Learning from VAE Errors to support ECG-based Differential Diagnosis of Myocardial Scar

In September 2026, researchers evaluated the effectiveness of electrocardiogram patterns derived from β-variational autoencoders in distinguishing patients with late gadolinium enhancement (LGE+) from those without LGE-. Using data from 300 patients with myocardial disease, they compared the performance of the basic ECGx.AI model with a shallow β-VAE combined with gradient boosting algorithm. The results showed that the latter had an ROC area of 0.577 and a sensitivity of 0.775; while the ECGx.AI combined with random forest had a higher ROC area of 0.686. Additionally, significant differences were observed in the dynamic time regularization reconstruction error between 10 groups of 12 leads. Using this error for logistic regression classification yielded an ROC area of 0.643. The study supports using VAE reconstruction error as a potential marker for scar-related electrocardiogram changes to assist in the differential diagnosis of myocardial scars.