AI reading ECG images spots 87% of coronary disease cases, but half of healthy patients flagged

The trade-off was many false alarms: at the 0.50 threshold, specificity was 50%, meaning 317 of 631 people without CAD were correctly classified. On a later, independent cohort of 2,168 samples collected prospectively from February to May 2024, ROC-AUC was 0.81, but sensitivity fell to 0.67 while specificity rose to 0.78.

The authors conclude that routinely acquired ECG images may support low-cost, scalable CAD risk stratification. The negative predictive value of 0.57 on the temporal cohort suggests a negative result would not reliably rule out disease in a population where 64% had CAD.

Standard ECG is cheap and widely available but performs poorly at detecting early and nonobstructive coronary artery disease when read using conventional criteria. The researchers asked whether a convolutional neural network applied to ECG images alone could identify angiographically confirmed disease in a real-world population.

The model, a DenseNet-121, was trained on ECG images paired with coronary angiography or coronary CT angiography results. Performance was assessed with ROC-AUC, PR-AUC and standard diagnostic metrics, and prevalence-dependent behaviour was summarised with PPV and NPV prevalence curves (AUCppv and AUCnpv).

Results were consistent in overall discrimination (ROC-AUC 0.80 on the test set, 0.81 on the later cohort) but the balance between catching disease and avoiding false positives shifted between datasets. The authors describe the findings as supporting risk stratification rather than a standalone diagnosis.