AI Analysis of Mammograms Predicts 10-Year Breast Cancer Risk Better Than Questionnaire-Based Clinical Tools

A new artificial intelligence model can analyze standard mammograms to estimate a woman’s 10-year risk of developing breast cancer more accurately than currently used clinical tools, according to a large international study. The advance could help more women at high risk qualify for preventive measures—such as lifestyle changes or risk-reducing medications—without needing extra questionnaires or genetic tests. The research, led by Mikael Eriksson at Karolinska Institutet and the University of Cambridge, was published today in Science Translational Medicine.

A team of researchers from Sweden, the UK, and the US trained an AI system using digital mammograms from the KARMA cohort in Sweden, then validated its performance in two separate population-based cohorts: KARMA (internal validation) and Olmsted County, Minnesota (external validation), together spanning nearly 8,700 women followed for up to 12 years. A further hospital-based validation was performed in the EMBED dataset from Atlanta.

The image-based AI model—designed specifically for long-term risk prediction of both invasive and in situ breast cancer—showed consistent ability to discriminate who would develop the disease. For invasive breast cancer, the model achieved a 10-year area under the curve (AUC) of 0.72 in both the Swedish and US validation cohorts. In the Swedish KARMA validation cohort—the only setting where Tyrer-Cuzick and BCSC scores were available—it showed a statistically significant improvement over those widely used clinical risk calculators (AUC 0.64 and 0.66, respectively). It also outperformed Mirai, an earlier AI model built for short-term risk, in all validation cohorts.

When applied according to US clinical guidelines that classify women with a 10-year risk of 6% or higher as high-risk, the AI tool identified 41% of invasive breast cancers in the KARMA validation group—compared with just 15% using Tyrer-Cuzick and 5% using BCSC. In the top 10% of women at highest risk, the AI model captured 33% of all future breast cancers, while Tyrer-Cuzick captured 23%, BCSC 20%, and Mirai 24%. The model was well calibrated, with observed breast cancer numbers closely matching predicted numbers in both the Swedish and US cohorts (expected-to-observed ratios near 1.0).

The tool performed similarly in white and Black women in the Atlanta cohort, though the researchers caution that the Swedish training population was predominantly of European ancestry, and further evaluation in more diverse populations is needed before clinical use. Another limitation: the model significantly overestimated risk in the very lowest-risk group in Olmsted County, and its performance for in situ cancers varied between settings, likely reflecting the lower proportion of in situ cases in the Swedish cohort used for training.

Nevertheless, the findings suggest that AI-based mammogram analysis could be integrated directly into existing breast screening programs, potentially identifying women who could benefit most from preventive therapies without the logistical barriers of collecting detailed family history and lifestyle data. With breast cancer incidence rising globally, better risk stratification could make prevention strategies more effective and cost-efficient.

"We show that an image-derived AI risk model can identify up to 40% of breast cancers as high-risk at study entry using clinical guidelines," the authors write, adding that the model "was well calibrated and showed significantly better discriminatory performance and selection of high-risk individuals than the Tyrer-Cuzick-v8 and BCSC-v3 risk tools used in clinical practice."

The team plans further studies to understand what mammographic features the AI is detecting and to test the model in more diverse screening settings. For now, the work strengthens the case for using AI to move breast cancer screening from a one-size-fits-all approach to a more personalized prevention strategy.

Reference: Eriksson M, Czene K, Scott C, et al. A long-term image-derived AI-based risk model for primary prevention of breast cancer in individuals at high risk. Science Translational Medicine. 2026;18(850). DOI: 10.1126/scitranslmed.ady7414