How AI estimates five-year breast cancer risk from mammograms

Mammograms are designed to detect breast cancer that may already be present. But the same images can now help estimate five-year breast cancer risk, even when a radiologist identifies no suspicious findings. Artificial intelligence can analyze characteristics of breast tissue on a mammogram and calculate the probability of a breast cancer diagnosis over the following five years.

As of October 2026, this type of assessment has become available to consumers across the United States through the digital health platform Everlywell. Women with a qualifying existing mammogram can request a risk assessment without undergoing another imaging examination specifically for the algorithm.

The technology is called Clairity Breast. Its underlying software, Allix5, received authorization from the US Food and Drug Administration in May 2025 following a review of clinical performance data.

The distinction between this technology and conventional medical imaging AI is fundamental. Some AI systems help radiologists identify potentially malignant abnormalities. Clairity Breast addresses a different question: how likely is a woman to receive a breast cancer diagnosis in the coming years?

What can AI learn from a mammogram?

A mammogram captures the arrangement of fatty, connective, and glandular breast tissue. These structures differ substantially between individuals.

Breast density has long been recognized as a risk factor for breast cancer. Dense tissue is associated with a moderately increased likelihood of disease and can also make existing cancers harder to detect on mammograms.

Yet density represents only one characteristic of the image.

Deep learning algorithms can examine far more complex combinations of visual features, including patterns that are difficult to describe using conventional radiological terminology.

These systems are trained using mammograms linked to subsequent health records. Researchers identify which women developed breast cancer during follow-up and which did not. The algorithm learns statistical relationships between features of the original images and later diagnoses.

When applied to a new mammogram, the trained model produces a risk estimate based on patterns learned from those previous cases.

It does not identify a microscopic cancer that is guaranteed to become visible in the future. Nor does it necessarily reveal the biological reason why a particular woman's risk is elevated. In many deep learning systems, the precise image features driving an individual prediction are not directly interpretable.

The distinction matters: predicting the probability of a diagnosis is not the same as detecting a cancer years before it develops.

What the FDA evaluation actually found

The pivotal evaluation supporting the authorization of Clairity Breast was a retrospective, multicenter study involving 77,511 mammographic examinations from 44,112 women at five US clinical sites.

Investigators compared risk predictions generated from the original mammograms with subsequent breast cancer outcomes.

Discrimination was measured using the area under the receiver operating characteristic curve, or AUC. A value of 0.5 indicates performance equivalent to random ranking, while 1.0 represents perfect separation of cases and noncases.

The overall AUC for Allix5 was 0.70.

When investigators excluded cancers identified shortly after an abnormal index screening examination, the AUC decreased to 0.66.

This distinction is important because the second analysis more directly addresses risk prediction in women without cancer already being identified through the initial screening process.

The model retained predictive information, but its accuracy does not support treating an individual score as a reliable forecast of whether a particular woman will develop cancer.

The FDA also identified specific limitations.

Predictions above 5% require caution because the model may overestimate risk at higher scores, particularly above 10%. Effectiveness was not demonstrated in women with breast implants.

The software also supports only certain directly acquired two-dimensional digital mammography images from specified equipment. It cannot be applied indiscriminately to every mammogram.

Official FDA decision summary: https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN240047.pdf

What a large scientific study found

Further evidence for mammography-based risk prediction was published in JAMA Network Open in May 2026.

Researchers at Massachusetts General Hospital and Harvard Medical School evaluated 123,091 screening mammograms from 67,019 women across five sites within one academic health system. Follow-up data allowed them to examine breast cancer diagnoses within five years.

Importantly, this study evaluated Mirai, a different deep learning model. Its performance figures should not be attributed directly to the commercial Clairity Breast product, although both approaches use mammographic images to estimate future risk.

The investigators divided examinations into low-, intermediate-, and high-risk categories using model-generated predictions.

Breast cancer was diagnosed within five years following 1.0% of examinations assigned to the low-risk group, compared with 6.2% in the high-risk group.

That was more than a sixfold difference in observed cancer incidence between the two groups.

The model was also compared with breast density, a conventional factor used in decisions about supplemental breast imaging.

Mirai achieved an AUC of 0.71 for predicting future breast cancer, compared with 0.53 for breast density alone.

The results suggest that information captured by image-based AI extends beyond the conventional density categories reported by radiologists.

For example, some women with nondense breasts were still classified as high risk by the algorithm. This indicates that breast density alone may fail to identify individuals whose mammograms contain other risk-associated features.

There are important qualifications.

The analysis was retrospective and involved facilities within a single healthcare system using imaging equipment from one manufacturer. Its findings require prospective evaluation in broader clinical settings.

Furthermore, the study established predictive discrimination, not improved patient outcomes from AI-guided clinical decisions.

One of the authors, Constance Lehman, disclosed consulting fees and equity in Clairity. This relationship is relevant when considering the wider development and commercialization of breast cancer risk prediction technology.

From a risk score to a clinical decision

Traditional breast cancer risk assessment considers factors such as age, family history, reproductive history, inherited genetic variants, and breast density.

AI analysis of mammograms offers another potential source of information.

A five-year risk estimate of 3%, for example, refers to an expected rate of approximately three breast cancer diagnoses among 100 women with comparable risk characteristics over five years. It does not mean that cancer is currently present or that the person receiving the estimate will necessarily develop the disease.

A higher risk assessment might prompt a discussion about additional imaging, such as breast MRI, or established risk-reduction strategies.

However, better prediction does not automatically translate into better health outcomes.

To establish clinical value, researchers must demonstrate that using these predictions to guide care actually improves outcomes, such as reducing diagnoses at advanced stages or preventing breast cancer deaths.

Overestimated risk can result in anxiety, unnecessary testing, and potentially unnecessary preventive interventions. Underestimated risk may create false reassurance and discourage appropriate surveillance.

The FDA therefore specifies that Allix5 is not intended to detect or diagnose breast cancer, provide treatment recommendations, or replace professional interpretation of screening mammograms.

Its output is meant to supplement clinical assessment, alongside a patient's medical history and other established risk factors.

Through Everlywell, eligible US consumers can now request a Clairity Breast assessment using a compatible screening mammogram obtained within the previous 12 months. Results are reviewed within a clinician-supported process and accompanied by information intended to guide discussions about screening and prevention.

This creates a new challenge for clinical medicine: determining when a statistically meaningful prediction should change what happens to the patient.

The same mammogram can now contribute to two different assessments. One addresses what a radiologist can identify today. The other offers additional information about the probability of disease over the years ahead.

Whether the second assessment ultimately improves outcomes will depend on what clinicians and patients can usefully do with that information.