Abstract P2-07-12: Associations between demographic and clinicopathologic characteristics and convolutional neural network-derived breast cancer risk scores among women with hormone receptor-positive breast cancer

Abstract Introduction: Deep learning tools applied to standard surveillance mammograms could serve as noninvasive pharmacodynamic biomarkers of response to adjuvant breast cancer (BC) therapies, including endocrine therapy (ET). We previously demonstrated that a convolutional neural network (CNN)-based model applied to surveillance mammograms is an accurate, independent predictor of BC risk, and that short-term change in a CNN-derived BC risk score with adjuvant ET was also associated with breast cancer relapse. We evaluated demographic and clinicopathologic characteristics associated with baseline CNN risk score and with change in CNN risk score over time on adjuvant ET among women with early-stage HR+ breast cancer. Methods: We conducted a retrospective cohort study among women diagnosed with stage I-III unilateral hormone receptor-positive (HR+) BC at Columbia University Irving Medical Center (CUIMC) from 2007-2017, who received adjuvant ET (tamoxifen, aromatase inhibitors [AIs) and had mammograms of the contralateral breast at diagnosis (baseline) and at 1-2 years on adjuvant ET (follow-up). We extracted demographics, body mass index (BMI, kg/m2), clinicopathologic characteristics including tumor stage and grade, and BC treatments from the electronic health record (EHR) and the New York-Presbyterian Hospital Tumor Registry. We applied the CNN risk model to baseline and follow-up mammograms to estimate baseline CNN BC risk score (range, 0-1, with 1 indicating highest risk), and to calculate absolute change in CNN risk score from baseline to 1-2 years on ET. We conducted multivariable linear regression models to evaluate for potential associations between baseline CNN risk score and change in CNN risk score and prognostic factors including age, race/ethnicity, BMI, tumor stage and grade, and BC treatments. Result: Among 749 evaluable women, mean age was 59.4 years (standard deviation [SD], 12.4 years), and 38% were non-Hispanic White, 14% non-Hispanic Black, 38% Hispanic, and 10% Asian. Thirty-six percent of women were obese (BMI >30 kg/m2), and three-quarters were postmenopausal at diagnosis. The majority of patients had stage I cancer (59.4%), and received adjuvant aromatase inhibitors (68.1%). After adjustment for covariates, baseline BMI was not significantly associated with baseline CNN risk score (regression slope estimate [ß]=0.001; p=0.424). However, increasing age (ß=0.001; p<0.001), Black race (ß=0.023; p<0.001), tumor stage III (ß=0.023, p=0.005) were associated with higher baseline CNN score, while high tumor grade was inversely associated with baseline CNN score (ß= -0.0150; p=0.028). In multivariable analysis, absolute change in CNN risk score from baseline to follow-up was inversely associated with change in BMI over that period (ß= -0.002; p=0.007) and with high tumor grade (ß -0.014; p=0.039), while older age (ß=0.001; p<0.001), Black race (ß=0.024; p<0.001), tumor stage III (ß=0.023; p<0.001), and low tumor grade (ß=0.022; p=0.024), had significant positive associations with change in CNN score. Conclusions: We found that poor prognostic factors such as advanced tumor stage and Black race were associated with higher baseline CNN risk score and increase in CNN score on adjuvant ET. However, factors typically associated with more favorable BC diagnosis, including lower tumor grade, older age, and a decrease in BMI on adjuvant ET, were also associated with higher baseline CNN score and an increase in CNN risk score. These associations should be further evaluated in future studies utilizing larger, diverse patient cohorts of patients with early-stage HR+ BC. Citation Format: Julia McGuinness, Juliet Rowe, Simukayi Mutasa, Vicky Ro, Samuel Pan, Jianhua Hu, Meghna S. Trivedi, Melissa K. Accordino, Kevin Kalinsky, Dawn L. Hershman, Richard S. Ha, Katherine D. Crew. Associations between demographic and clinicopathologic characteristics and convolutional neural network-derived breast cancer risk scores among women with hormone receptor-positive breast cancer [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P2-07-12.