AI Reads Mammograms and Finds Heart Disease: Tel Aviv Team's 97,000-Scan Breakthrough at ESC 2026

The most underused medical image in cardiology may turn out to be the mammogram. At the European Society of Cardiology’s annual congress in Munich — ESC Congress 2026, the world’s largest gathering of heart specialists — a team from Tel Aviv University presented a machine-learning model that reads routine breast cancer screening scans and reliably flags women who have cardiovascular disease. The findings, reported by The Guardian on August 27, point toward a future in which the hundreds of millions of mammograms performed worldwide each year become a free-standing cardiovascular screening programme riding on infrastructure that already exists.

The study in numbers

The retrospective cohort study, presented by Dr. Viana Copeland of Tel Aviv University, is one of the largest of its kind. The researchers analysed 97,364 mammography examinations from 29,921 women with an average age of 54, drawn from Israeli screening records. By cross-referencing the women’s medical histories, the team established the ground truth: 16% of the women had high blood pressure, 2.5% had coronary heart disease, and 2.5% had suffered a stroke.

A machine-learning model was then trained to detect those conditions from the mammogram images alone — no blood pressure cuffs, no cholesterol panels, no additional imaging. The results were striking:

• Stroke: identified correctly 86% of the time
• High blood pressure: 79% accuracy
• Coronary heart disease: 78% accuracy

Crucially, the model’s performance held up regardless of the women’s age and whether or not they also had cancer. That robustness matters: it suggests the model is picking up on genuine anatomical signal in the breast tissue and vasculature rather than incidental correlations with demographics or disease burden.

Why the heart hides in a breast scan

The intuition behind the work is that a mammogram is, among other things, an X-ray of the chest wall’s vasculature. Breast arterial calcification — calcium deposits in the arteries of the breast — has been recognised for decades as a marker of systemic vascular disease, yet it is almost never reported because radiologists reading screening mammograms are looking for tumours, not arteries. Machine learning changes the economics of that omission: software can quantify subtle vascular features across tens of thousands of images at essentially zero marginal cost, without adding a single extra examination, radiation dose, or clinic visit for the patient.

It also arrives at exactly the right moment in a woman’s health trajectory. As Dr. Copeland told delegates in Munich, mammography reaches many women in midlife — a critical window for recognising and addressing cardiovascular risk — and it reaches women who are not seeking care for heart symptoms at all.

A year of converging evidence

The Tel Aviv study is not an isolated result but the newest and most clinically direct entry in a fast-accumulating literature. In March 2026, a team led by Dr. Hari Trivedi at Emory University published a study of 123,762 women in the European Heart Journal showing that AI-quantified breast arterial calcification was an independent predictor of serious cardiovascular events: women with mild calcification were about 30% more likely to suffer a major cardiac event, moderate calcification pushed the risk above 70% higher, and severe calcification doubled or tripled it. The Emory work predicted future risk; the new Israeli work goes a step further by identifying existing disease — hypertension, coronary heart disease, stroke — straight from the image. The two approaches are complementary: one grades risk before symptoms appear, the other surfaces diagnosed-but-undetected conditions in women the health system has already lost track of.

The underdiagnosis problem this attacks

Cardiovascular disease is the leading cause of death in women worldwide, yet it remains persistently and disproportionately underdiagnosed and undertreated in women compared with men. The “man’s disease” myth still shapes both clinical suspicion and women’s own health-seeking behaviour: by the time many women present with symptoms, their disease is already advanced. Dr. Sonya Babu-Narayan, consultant cardiologist and clinical director of the British Heart Foundation, captured the significance: if the Israeli approach is further proven, “AI could one day allow breast cancer screening programmes to become dual-purpose, helping to flag women with the highest risk of dangerous cardiovascular disease, as well as spotting breast cancer early.”

The scale of the opportunity is enormous. Roughly two-thirds of women aged 50–69 in the European Union report having had a mammogram within the previous two years, and in the United States nearly 70% of women over 45 are up to date with screening. Fewer than 40% of women know their own cholesterol levels. A screening platform with near-universal midlife penetration that could simultaneously triage cardiovascular risk would address that mismatch without building anything new.

From poster to practice

Enthusiasm among ESC experts comes with caveats they are careful to spell out. Elena Arbelo, an expert member of the ESC communication committee, called the findings “compelling” but stressed that “the challenge now is to establish accuracy and reliability — to move from experimentation to clinical implementation.” The Israeli team itself is still iterating: its next steps include improving the model’s accuracy, reducing false positives and false negatives, and expanding the range of heart conditions it can detect. Retrospective cohort data, however large, must survive prospective validation — the test of whether flagging women actually changes outcomes, not just classifications.

The pathway to adoption, though, is unusually clear. Because the analysis runs on scans that are already being taken for cancer screening, integration means software in the imaging workflow plus guidelines for what to tell patients and their doctors when the model flags a finding. That is a regulatory and operational problem rather than a hardware one — the same lesson the Emory group drew when it called for AI-based calcification reporting to be folded into existing mammography programmes.

At a congress where AI is the spotlight theme, the Tel Aviv result is a reminder of where medical AI is most immediately useful — not in replacing doctors, but in extracting public-health signal from images the world is already producing at a rate of hundreds of millions per year, and in doing so for the patients medicine has historically listened to least.