Will Improving Breast Cancer Risk Assessment With AI Lead to Better Outcomes?

The US Preventive Services Task Force recommends that clinicians offer risk-reducing medications to women who are at increased risk of breast cancer and at low risk for adverse medication effects. However, clinical tools used to estimate risk of breast cancer, such as the Breast Cancer Risk Assessment Tool​ and the Tyler-Cuzick Risk Model​, perform only modestly better than random chance, with lower performance in African American and Hispanic women​. The limitations of these tools have stimulated the development of artificial intelligence (AI) models that analyze patterns in screening mammography images​ to more accurately predict breast cancer risk in all populations.

A 2023 systematic review​ of 16 studies found that the median area under the receiver operating characteristic curve (AUROC) for AI imaging tools was 0.72, compared to 0.61 for breast density or clinical risk factor-based tools. Of note, in 7 of 8 studies that compared AI risk prediction to assessments that combined AI with clinical risk factors, the latter did not produce more accurate risk estimates than AI alone.

In a later study​ of an AI model that utilized retrospective data from more than 300,000 patients aged 50 to 70 years who had screening mammograms in the United Kingdom, the overall AUROC was 0.7, with no differences in performance across ethnic or age groups. Similarly, in a recent US study​, a deep learning model involving mammograms from more than 67,000 patients produced an AUROC of 0.71, compared to an AUROC of 0.53 for breast density.

Last year, the US Food and Drug Administration approved an AI test that analyzes a single screening mammogram performed within the past 12 months to classify patients as at average (less than 1.7%), intermediate (1.7% to 3%), or high (greater than 3%) risk of developing breast cancer over the next 5 years. Everlywell, a digital health company, is charging $249 for this test​, which does not require a doctor’s order and is not covered by insurance. Patients are directed to share their report with their primary care clinicians to discuss next steps, which for those at intermediate or high risk could include more intensive screening or risk-reducing medications.

The trouble is, we don’t know if more accurate cancer risk prediction improves patient-oriented outcomes. Previous trials of risk-reducing medications such as tamoxifen and raloxifene used eligibility thresholds based on clinical risk assessments and showed reductions in breast cancer diagnoses, but not breast cancer deaths. Lifestyle modifications such as increasing physical activity, healthy eating patterns, and limiting alcohol consumption also reduce breast cancer risk, but whether women with increased risk will be more motivated to adopt these behaviors is unclear. In the absence of such evidence, AI tools may only increase preventive care costs without producing health benefits.