Mass General Brigham and VA built an AI that spots hidden heart disease in old CT scans

Millions of people get routine chest CTs every year to check for lung cancer or pneumonia, and once the lungs clear, those scans sit untouched in hospital archives.

Now, researchers from Mass General Brigham and the U.S. Department of Veterans Affairs (VA) just built an AI tool that digs through those old scans to spot hidden heart disease.

Published in NEJM AI, the new deep learning algorithm, dubbed AI-CAC, detects coronary artery calcium (CAC) on standard chest CTs without requiring specialized cardiac imaging.

Mining routine scans for hidden risks

The gold standard for measuring artery calcium normally requires gated CT scans, which sync with a patient's heartbeat to freeze cardiac motion.

However, the vast majority of chest imaging done in clinical care is nongated. In the VA healthcare system alone, there are millions of routine nongated scans sitting in archives, compared to only about 50,000 gated studies.

Researchers trained the AI model using real-world scans collected across 98 VA medical centers, then tested it on 8,052 scans to simulate everyday opportunistic screening.

The tool hit an 89.4% accuracy rate in detecting whether calcium deposits were present. When calcium was found, it was 87.3% accurate at classifying whether the score exceeded 100, the typical threshold for moderate cardiovascular risk.

It also accurately flagged long-term outcomes: patients with AI-detected calcium scores over 400 faced a 3.49 times higher risk of all-cause mortality over a 10-year span compared to those with a score of zero. Four independent cardiologists confirmed that 99.2% of patients flagged in that high-risk bracket would directly benefit from preventive lipid-lowering medication.

Why it matters

We think the real brilliance of this approach is that it requires zero extra hardware and exposes patients to zero additional radiation. Instead of asking patients to schedule and pay for another specialist visit, it opportunistically mines routine scans they have already completed for other reasons.

For patients and doctors, this flips the traditional cardiac care model. Catching severe arterial calcification years before an actual heart attack occurs allows clinicians to prescribe standard statins proactively, effectively preventing high-cost, life-threatening emergencies.

That said, there is an important caveat: the algorithm was trained and tested exclusively on U.S. military veterans, a group that heavily skews older and male. We will need to see whether the model maintains its 89% accuracy when deployed across a broader, more diverse general population, but as a blueprint for repurposing existing hospital data, it looks remarkably effective.