Abstract P2-10-03: Improving breast cancer risk prediction using a convolutional neural network-based mammographic evaluation in combination with clinical risk factors

Abstract

Abstract Background: Accurate assessment of a woman’s individualized breast cancer risk is necessary to inform shared decision-making regarding screening and risk-reducing strategies. Recently, deep learning techniques, including convolutional neural networks (CNN), have shown better predictive potential for breast cancer risk compared to mammographic density (MD). We evaluated whether combining clinical factors in the Breast Cancer Surveillance Consortium (BCSC) model, including MD, with a novel CNN-based mammographic evaluation more accurately predicts breast cancer risk than the BCSC model alone in a cohort of racially/ethnically diverse women. Methods: We conducted a retrospective cohort study of 23,552 women, age 35-74 years, who underwent screening mammography from 2014 to 2018 at Columbia University Irving Medical Center in New York City. We extracted data from the electronic health record (EHR) on breast cancer risk factors (age, race/ethnicity, prior benign breast biopsy, first degree family history of breast cancer, and MD). From this cohort, we identified 206 women who developed breast cancer by linkage to the tumor registry. We calculated 5-year invasive breast cancer risk using the BCSC model. We applied CNN-based breast cancer risk model to full-field craniocaudal mammographic views of both breasts, with an output of a risk score (range, 0-1). We used logistic regression models with breast cancer status as the outcome and predictors including clinical factors only (BCSC model) or combined with CNN risk score (hybrid model). We compared the prediction performance of these models via area under the receiver operating characteristics curves (AUCs) based on the DeLong test. We also calculated each model’s AUC for subgroups of age and race/ethnicity. Results: Among 23,552 evaluable women, mean age was 55.9 years (standard deviation [SD], 9.5) with 27% non-Hispanic White, 9% non-Hispanic Black, 36% Hispanic, 5% Asian, and 23% Other/Unknown race/ethnicity. Four percent had a first-degree family history of breast cancer, 10% had a prior benign breast biopsy, 45% had heterogeneously or extremely dense breasts on mammography, and 22% met high-risk criteria based upon a 5-year invasive breast cancer risk 1.67% according to the BCSC model. Mean CNN risk score was higher among breast cancer cases compared to unaffected controls (0.477 vs. 0.466, p=0.077). We found that the hybrid model outperformed the BCSC model (AUC of 0.676 vs. 0.640, respectively; p=0.003). In subgroup analyses, the hybrid model more accurately predicted breast cancer risk compared to the BCSC model among women age<50 (AUC of 0.713 vs. 0.645, respectively; p=0.078) and age>=50 (AUC of 0.663 vs. 0.625, respectively; p=0.026); non-Hispanic Black women (AUC of 0.794 vs. 0.663, respectively; p=0.028) and Hispanic women (AUC of 0.666 vs. 0.621, respectively; p=0.060). Conclusion: Among women undergoing screening mammography, a hybrid model incorporating a CNN-based mammography evaluation with clinical factors from the BCSC model more accurately predicted breast cancer risk relative to the BCSC model alone, particularly among racial and ethnic minorities. Combined with clinical risk factors, our CNN model may be used to efficiently predict breast cancer risk and inform risk-stratified breast cancer screening and prevention strategies. Comparing prediction performance for breast cancer risk of the BCSC model vs. hybrid modelBCSC ModelHybrid ModelP-value*AUC95% CIAUC95% CIAll patients0.6400.602-0.6830.6760.640-0.7110.003Age (years)<500.6450.562-0.7280.7130.648-0.7780.078>=500.6250.579-0.6710.6630.623-0.7030.026Race/ethnicityNon-Hispanic White0.6880.623-0.7540.7040.643-0.7660.230Non-Hispanic Black0.6630.539-0.7880.7940.704-0.8830.028Hispanic0.6210.560-0.6860.6660.597-0.7330.060Asian0.6160.440-0.7910.6510.461-0.8410.694 Citation Format: Alissa Michel, Vicky Ro, Julia E McGuinness, Simukayi Mutasa, Richard Ha, Katherine D Crew. Improving breast cancer risk prediction using a convolutional neural network-based mammographic evaluation in combination with clinical risk factors [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P2-10-03.