AI Model Offers Clearer Prediction Than Breast Density After a Negative Mammogram

An AI-based risk model may provide a more accurate picture of a woman’s future breast cancer risk than BI-RADS breast density alone, according to a large 20-year study presented Monday by researchers at the Kaiser Permanente Division of Research. The findings suggest that tools like the Mirai AI model, a deep learning algorithm, could eventually help personalize screening decisions for women with dense breasts.

“Many key clinical decisions in breast cancer care depend on accurate risk models,” said presenter Tejomay Gadgil, MS, a data scientist at Kaiser Permanente. “We were drawn to this work because better risk models have the potential to help patients and clinicians make more informed decisions that lead to better outcomes.”

The study analyzed more than 183,000 women with negative screening mammograms between 2003 and 2022. Using four standard mammographic views, the team generated Mirai risk scores and compared their performance with BI-RADS breast density across short-term (two-year) and long-term (10-year) periods. Over 15 years, nearly 7,000 women developed breast cancer.

Mirai consistently outperformed breast density in predicting both short- and long-term cancer risk. BI-RADS density showed only modest predictive ability, while Mirai demonstrated much higher accuracy. The combined model—Mirai plus density—did not improve prediction beyond Mirai alone, Gadgil noted.

“There has been an assumption that the breast cancer risk AI identifies in mammograms is synonymous with breast density,” Gadgil said. “One of the surprising results of our work is that AI risk is an independent predictor. This suggests there are other risk factors AI is picking up on a mammogram.”

Strong Performance Across All Breast Density Categories

A central question was whether AI risk tools perform consistently in women with dense breasts—the group most often encouraged to consider supplemental screening. In the long-term 10-year window, Mirai’s performance was nearly identical across BI-RADS density categories, suggesting that breast density may be less relevant to long-term cancer risk than once believed.

“In zero to 10 years, Mirai’s performance was similar across breast density,” Gadgil said. “This suggests the risk for future incident breast cancer is independent of breast density.”

This contrasts with extremely dense breasts, which—although long considered a high-risk category—showed only a modest increase in risk in this cohort. Mirai identified roughly twice as many cancers at equivalent thresholds, providing more actionable risk stratification for clinicians and patients.

“We want women with dense breasts to know that clinicians and researchers are aware that they must often make decisions about supplemental screenings,” Gadgil said. “Using AI to help radiologists read mammograms could give patients and their providers more information to guide decisions and improve outcomes.”

Early-Stage Work With Potential to Shape Future Screening

Although the results are compelling, Gadgil said future studies need to test newer algorithms, larger cohorts, and specific patient subgroups to determine whether AI tools should influence recommendations for MRI, ultrasound, or screening intervals.

Improving risk estimation could reduce confusion for women who receive dense breast notifications and help clinicians offer more individualized guidance. “Our goal is to reduce the confusion women experience when they learn they have dense breasts by developing methods to identify those who are at higher risk more accurately,” Gadgil said.

Gadgil hopes RSNA attendees come away with a deeper understanding of the emerging evidence related to AI-based risk models and how these tools may support more personalized screening decisions in the future.

“We hope our work will spark new conversations that provide new perspectives to advance the field,” he said.

Access the presentation, “Mammography AI Risk Score Performance Stratified Across BI-RADS Breast Density Among A Cohort of 183,441 Women with a Negative Screening Mammogram over 15 Years,” (M7-SSBR03-3) at RSNA.org/MeetingCentral.