Currently, screening mammography recommendations are age-based, but a new risk-based screening paradigm is gaining traction. This approach tailors screening initiation, frequency, and the need for supplemental services, aiming to minimize harm for low-risk women and maximize benefits for high-risk women. Over the past few decades, several risk models have been developed to quantify an individual’s risk of developing breast cancer. However, these traditional, questionnaire-based models have achieved only modest performance in discriminating between and predicting future events, in part due to their reliance on self-reported information, which is susceptible to recall bias.
Polygenic risk scores (PRS), which predict breast cancer risk based on the combined effects of common genetic variants, have been shown to enhance disease risk discrimination. Recently, a novel mammographic image-based deep learning model, Mirai, has also shown promise for future breast cancer prediction.
In a study in the British Journal of Cancer, Dr. Shadi Azam, postdoctoral associate in population health sciences, Dr. Rulla Tamimi, chief of the Division of Epidemiology and professor of population health sciences, and colleagues investigated whether the Mirai 5-year risk score outperforms the traditional Gail 5-year risk model, a commonly used questionnaire-based risk assessment tool. They further examined whether incorporating a PRS could improve Mirai’s performance using data from the Nurses’ Health Study II. The Nurses’ Health Study II, led by Brigham and Women’s Hospital, Harvard Medical School, and the Harvard T. H. Chan School of Public Health, is among the largest investigations into the risk factors for major chronic diseases in women. With this data, Dr. Azam, Dr. Tamimi, and colleagues compared the performance of the Mirai 5-year risk score, with and without PRS, to that of the Gail 5-year score.
Researchers found that the Mirai model, which uses information from mammograms, was better at predicting future breast cancer risk than the traditional Gail model, which is based on questionnaire data. When a PRS was added, predictions improved for both models. However, the combination of Mirai and genetic data performed the best overall. This suggests that using advanced imaging tools together with genetic information may provide a more accurate way to assess breast cancer risk than traditional methods alone.
As such, integrating PRS with deep learning models may enhance individualized risk assessment in select clinical settings. These findings align with prior work demonstrating that PRS can enhance predictions from traditional risk models. Though there are challenges to implementing PRS testing, which is not yet reimbursed by most insurers and remains costly, this research supports prospective studies to evaluate the utility and scalability of this approach.
- Highlights