StenCE framework detects coronary stenosis from ECGs with high accuracy
A new pretraining model learns stenosis signals from routine ECGs, outperforming prior methods.
A team of researchers from multiple institutions (including Nikola Cenikj, Özgün Turgut, Alexander Müller, and others) have developed StenCE, a pretraining framework that uses cross-modal contrastive learning to align electrocardiogram (ECG) signals with coronary angiography images. The goal is to enable detection of severe coronary artery stenosis—a major risk factor for heart attacks—directly from ECGs, which are fast, cheap, and routinely collected even in asymptomatic patients. Currently, stenosis diagnosis relies on invasive coronary angiography, limiting screening to high-risk symptomatic individuals.
StenCE works by training an ECG encoder to produce representations that match those derived from angiography of confirmed stenosis cases, effectively teaching the model to identify subtle ECG patterns linked to blocked arteries. The framework was evaluated on datasets with varying stenosis severity thresholds and also on additional ECG disease classification tasks. Results show consistent performance improvements across different ECG encoder architectures, and StenCE is the first method to achieve high accuracy in severe stenosis classification directly from ECG signals. The source code is publicly available, enabling further research and clinical validation.
- StenCE uses cross-modal contrastive learning to align ECG and angiography features, enabling stenosis detection from ECGs alone.
- The framework achieves the first high-performance classification of severe coronary stenosis from non-invasive ECGs.
- Evaluations across multiple severity thresholds and ECG encoders show consistent improvements over previous methods.
Why It Matters
Enables early, non-invasive screening for heart disease from routine ECGs, potentially catching silent stenosis.