A deep learning model can predict a woman’s risk of developing breast cancer by analyzing changes in her mammograms over time, according to new research published in Radiology. The research, led by researchers at Harvard Medical School, found that the AI-generated risk scores increased gradually among women who later developed cancer, sometimes detectable as early as six years before diagnosis.
The model works by examining the entire mammogram image, rather than focusing on a single feature like breast density. That’s a shift from older risk models, which often rely on limited data points and have trouble accurately predicting the disease in a general screening population.
“We observed clinically relevant differences in risk trajectories between women who did and did not develop the condition,” said Dr. Constance D. Lehman, a professor of radiology at Harvard Medical School and the study’s lead researcher.
Related: Molecule may reverse brain’s Alzheimer’s defenses
The study drew from a large, diverse group of patients. It included women who had screening imaging exams between 2009 and 2019 across six imaging sites, covering urban, community, and rural practices.
After exclusions, the final group included 54,014 women with a median age of 61. Of those, 817 were eventually diagnosed with breast cancer, and 53,197 served as disease-free controls. Each woman contributed one final mammogram and up to six prior annual exams, totaling 158,807 such scans.
The open-source AI tool gave each imaging exam a five-year risk score. It used no demographic, clinical, or historical data — just the image itself.
Among patients with the condition, the median measure climbed from 2.1 in the earliest years of the research to 6.6 at the time of diagnosis. For disease-free women, the ratings stayed flat, ranging from 1.8 to 2.2.
Related: The Healing Power of Shadow Work Journaling for Childhood Trauma
The increase was most dramatic in the two years right before a diagnosis of the condition. “The increase in ratings among those with the condition was detectable as early as six years prior to diagnosis and became more pronounced over time,” Dr. Lehman said.
Most women who develop the condition don’t have a strong family history or known genetic mutations. About 85% of cases are considered sporadic. “These findings demonstrate signals, invisible to the human eye, in the image alone can predict future risk,” she said.
The system’s performance held up across different age groups and breast density levels. That’s important because screening disparities exist across patient populations, and a tool that doesn’t rely on self-reported data could help close that gap.
“A dynamic biomarker approach grounded in the imaging data could mitigate some of these disparities by enabling risk-based personalization,” the lead researcher noted.
Related: Bharat Bhise Looks at The Validity of The Growing Demand for Health Supplements
AI image-based risk scores are already included in the 2026 National Full Cancer Network guidelines. Those guidelines suggest that women with an raised five-year rating — greater than 1.7% — should consider breast MRI starting at age 35, in addition to annual mammography.
An FDA-approved version of this type of scoring system is already in use at some U.S. healthcare institutions.
She compared the approach to how doctors manage high cholesterol or hypertension. “Having a dynamic rating opens up a whole new domain of more effective preventive therapies for the condition,” she said.