significant in predicting disease-free axilla and low-burden disease. A nonlinear association between age and nodal status has previously been shown, with a low probability of nodal metastasis in those aged< 70 years, and increased probability in those aged > 70 years [30, 31]. In this study, positive ER and PR status was predictive of nodal metastasis, in agreement with literature; the TNBC subtype, although more aggressive, infrequently metastasizes to the axilla [32]. While a negative PR status has been shown to independently lower the risk of nodal metastasis [33], Ki-67 positivity has been associated with nodal metastasis [34, 35], and the present results are in agreement. Interestingly, alterations in the distribution of the breast cancer intrinsic subtypes has been reported to occur from the premenopausal state to the postmenopausal state [36, 37]. It is also noteworthy that menopausal status, in addition to hormone receptor status and age, was predictive of disease burden. Some publications have suggested a higher proportion of nodal metastasis in lobular cancer than that in the breast carcinoma of no special (ductal) type [38]; however, others have either found no significant differences [39], or have implied a lower incidence of nodal metastasis in the lobular than that in the ductal type [40]. The present results revealed a nonlinear association between histological type and nodal status. As supported by previous reports, the upper-outer quadrant localization of the tumor was the most common [41], and multiple lesions were predictive of axillary metastasis [33]. In accordance with published data [42], a tumor location in the inner quadrants, in comparison with that in the upper-outer quadrants, was predictive of disease-free axilla. However, medial tumor localization has also been related to increased risk of relapse [43]. Differences in lymphatic drainage patterns from the breast have been reported between palpable and nonpalpable lesions [44]. Of note, about 63% of the cases in the current cohort were diagnosed by mammography screening, and mode of detection was a significant factor in the prediction of axillary metastasis. Although the value of mammography screening is extensively debated [45], the current findings supported the notion that mode of detection complements information on tumor features and biology [46].

The ACOSOG Z0011 trial results [4, 5], were supportive of less extensive axillary surgery in patients with 1–2 metastatic sentinel nodes [47], and underlined the importance of distinguishing between low- and high-burden metastatic involvement; accordingly, ALND was omitted in women with cT1-2 N0 disease, and those with < 3 positive sentinel nodes underwent breast-conserving therapy with adjuvant treatments. As with the ACOSOG Z0011 trial, a negative clinical axillary status was a criterion in the present study. However, eligibility criteria for the current study were independent of the surgical intervention to the breast (mastectomy or breast-conserving surgery).

Predicting low-burden disease (≤3 metastatic nodes) was more challenging than was predicting disease-free axilla or high-burden disease. Nevertheless, identifying presence of metastatic burden is valuable. On an average, 2–3 lymph nodes are removed if SLNB alone is performed for nodal staging [48] and most metastatic nodes are identified with the excision of the first three sentinel nodes [49]. To improve the accuracy of predicting 1–3 metastatic lymph nodes, inclusion of imaging features from magnetic resonance imaging into the model might be beneficial. However, MRI is not always available in the preoperative setting whereas the chosen predictors in our model are. An accurate preoperative prediction of ≤3 metastatic nodes could provide clinicians with important information supporting SLNB staging procedure but spare a majority patients from completion ALND. For patients predicted to have N2 disease, the option of neoadjuvant therapy or upfront ALND can be discussed with the patient. The proportion of patients with node-positive disease is declining, and alternative non-invasive methods to surgical staging are increasingly being explored and our prediction model of N0 aims to add knowledge in this field. With a cut-off at maximized NPV to identify those with disease-free axilla, 7.50% of patients would be spared unnecessary SLNB. In comparison, the reported FNR for the SLNB procedure has been 5–10% [1, 48] and applying these FNR cut-offs for current N0 prediction would bring the SLNB reduction rate to 17.75–27.25%. While adopting the clinically accepted false negative rate of 10% for the SLNB procedure, nearly one third of all node negative patients with a predicted N0 by the model could have been spared a surgical staging procedure.

The present study has several limitations. Besides its retrospective nature, the models were developed from a single-center cohort. Furthermore, high-burden axillary metastasis was uncommon, which impacts the generalizability of the outcome. However, the cohort originated from a prospectively maintained database, which represents a contemporary population with access to a well-established public mammography-screening program. On the other hand, the possibility to obtain information on the risk variables from a monitored source strengthened the study. Unlike results relying on diverse registries, all histopathological characteristics analyzed in the study were managed by a single breast pathologist, which helped to minimize inaccuracies. However, the optimal preoperative utility of the models requires key variables such as LVI, which may not always be achievable on core-needle biopsy [50]. Although meticulous internal validation was performed with results supporting model robustness, further external validation in an independent cohort is necessary to confirm the utility of the models as guidance tools.

The current study showed that nodal status is related to several independent patient and tumor characteristics, and that a nonlinear association exists between preoperatively obtainable clinicopathological variables and degree of axillary metastatic involvement. ANN models proved especially favorable in distinguishing high-burden disease and disease-free axilla and could thus be useful as a clinical decision tool in the preoperative setting imputing selected risk variables. If a threshold for classification of node-negativity were applied for high NPV and low FNR, individuals with a very low probability of axillary disease would not have been selected for SLNB by the model, and would be spared from unbeneficial axillary surgery.