What Impact Will AI Have on Traditional Breast Cancer Risk Calculators?

MammoNavigate · Editorial Analysis · Richard D. Lippert Jr., President & Founder, Mammologix · Breast Imaging Operations since 1995

Artificial intelligence can now directly assess the risk of breast cancer from a mammogram, and the evidence supporting this is quite robust. However, this new technology won't replace the traditional Gail or Tyrer-Cuzick risk assessment scores. Instead, it will complement them, raising a crucial question: who will oversee the increasing amount of risk information, and how will it be managed?

---

Ask what artificial intelligence will do to the breast cancer risk calculators that programs have leaned on for thirty years, and the expectation usually runs to replacement. A better algorithm arrives, the old questionnaire retires, and the score gets sharper. That is not what the evidence points to. The Gail and Tyrer-Cuzick models are not going to be switched off. They are going to be joined: first by a risk number the mammogram can now produce on its own, then by others.

The impact of AI here is not a single cleaner answer. It is more information, arriving more often, that someone in the program must receive, reconcile, and act on.

What the questionnaire actually costs

Calculating a traditional risk score is a bit like putting together a puzzle. A staff member gathers information about a patient's family history, reproductive history, and past biopsies, and then uses a calculator to come up with a score. They also look at a density category from the radiologist's report. But here's the thing: none of this information stays the same. When a patient comes in for a visit, they might remember something about their family history that they didn't recall last time. For example, they might suddenly remember that a relative had cancer, and how old they were when they got it. Or maybe they'll forget to mention a biopsy they had a few years ago. And to make things even more complicated, a different radiologist might read the patient's film and come up with a different density category.

So, every time the patient comes in, the staff member must rebuild the risk score from scratch, using new information that might not match what they had last time. The score gets recalculated, by hand, every time the patient returns, using inputs that might quietly disagree with what was used last year.

That costs in two ways, and programs usually see only the first. The visible cost is labor. Someone collects the history, someone keys it in, someone reconciles this year's answers against the chart when they do not match, and someone re-validates a number that drives a real decision. The hidden cost is the decision itself. A family-history field entered inconsistently, or a density read that drifted half a category, can move a woman across the line that decides whether she is offered supplemental MRI or sent for genetic counseling. When the input is wrong, the eligibility is wrong, and no one in the room knows it happened. That is the cost that never appears in a workflow diagram: the woman who should have been flagged and was not, because the score meant to flag her was built from numbers that would not stay put.

It's the starting point, and it comes with a hefty price tag, a lot of manual labor, and it's already a challenge to maintain consistency. Now, let's throw some more variables into the mix and see how that affects things.

The image carries signal the forms were guessing at

Reading risk from the mammogram itself is not a new or fragile idea. Mammographic density has been an established breast cancer risk factor for decades. What has changed is how much more of the risk signal the image turns out to hold once a model reads the whole picture instead of a single density grade, and the evidence base is now large.

The 2026 JNCI systematic review by Lowry and colleagues pooled 41 studies of mammography-based AI risk models. Across prediction windows out to five years and beyond, the median area under the curve was 0.71, and the image-based models outperformed the questionnaire-and-density tools, which an earlier review put at a median near 0.61 against 0.72 for the image models on the same outcomes.¹

One finding inside that review should change how a program thinks. Adding the clinical questionnaire variables on top of an image model produced little or no improvement in discrimination.¹ That does not mean imaging replaces clinical history. Calibration, counseling, genetics, and high-risk management still need the clinical picture, and the review is explicit that calibration and prospective testing are unfinished.¹

What it means is that the mammogram already carries most of the information the questionnaire was assembled to estimate. The program has been collecting that information twice: once, laboriously, from the patient's memory, and again, at no extra cost, in an image it was already taking.

There is a catch the same review surfaces, and it matters for how a high score gets described. Studies that counted cancers found at the index mammogram reported a median AUC of 0.75. Studies that excluded them dropped to 0.68. The authors read that gap as a sign that part of what these models score as future risk is cancer already visible on the image.¹

Some of a high five-year risk, in other words, is early disease the model is catching now. That is useful, but it is detection wearing a risk label, and a program that treats every elevated score only as a reason to screen harder next year can miss that some of those women need a closer look this cycle.

