AI-Quantified Breast Arterial Calcification Can Predict Heart Disease Risk From Mammograms

Despite being the leading cause of death in women globally, cardiovascular disease (CVD) tends to be underdiagnosed and undertreated in women [1].

Predicting CVD risk is complex. It involves integrating multiple different kinds of data—from demographic information to laboratory scores to medical history—using a tool such as the PREVENT score to quantify that risk [2,3]. Data are often incomplete, and for women, accuracy and timely detection tend to be further limited because the ways CVD can present in women specifically have historically been underrepresented and underrecognized.

A recent retrospective study of 123,762 women [4] suggests a new, efficient way to address some of these limitations and improves how women’s cardiovascular risk is assessed: applying artificial intelligence to routine mammograms.

The Study

Breast arterial calcification (BAC)—a buildup of calcium in breast tissue—is an increasingly established risk factor for CVD [5-9]. Researchers analyzed the data of women from two health care systems in the United States who had undergone screening mammography—a cohort of 74,124 women aged 55.5 years on average, from Emory Healthcare, and a cohort of 49,638 women aged 59.5 years on average, from Mayo Clinic—and assessed the images and quantified the extent of BAC present.

To do this, they developed, trained, and validated a deep learning artificial intelligence model, using it to segment BAC on mammograms and automatically quantify risk severity. Where previous studies have often assessed BAC as a binary variable (presence vs absence) [10], this new model—a “transformer-based segmentation model” capable of maintaining high sensitivity while reducing false positives—permitted finer-grained risk categorization. BAC severity was categorized as zero (0 mm²), mild (>0‐10 mm²), moderate (>10‐25 mm²), and severe (>25 mm²).

BAC as an Independent Risk Factor

BAC was ultimately detected in 16.1% (Emory cohort) and 20.6% (Mayo cohort) of study participants. Increased BAC predicted increased cardiovascular risk over and above the PREVENT scores, confirming that it is an independent risk factor offering significant additional prognostic value.

“The model returns a continuous area score—an amount of BAC millimeter squared from zero to anything, and every millimeter squared of BAC increases risk of CVD by 1%. If you have 20 millimeters of BAC, your hazard ratio is 20% increased risk,” says Hari Trivedi, MD, Associate Professor, Radiology and Biomedical Informatics at Emory School of Medicine and lead study author.

He adds that since reporting a continuous score can be “tricky” and “overwhelming in a clinical setting” and for ease in statistical analysis, the researchers grouped the numbers into mild, moderate, and severe categories. “The model returns an actual number, but we take that and put it into mild, moderate and severe categories, similar to what’s done for cardiac CT [computed tomography].”

The researchers also observed a higher prevalence of antihypertensive medication use and statin use as BAC severity increased, and higher prevalence of some cardiovascular risk factors, including diabetes mellitus, higher systolic blood pressure, higher body mass index, and lower eGFR (estimated glomerular filtration rate), consistent with previous BAC studies [10,11]. Interestingly, increased BAC severity was not associated with higher smoking prevalence.

“There have been varying results on the link between BAC and smoking…Smokers have lower BAC in our study,” says Trivedi. “Smoking data is notoriously unreliable…Even if BAC is inversely related to smoking, we have already shown BAC is an independent risk factor and smoking is an independent risk factor. The nice thing about independent risk factors is it doesn’t matter if they relate to each other.”

Cohort Differences

In addition to the BAC prevalence differences between cohorts—likely attributable to differences in age—the study found a statistically significant association between BAC and myocardial infarction in the Mayo cohort, but not in the Emory cohort. Trivedi thinks population differences may explain this.

“All the trends that we observed with mild, moderate, severe, the hazard ratios, etc, the trends are the same (in both cohorts). Now, the degree of severity changes. At Emory, having a severe BAC score confers a 2.2 times risk of CVD. At Mayo, that number is slightly different. And that’s to be expected,” says Trivedi. Controllable and uncontrollable risk factors for CVD including, for example, lifestyle, existing health conditions, age, race, and ethnicity, are well-established [12]. “The model, segmentation, and score works across multiple scanners and multiple populations. Emory’s population is more racially diverse. It is almost half African American, whereas Mayo’s is not. Mayo has more Hispanic patients than Emory. Emory’s in the low single digits for Hispanic patients. Mayo’s population is older than Emory’s patient population.”

Earlier Screening?

The research also found the prognostic value of BAC especially noteworthy in women under 50. While BAC is more prevalent with age, its presence in younger women is a particularly strong indicator of elevated cardiovascular risk [5,13,14].

The analysis suggests that moderate to severe BAC in this younger cohort was associated with significantly lower major adverse cardiovascular event survival, identifying a high-risk group that might otherwise be overlooked by age-weighted risk models.

“Women in the United States are recommended to start screening mammography at age 40. Women under 50 are typically not considered at risk for CVD, and may not even be screened. So the fact that we see a link between BAC and CVD even in women under 50 is really important,” says Trivedi.

Next Steps

Trivedi and team’s model holds a great deal of promise for rapid risk assessment and early detection of CVD in women at scale. It can be readily implemented into routine mammography screenings, permitting more efficient, lower-barrier dual-purpose care. The researchers are working on continued deployment and validation.

“We are actively pursuing three international cohorts in Greece, Brazil, and the UK for evaluation, not of the model’s technical performance, but of cardiovascular outcomes. We are also working on a prospective clinical deployment inside of Emory. We have a lot of demand and support from the cardiologists,” says Trivedi.

Recently, the model was also integrated and deployed at the Mayo Clinic in Arizona. “Emory does 150,000 screening mammograms per year. Of those, 2%‐3% of women roughly have severe BAC—that’s 3000 women a year. So part of the challenge is going to be, how are we going to handle all of this potential extra volume? We anticipate doing a clinical deployment and then measur[ing] the impact of this. We sort of technically validated the model. Now we need to validate the clinical impact of having the information,” says Trivedi.

“You can screen women at scale, essentially without any cost, without going to an extra doctor’s visit, without getting your lipids drawn, but this is like truly exhaustive data that comes out of an existing mammogram that 40 million women are doing a year, and you can immediately risk stratify all of them,” he says. “Now that we have this tool, we can begin studying it.”