Doctors may soon be able to recognize a difficult-to-diagnose form of heart disease within seconds using an AI model developed at the University of Michigan. This is according to a recent study. The model analyzes a standard electrocardiogram (ECG) and detects coronary microvascular dysfunction (CMVD). This is a condition that often goes unnoticed with existing methods.
CMVD affects the smallest blood vessels of the heart and causes symptoms such as chest pain and an increased risk of heart attack. Unlike classic coronary artery disease, the condition is not visible on an angiogram. A reliable diagnosis usually requires advanced and expensive PET scans, which are mainly available in specialized centers.
CMVD diagnosis
The new AI model, published in NEJM AI, breaks through that limitation. The researchers trained the system to recognize CMVD based on an ECG recording of just ten seconds. The model performed significantly better than previous ECG AI applications, including in predicting myocardial flow reserve, the gold standard for diagnosing CMVD.
According to lead researcher Venkatesh L. Murthy, the model offers clinicians a practical tool for identifying a condition that is currently often missed, particularly in emergency departments. Every year in the US alone, around 14 million people report to the emergency department or outpatient clinic with chest pain, with CMVD regularly going undetected.
Self-supervised learning
One challenge was the limited number of PET scans available for training. The team solved this with self-supervised learning. First, a deep learning model was pre-trained on more than 800,000 unlabeled ECG signals, so that it could independently learn to recognize the heart's “electrical language.” The model was then refined using a smaller dataset of PET scans.
The result is an AI system that not only reliably predicts CMVD, but also improves diagnostic accuracy for other, more common heart conditions. It is striking that ECGs at rest proved to be almost as informative as exercise ECGs.
Enriching diagnostics with AI
This technology can be of great value to hospitals with limited resources. “In patients with chest pain, an angiogram may be ‘clean’ even though CMVD is present,” said co-author Sascha N. Goonewardena. “With this AI ECG, we can quickly, non-invasively, and cost-effectively determine who would benefit from further investigation.”
The study highlights how AI can enrich existing, accessible diagnostics and thus improve access to specialized care. For cardiology, this represents an important step toward earlier and more equitable detection of complex heart disease.
AI ECG technology
A few months ago, an international study showed that an AI-driven ECG system can detect serious heart attacks (STEMIs) faster and more accurately, with significantly fewer false positives. STEMIs require rapid treatment; without intervention within 90 minutes, the risk of mortality increases significantly. In practice, this proves difficult to achieve, especially outside PCI centers.
The AI ECG model, known as Queen of Hearts, recognizes both classic and atypical patterns of acute coronary artery occlusion. In an analysis of 1,032 patients in three European PCI centers, the AI detected 553 of the 601 STEMIs, compared to 427 with standard ECG assessment. At the same time, the proportion of false positives fell from 41.8 to 7.9 percent.
