Purpose: Neoadjuvant systemic treatment elicits a pathologic complete response (pCR) in an average of 35% of women with breast cancer. In such cases, breast surgery may be considered overtreatment. However, imaging and vacuum-assisted biopsy (VAB) alone showed high rates of missed cancer compared to standard breast surgery. We therefore evaluated multivariate algorithms using patient, tumor, and VAB variables to accurately identify patients with breast pCR.

Methods: We developed and tested three multivariate approaches: elastic net regression, Support Vector Machines (SVM), and a deep neural network. We analyzed 452 patients, randomly partitioned into training and test samples (2:1 ratio), who participated in three prospective studies assessing the feasibility of VAB to accurately detect residual disease after neoadjuvant systemic treatment (NST). The studies were conducted at 23 sites in the United States, Germany, and South Korea. The trials enrolled women who presented with clinical stage I-III breast cancer of any biological subtype and a partial or complete response to NST confirmed by ultrasonography, mammography, or magnetic resonance imaging; all patients underwent guideline-adherent surgery. We compared the performance of the multivariate algorithms to the histopathologic evaluation of disease response in the surgical specimen (reference standard) - false-negative rate (FNR, missed residual cancer) and specificity (identification of breast pCR) were the main outcome measures. The best performing algorithm on the test set with respect to sensitivity and specificity was validated using data of an independent fourth trial. We compared the performance of the multivariate approaches to the performance of imaging and/or VAB.

Results: In the test set (n=152), elastic net regression, SVM and the neural network revealed an FNR of 1.2% (1 of 85 patients with missed residual disease). Specificity of the elastic net regression was 46.3% (31 of 67 women with surgically confirmed breast pCR identified), of the SVM 62.7% (42 of 67) and of the neural network 67.2% (45 of 67). All multivariate algorithms performed better than imaging or VAB: FNR 25.9% (22 of 85) and 16.5% (14 of 85), respectively. Subsequent external validation (n=50) of the neural network algorithm showed a false-negative rate of 0% (0 of 27) and a specificity of 65.2% (15 of 23). The area under the ROC curve for the deep neural network was 0.97 (95% CI, 0.94 to 1.00). Analyzing the coefficients of the elastic net regression (regularized beta; ß) showed that the lesion diameter on imaging after NST (ß = 0.31) and VAB results (ß = 0.49) were the most important variables in the prediction of residual tumor. Other variables were also important: age (ß = 0.18), in-situ in the initial diagnostic (not VAB) biopsy (ß = 0.11), difficulties during the pathologic evaluation of the VAB specimen (ß = 0.11); needle size 7G (ß = -0.06, as opposed to 8G, 9G, 10G), multicentricity on imaging after NST (ß = 0.06), hormone-receptor positivity (ß = 0.01), and a clip marker positioned within the (former) lesion (ß = -0.01, as opposed to a clip marker positioned <5mm or >5mm from the lesion

Conclusion: A multivariate algorithm can accurately select breast cancer patients without residual disease after neoadjuvant treatment. This finding may pave the way to study omission of breast surgery in these patients in the future.

Performance of multivariate algorithms compared to imaging and vacuum-assisted biopsy
False-negative rate - value (95% CI)Specificity - value (95% CI)Negative predictive value - value (95% CI)Positive predictive value - value (95% CI)
Test set (n=152)     
Imaging 25.9% (17.0-36.5%) 61.2% (48.5-72.9%) 65.1% (52.0-76.7%) 70.8% (60.2-79.9%) 
VAB 16.5% (9.3-26.1%) 89.6% (79.7-95.7%) 81.1% (70.3-89.3%) 91.0% (82.4-96.3%) 
Imaging + VAB 5.9% (1.9-13.2%) 52.2% (39.7-64.6) 87.5% (73.2-95.8%) 71.4% (62.1-79.6%) 
Elastic net regression 1.2% (0.0-6.4%) 46.3% (34.0-58.9%) 96.9% (83.8-99.9%) 70.0% (61.0%-78.0%) 
Support Vector Machine 1.2% (0.0-6.4%) 62.7% (50.0 - 74.2%) 97.7% (87.7-99.9%) 77.1% (68.0-84.6%) 
Deep Neural Network 1.2% (0.0-6.4%) 67.2% (54.6-78.2%) 97.8% (88.5-99.9%) 79.3% (70.3-86.5%) 
Validation set (n=50)     
Deep Neural Network 0.0% (0.0-12.8%) 65.2% (42.7-83.6%) 100% (78.2-100%) 77.1% (59.9-89.6%) 
Performance of multivariate algorithms compared to imaging and vacuum-assisted biopsy
False-negative rate - value (95% CI)Specificity - value (95% CI)Negative predictive value - value (95% CI)Positive predictive value - value (95% CI)
Test set (n=152)     
Imaging 25.9% (17.0-36.5%) 61.2% (48.5-72.9%) 65.1% (52.0-76.7%) 70.8% (60.2-79.9%) 
VAB 16.5% (9.3-26.1%) 89.6% (79.7-95.7%) 81.1% (70.3-89.3%) 91.0% (82.4-96.3%) 
Imaging + VAB 5.9% (1.9-13.2%) 52.2% (39.7-64.6) 87.5% (73.2-95.8%) 71.4% (62.1-79.6%) 
Elastic net regression 1.2% (0.0-6.4%) 46.3% (34.0-58.9%) 96.9% (83.8-99.9%) 70.0% (61.0%-78.0%) 
Support Vector Machine 1.2% (0.0-6.4%) 62.7% (50.0 - 74.2%) 97.7% (87.7-99.9%) 77.1% (68.0-84.6%) 
Deep Neural Network 1.2% (0.0-6.4%) 67.2% (54.6-78.2%) 97.8% (88.5-99.9%) 79.3% (70.3-86.5%) 
Validation set (n=50)     
Deep Neural Network 0.0% (0.0-12.8%) 65.2% (42.7-83.6%) 100% (78.2-100%) 77.1% (59.9-89.6%) 

Citation Format: André Pfob, Chris Sidey-Gibbons, Han-Byoel Lee, Marios Konstantinos Tasoulis, Vivian Koelbel, Michael Golatta, Gaiane M. Rauch, Benjamin D. Smith, Vicente Valero, Fiona MacNeill, Wonshik Han, Walter Paul Weber, Geraldine Rauch, Henry Kuerer, Joerg Heil. Identify breast cancer patients with pathologic complete response in the breast after neoadjuvant systemic treatment - an international, multicenter analysis [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS2-42.