A deep learning algorithm that only uses data from mammogram images along with age may predict major cardiac events as accurately as traditional cardiovascular risk calculators, new data suggest. 

The findings were published online in Heart.

The algorithm was developed by researchers at The George Institute for Global Health in collaboration with the University of New South Wales and University of Sydney in Sydney, Australia.

Lead author Jennifer Barraclough, PhD, a research fellow at The George Institute, said using information from a screening process already widely used by women means this model could help predict cardiovascular risk for women in diverse communities globally. 

“We hope this technology will one day provide greater and more equitable access to screening in rural areas,” she said in a press release.

An ‘Innovative’ Approach

Barraclough and colleagues designed and validated the model, called DeepSurv, using routine mammograms from 49,196 women in metropolitan and rural areas of Victoria, Australia, linked to individual hospital and death records. Median follow-up was 8.8 years. They then compared the model to traditional models that rely on multiple data points based on known cardiovascular risk factors, such as blood pressure and cholesterol levels.

The new model had a concordance index of 0.72 (95% CI, 0.71-0.73), with similar performance to current models, including the New Zealand PREDICT tool and the American Heart Association PREVENT equations that incorporate age and clinical variables.

To date, researchers have developed algorithms based on breast arterial calcification from mammographic data, but they have limitations, particularly in predicting cardiovascular risk in older women, according to the paper. 

“This is the first deep learning model incorporating all breast characteristics/architecture from routine screening mammograms, as opposed to breast arterial calcification alone,” the authors wrote.

For their model, Barraclough and colleagues used the first available set of right and left breast digital mammographic images and, whenever possible, used image data for mammograms performed on two different occasions. In 9% of cases, only one set of mammographic images was available before the first cardiovascular event and was therefore used in isolation. They then used an image encoder to extract features from the mammogram data, which were combined in a neural-network-based model to predict risk for major adverse cardiac events over 10 years.

“Being able to use multiple features of a mammogram with just age and not knowing anything else about the patient to be able to predict risk is definitely innovative,” Rupa Sanghani, MD, director of the Rush Heart Center for Women at Rush University Medical Center in Chicago, told Medscape Medical News.

Underscreening of Cardiovascular Risk in Women

Sanghani noted additional cardiovascular risk models are much needed for women. 

“We know that heart disease is underestimated in women. And we know that women don’t come in as much as they should to be seen,” she said, though the rates of mammogram screening are relatively high.

“Your chances of dying from breast cancer are about 1 in 8 whereas your chances of dying from heart disease are 1 in 3,” Sanghani added.

The model has potential for use in the US, Sanghani said, but she noted it’s still in the early stages and the paper doesn’t list all the components on the mammogram used in the model.

“Mammography is increasingly being looked at as a tool to predict heart disease risk in women, which may allow for earlier detection of heart disease. I think this is a great example of how artificial intelligence could be used in healthcare, but we still need research on cost, implementation, outcomes, and whether it truly changes their care,” Sanghani said.

In an accompanying editorial, Gemma A. Figtree, MBBS, an interventional cardiologist with the Kolling Institute of Medical Research, and Stuart M Grieve, MBBS, a radiologist with the Charles Perkins Centre, both at the University of Sydney, wrote that using mammography data “represents little direct cost and perhaps avoids the risk of ‘losing the moment’ of a woman’s interaction with the health system at breast screening.”

They highlight substantial “underappreciation of heart disease as a threat to women by both women and the health system” as well as frustration of the “suboptimal performance of traditional risk factor algorithms.”

They noted that more than 67% of women in the US and the UK get screened for breast cancer.

“Mammography may therefore represent a ‘touch point’ for raising awareness about cardiovascular risk and disease in women,” Figtree and Grieve wrote.

Barraclough and Sanghani reported no relevant conflicts of interest. Figtree reported receiving grants from National Health and Medical Research Council (Australia) and multiple pharmaceutical companies and holding several patents related to cardiovascular disease. Grieve reported serving as company lead and founder of iCoreLabs. 

Marcia Frellick is a Chicago-based, independent healthcare journalist and a regular contributor to Medscape.