JMIR: AI Learns to Detect Heart Disease Risk from Mammography Data
According to a publication in JMIR Publications, artificial intelligence enables routine mammography to assess cardiovascular disease risk. By analyzing breast arterial calcification, AI can identify high-risk women without additional tests, which is especially relevant for patients under 50.
The Gist: What's Really Happening
Researchers from Emory University and Mayo Clinic have developed a transformer-based AI model that turns routine mammography into a cardiovascular risk screening tool. The model segments breast arterial calcification (BAC) and outputs a continuous metric—each additional square millimeter of calcification corresponds to a 1% increase in MACE risk. The cohort is massive: 123,762 women across two independent samples. BAC was found in 16.1% of participants in the Emory cohort and 20.6% in the Mayo cohort.
The key takeaway fits in one sentence: BAC is an independent risk factor for cardiovascular events that adds predictive value on top of the PREVENT score, the current clinical standard for risk calculation. For women under 50—a demographic group that traditional cardiac models may dismiss as low-risk—moderate or severe BAC correlates with significantly lower MACE-free survival. These women typically don't get lipid panels or see a cardiologist, but 40 million women in the US undergo mammography annually, starting at age 40.
Timeline and Context
The history of BAC is not new—radiologists have seen these calcifications on images for decades. But as lead author Hari Trivedi notes, "Practice guidelines do not require radiologists to report this finding." The result is absurd: the information exists in the image pixels but never makes it into the report or clinical decision. According to some data, BAC is present in 12–42.5% of screening mammograms but is reported in less than 5% of cases.
The turning point was the PREVENT calculator, released by the American Heart Association in 2023. It replaced the Pooled Cohort Equations and is designed to assess cardiovascular, kidney, and metabolic risk. But PREVENT is a purely clinical-laboratory tool; it does not include anatomical vascular imaging. BAC fills exactly this niche—it is an anatomical marker of calcium burden visible on existing images.
The JMIR study advances the methodology significantly. Previous work assessed BAC as a binary variable—present or absent. Trivedi's model outputs a continuous score in mm² and categorizes it: mild (0–10 mm²), moderate (10–25 mm²), severe (>25 mm²). Validation was performed on two fundamentally different populations: Emory in Atlanta, nearly half African American; Mayo, older and including more Hispanic patients. The model works cross-scanner and cross-population.
In parallel, a commercial ecosystem is developing. Onsite Women's Health, the largest provider of office-based mammography in the US (over 175 practices in 28 states, projecting 600,000+ exams in 2026), launched the "Mammo with heart" program based on FDA-cleared AI from CureMetrix. CMS created new reimbursement codes for algorithmic analysis of coronary calcium on chest CT effective April 1, 2026—a precedent for paying for AI-based calcification analysis already exists. In 2026, AI-specific CPT codes for mammography were also introduced: 3321F–3323F for automated comparison with prior images.
Who Wins and Who Loses
Winners
Women aged 40–50. These are the primary beneficiaries. They get mammograms on schedule, but their cardiovascular risk is systematically underestimated. Moderate or severe BAC in this age group is a powerful predictor of MACE that standard algorithms, which penalize young age, miss. They get a "two-for-one exam" without additional radiation, doctor visit, or blood test.
Screening mammography providers with AI integration. Onsite Women's Health turned mammography from one screening into two. CEO Jillian Wright puts it bluntly: "We believe it's our responsibility to extract the maximum from every mammogram." This is a differentiation strategy: when competition among mammography centers is based on convenience and speed, "Mammo with heart" adds clinical value that competitors without AI cannot replicate.
CureMetrix. Their cmAngio, an FDA-cleared solution for BAC analysis, is now being deployed across a network of 175+ practices. Onsite is an ideal distribution channel.
Cardiologists receiving referrals. Trivedi acknowledges: "We have a lot of demand and support from cardiologists." 3,000 women per year at Emory alone will be identified as having severe BAC. These patients will enter cardiology clinics earlier than without screening. For cardiologists, this means new patients; for the system, it shifts spending from late-stage intervention to prevention.
Researchers in gender-specific medicine. BAC, for the first time at this scale and with this accuracy, documents an anatomical risk marker specific to the female population. This strengthens the evidence base for gender-specific approaches in preventive cardiology.
Losers
Traditional cardiac risk calculators (PREVENT, ASCVD Risk Estimator). They don't lose clinical validity, but they lose their monopoly. Trivedi and team showed that BAC adds predictive value on top of PREVENT. For clinicians, this creates a dilemma: PREVENT gives one risk number, BAC gives another; how to integrate them and which to prioritize when prescribing statins?
Clinicians with a high threshold for prescribing statins to women. Historically, statins are underprescribed to women. BAC screening, showing objective anatomical risk, may change physician decisions and increase statin prescriptions. For proponents of minimal medication intervention, this is an unwelcome shift.
Insurers. 3,000 additional cardiology referrals per year from one medical center means increased costs. In the short term, additional screening and subsequent cardiologist visits cost money. Insurers will demand evidence that early intervention pays off through prevented heart attacks and strokes—data from longitudinal outcome studies that don't yet exist.
