Learning from VAE Errors to Support ECG-Based Differential Diagnosis of Myocardial Scar
Original title:Learning from VAE Errors to support ECG-based Differential Diagnosis of Myocardial Scar
Myocardial scar identification typically depends on late gadolinium enhancement CMR, a modality constrained by high cost and accessibility. Evaluating a cohort of 300 cardiomyopathic patients, researchers tested whether reconstruction failure in standard 12-lead ECGs could function as an alternative marker. Using a shallow β-VAE trained strictly on normal heartbeats from PTB-XL alongside the ECGx.AI foundation model, Dynamic Time Warping reconstruction errors exhibited statistically significant divergence across 10 of 12 leads, achieving an AUROC of 0.643 via logistic regression.
Why it's worth reading
It highlights a pragmatic anomaly-detection approach, showing that reconstruction deviations of normal-trained VAEs across standard ECG leads can reliably reveal hidden scar tissue.