What has arrived so far

This is no longer hypothetical. In May 2025 the FDA created a new device class and authorized the first entrant: Allix5, marketed as Clairity Breast, a Class II radiological software that produces a five-year breast cancer risk probability from a bilateral screening mammogram, for patients without a known cancer at the time of the exam.² The order is careful about the boundaries. The device does not diagnose or detect cancer, does not make care recommendations, is not a sole basis for clinical decisions, and is not meant to guide how the radiologist reads the film. Its output is considered after the read. It works only on directly acquired 2D images from specified Hologic systems, not synthetic 2D.² And it carries an obligation most software does not: the FDA's special controls require the maker to monitor performance after deployment, detect drift over time, and report current performance back to users, with discrimination and calibration broken out by demographic, equipment, and site subgroups.² The regulator built in an assumption that the number will not stay accurate on its own.

The thresholds a score like this would feed are already in place. Supplemental MRI has long been tied to a lifetime risk near 20 percent, and as DenseBreast-info summarizes, the March 2026 NCCN breast screening guidance now says to consider MRI for a woman whose five-year risk reaches 1.7 percent, the same cutoff the Gail model has used for years.⁴ Whether stratifying screening this way changes outcomes is being tested directly.

The WISDOM trial in JAMA found risk-based screening noninferior to annual screening for late-stage cancers, stage IIB or higher, at 30 versus 48 per 100,000 person-years, without reducing biopsies.³ The American College of Radiology pushed back, noting that mammography use was low and nearly identical in both arms, which makes the trial a weak basis for changing policy.⁵ The direction is real. The proof is not finished.

The impact is more information, and it lands on someone

Put the pieces together and the impact of AI on the old calculators comes into focus, and it is not the one most people expect. The mammogram-derived score does not subtract a step. It adds one.

A program that adopts it now runs two risk estimates that can disagree: the questionnaire-and-density number it has always produced, and an image-derived number that updates every time the patient is screened and that the FDA expects to be recalibrated over time. Before long a polygenic score joins them. Each addition is more information flowing into the same program, and information does not manage itself.

The operational questions are concrete, and most programs have not answered them. Who receives the image-derived score and reconciles it with the questionnaire number when the two point in different directions? Once a woman crosses 1.7 percent, who owns her name on that list, decides what she is offered, documents her decision, and carries the open cases forward so they are not quietly lost between visits? The scheduler, the technologist, the radiologist, and the risk coordinator each touch part of this, and none of them owns the whole of it by default. A traditional calculator becoming one input among several sounds like a small change. In the workflow it means someone now has to integrate several inputs and stand behind the result, and that someone has to be named before the inputs start arriving.

Where the old calculators end up

Over a longer horizon, the likely fate of Gail, Tyrer-Cuzick, and BCSC is not deletion. It is absorption. They become inputs inside a larger model, recalibrated to the local population and embedded in the workflow, where they help trigger a screening interval, an MRI discussion, or a genetics referral. That part is analysis more than settled fact, and the evidence supports the direction better than the timing. But the familiar number losing its place as the engine does not make the work smaller. It moves the work from producing one score to managing several, and managing several is harder than producing one.

The question worth sitting with is not whether the AI number is good. The better models already beat the questionnaire on the measure everyone quotes. The question is who in the program will hold a growing pile of risk information and turn it into a decision a patient can act on. That has always been the hard part of risk assessment. AI does not solve it. It raises the volume and the stakes, and it does that whether anyone has been assigned the job.

---

Sources

1. Lowry KP, Jeong HE, Kim KH, Hughes KS, Lee CI, Yala A, Kerlikowske K, Vachon CM. Current state of mammography-based artificial intelligence for future breast cancer risk prediction: a systematic review. JNCI J Natl Cancer Inst. 2026;118(3):392–403. doi:10.1093/jnci/djag002

2. US Food and Drug Administration. De Novo classification order, Allix5 (Clairity Breast), DEN240047. Issued May 30, 2025. accessdata.fda.gov/cdrh_docs/pdf24/DEN240047.pdf

3. Esserman LJ, Fiscalini AS, Naeim A, et al. Risk-based vs annual breast cancer screening: the WISDOM randomized clinical trial. JAMA. 2026;335(9):763–774. doi:10.1001/jama.2025.24784

4. DenseBreast-info. New NCCN guidelines add 5-year risk to breast screening: what it means for your patients. Referencing NCCN Clinical Practice Guidelines in Oncology, Breast Cancer Screening and Diagnosis, 2026. densebreast-info.org

5. American College of Radiology. ACR statement on WISDOM breast cancer screening trial results. December 12, 2025. newswise.com