Disentangled representational learning for anomaly detection in single-lead electrocardiogram signals using variational autoencoder

Maximilian Kapsecker, Matthias C. Möller, Stephan M. Jonas

Abstract: Highlights•β total correlation variational autoencoder encode atomic electrocardiogram features.•Anomalies arise as linear combinations of outliers along interpretable axis encodings.•Fine-tuning ECG models per subject personalizes and mitigates data heterogeneity.•Performance remains competitive, despite the model being unsupervised and explainable.