Women with false-positive results. Trivedi does not discuss this in the interview, but false-positive BAC signals are inevitable. If the model overdiagnoses calcification, a woman gets unnecessary stress, unnecessary cardiac workup, and potentially unnecessary medication. The false positive rate for BAC AI is a critical unpublished parameter.
What the Media Miss
Insight #1: CMS reimbursement for AI-based calcification analysis already exists—but not for BAC. This fact has escaped all commentators. CMS has already created a national billing code for algorithmic analysis of coronary artery calcium (CAC) and aortic valve calcium (AVC) on chest CT, effective April 1, 2026. No analogous code exists yet for mammographic BAC. But the regulatory precedent is set: CMS has recognized that algorithmic calcium metrics are a billable service. Once Trivedi's team and commercial players like CureMetrix build sufficient evidence, creating a BAC-specific CPT code will be a matter of months, not years. This is an accelerator the market has not yet priced in.
Insight #2: 3,000 women per year with severe BAC is just one medical center. Scale that to 40 million mammograms per year. Emory performs 150,000 mammograms per year, of which 2–3% show severe BAC, about 3,000 women. Extrapolate to the national 40 million screening mammograms: that's 800,000–1,200,000 women with severe BAC annually. Identifying them will create a wave of cardiology referrals that will either overwhelm the system or be filtered through automated stratification protocols. Trivedi admits: "We foresee challenges—how are we going to handle this potential extra volume?" This is not a technical problem but a capacity planning problem that no one is solving yet.
Insight #3: Smoking data invert expectations—and reveal a methodological bombshell. "BAC is inversely associated with smoking in our study—smokers have less BAC," says Trivedi. This is counterintuitive, since smoking is a powerful risk factor for both atherosclerosis and MACE. Trivedi attributes it to "unreliability of smoking data" and claims it's not a problem because BAC and smoking are independent risk factors. But for a clinician, this creates a dangerous situation: a female smoker with low BAC could be mistakenly classified as low-risk, even though smoking elevates her actual risk. BAC screening does not replace behavioral risk factors, but if the model suppresses the weight of one factor due to collinearity with another, predictive accuracy drops.
Forecast: Next 30 Days and 90 Days
30 Days (by June 6, 2026)
Trivedi and team will begin prospective clinical deployment of the model within Emory Healthcare. This is not technical validation—that's done. It's measuring clinical impact: how many patients with severe BAC are actually referred to a cardiologist, how many receive statins, how lipid panels change, were there false alarms? This data will form the basis for a CPT code application.
Meanwhile, Onsite Women's Health is accelerating the expansion of "Mammo with heart." Currently, the program is available at one affiliated practice—Intermountain Medical Clinic. In the next 30 days, announcements of new sites will follow. Given their network of 175 practices in 28 states, scaling will be rapid. CureMetrix will receive a growing stream of real-world clinical performance data for cmAngio.
The JMIR publication will trigger a wave of inquiries from radiology departments. Every major medical center will ask: "Does our PACS have the capability to run this AI module? Which vendors offer integration?" At least 20–30 centers will initiate internal discussions about BAC screening.
90 Days (by August 5, 2026)
Results from three international validation cohorts will become known. Trivedi reported that researchers are working with cohorts in Greece, Brazil, and the UK, testing not technical model performance but cardiovascular outcomes. If BAC shows comparable added predictive value on three continents, it will transform the study from an "interesting finding" into a "globally validated biomarker." Such evidence is a direct path to inclusion in clinical guidelines.
AHA/ACC will begin informal discussions about incorporating BAC into PREVENT 2.0 or a parallel risk calculator. This process is slow—from publication to guideline update takes 2–4 years. But given that PREVENT already lacks anatomy and BAC is precisely an anatomical parameter, pressure on guideline committees will mount. Expect the first editorial in Circulation or JACC calling for a revision of approaches.
CMS will receive a request from professional societies (Society of Breast Imaging, American College of Radiology) to create a CPT code for BAC analysis. This won't happen instantly—the CPT Editorial Panel meets three times a year, and a code could enter the 2027 cycle. But in parallel, clinics will begin piloting reimbursement through codes for incidental findings or through the calcium codes created for CT.
The AI radiology market will segment into two camps: companies adding BAC as a second function (mammography AI with a claimed "two-for-one") and companies focusing exclusively on breast cancer. Those with FDA-cleared BAC detection (CureMetrix) will gain a competitive window of 12–18 months while others go through the regulatory process.
The key unknown is patient compliance. A woman comes for a mammogram and is told: "No cancer detected, but you have moderate breast arterial calcification." What does she do? See a cardiologist? Ignore it? Panic? The answer to this question will determine whether BAC screening becomes a real prevention tool or a technology that generates data without clinical consequences. Trivedi understands this: "Now that we have this tool, we can start studying it." The tool is here. The data is here. But patient and physician behavior is the last mile that separates innovation from real impact. And the last mile is always the hardest.
— Editorial Team
