Abstract P2-01-18: Preoperative prediction of axillary lymph node status based on artificial neural network algorithm model

Abstract

Abstract Background Axillary lymph node status remains the single most significant prognostic factor for patients with primary breast cancer. Clinical and pathologic data have been used to develop statistical models to predict the axillary nodal status; the diverse accuracy may reflect the complexity of factors related to axillary metastasis. Artificial neural network (ANN) is a computational method proposed as a supplement to standard statistical models for predicting complex biological phenomena. ANN is composed of artificial neurons and interrelated by synaptic weights, effective in multifactorial analysis and has the ability to explore underlying nonlinear relations of interconnected variables. The aim of this study was to create an ANN-based preoperative decision tool for prediction of nodal axillary status (N0, N+ with 1-3 positive lymph nodes and N+ with ≥ 4 positive nodes). In the clinical setting, this may contribute to improved selection of patients for no axillary staging for those predicted with disease-free axilla (N0), sentinel node biopsy for patients with predicted 1-3 nodal metastases, and axillary lymph node dissection or neoadjuvant therapy for patients displaying four or more involved axillary lymph nodes. Methods The cohort constituted of consecutive patients diagnosed with primary breast malignancy between January 2009 and December 2012 at Skåne University Hospital in Lund, Sweden. The exclusion criteria were palpable axillary nodes or cytology-verified axillary metastasis and neoadjuvant chemotherapy. Data on mode of detection were retrieved. Clinical parameters included age, BMI, menopausal status and the location of the tumor within the breast. A breast pathologist extracted the histopathological variables of the tumor and lymph node. The ANN consisted of 3 layers, an input layer using 1-22 separate variables, one hidden layer, and a single node output layer, with the nodal status N0, macro-metastases N+1-3 and N+≥ 4, respectively as output. The ANN was trained by back-propagation using gradient descent of a cross entropy error, for 2000 epochs. Area under the receiver operating characteristic curve (AUC) was used to assess the performance of the ANN-based predictive models for axillary nodal status. Evaluation was performed in a stratified 5-fold cross validation scheme, repeated 10 times. Sensitivity and specificity are given for a cutoff value corresponding to optimal balanced accuracy. Results The cohort consisted of 800 patients, classified into N0 64 % (n=514), N+1-3 positive nodes 29% (n=232) and N+≥ 4 positive nodes 7% (n=54). The AUC was 0.72 for prediction of node negativity; sensitivity 74% and specificity 64%. AUC was 0.77 for N+ ≥ 4; sensitivity 67% and specificity 79%. The predictive model for N+1-3 macro-metastases is in progress. Tumor size and lymphovascular invasion are the two principal risk variables selected to construct the ANN predictive model for N+ ≥ 4. However, for N0, the predictive model is characterized by a complicated integration of numerous clinicopathological risk variables. Conclusions Based on clinicopathological and mammography-screening data, ANN can be valuable in predicting axillary nodal status and as a guidance tool for directing patients to personalized treatment of the axilla. Citation Format: Dihge L, Edén P, Rydén L. Preoperative prediction of axillary lymph node status based on artificial neural network algorithm model [abstract]. In: Proceedings of the 2016 San Antonio Breast Cancer Symposium; 2016 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2017;77(4 Suppl):Abstract nr P2-01-18.