AI trained on mammograms predicts heart attack and stroke risk, study finds
Australian researchers report a deep-learning model that uses routine breast screening images to estimate cardiovascular risk as accurately as traditional calculators
Researchers in Australia reported that an artificial intelligence algorithm trained on routine mammograms can predict the risk of major cardiovascular events such as heart attack and stroke with performance comparable to conventional clinical risk calculators.
The deep-learning model was developed and tested using images from 49,196 women enrolled in the Victoria Lifepool cohort registry, a breast cancer research initiative. The cohort had an average age of 59; about one-third were taking medication for high cholesterol and 27% were on treatment for high blood pressure.
During an average follow-up period of almost nine years, 2,383 women experienced a heart attack, 731 had heart failure and 656 had a stroke. The study, published in the journal Heart, found that the mammography-based algorithm performed about as well as established calculators that use age and clinical variables to assess cardiovascular risk.
Investigators said mammographic features such as breast arterial calcification and tissue density are associated with cardiovascular risk and can be exploited by machine-learning techniques. "We developed and tested a deep learning algorithm for cardiovascular risk prediction based on routine mammography images," the authors wrote. They noted a key advantage of the model was that it did not require additional history taking or electronic medical record data.
The researchers suggested that mammography—already widely used for breast cancer screening in midlife—may offer a "cost-effective 'two for one' opportunity to screen women for breast cancer and cardiovascular risk, enabling broader cardiovascular risk screening."
The study's findings build on growing interest in applying machine learning to cardiovascular risk prediction while using data sources not traditionally intended for heart disease screening. The authors acknowledged the approach is novel and placed their work in the context of an expanding body of research exploring nontraditional biomarkers and imaging features for risk stratification.
As an example of current screening practice in the United Kingdom, women registered with a general practitioner are automatically invited for a mammogram every three years between the ages of 50 and 71; each examination typically includes two X-rays of each breast to detect signs of cancer. The Australian team and others noted that such routine imaging encounters provide an opportunity to assess additional health risks without separate appointments or tests.
The study did not report changes to clinical practice; broader clinical adoption would require further validation in diverse populations, assessment of how best to integrate the findings into screening workflows, and evaluation of cost-effectiveness and potential implications for follow-up care. The authors said the use of mammography images to predict cardiovascular risk is novel but that the use of machine-learning models to do cardiovascular risk prediction is gaining traction.