Performance of an image-only deep learning breast cancer risk model with the addition of a polygenic risk score.
Authors
Affiliations (10)
Affiliations (10)
- Department of Population Health Sciences, Weill Cornell Medicine, New York, USA. [email protected].
- Division of Breast Imaging, Department of Radiology, Mass General Brigham, Boston, USA.
- Departments of Nutrition and Epidemiology, Harvard T H Chan School of Public Health, Boston, USA.
- Channing Division of Network Medicine, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, USA.
- Department of Surgery, Emory Winship Cancer Institute, Atlanta, Georgia, USA.
- Division of Surgical Oncology, Mass General Brigham, Boston, USA.
- Transdivisional Research Program, Division of Cancer Epidemiology and Genetics, National Cancer Institute, National Institutes of Health, Rockville, MD, USA.
- Department of Epidemiology, University of Washington, Seattle, WA, USA.
- Public Health Sciences Division, Fred Hutchinson Cancer Center, Seattle, WA, USA.
- Department of Population Health Sciences, Weill Cornell Medicine, New York, USA.
Abstract
Mammograms contain imaging biomarkers that can predict future breast cancer risk using deep learning (DL) models. We evaluated whether adding a polygenic risk score (PRS) improves performance of the image-only DL breast cancer risk model Mirai. This nested case-control study within the Nurses' Health Study 2 included 902 women (270 cases, 632 controls) who underwent bilateral 2D digital screening mammography between 2001-2017. Risk was assessed using Mirai and, for clinical comparison, the Gail 5-year model. A PRS was calculated using 313 breast cancer-associated single-nucleotide polymorphisms. The primary outcome was incident breast cancer within five years of the index mammogram. Discrimination was evaluated using area under the receiver operating characteristic curve (AUC), with comparisons using the DeLong test. Mean age was 55.5 years(SD 5.3). Among cases, median time from index mammogram to diagnosis was 2.0 years (IQR0.5-4.0). Mirai alone achieved an AUC of 0.66 (95% CI: 0.62-0.70), increasing to 0.73 (95% CI 0.67-0.78; Pā=ā0.05) with PRS. The Gail model improved from 0.52 (95% CI: 0.47-0.57) to 0.69 (95% CI: 0.62-0.76; Pā<ā0.001) with PRS. Mirai+PRS significantly outperformed Gail+PRS (Pā<ā0.001). Integrating PRS with DL-based mammographic models modestly improves risk discrimination and may enhance personalized screening.