• generative models
• deep learning
• biomedical time series

Computer-aided diagnosis systems (CAD) usually require a huge volume of labeled clinical data giving rise to several challenges, patients’ privacy and data anonymization are one of those which need to be solved. The open access to medical data is restricted by this problem which discourages also technological advancement and deep understanding of patterns in recorded electrocardiograms (ECG) showing rhythm anomalies. The main objective of this work is to generate artificial ECG data agreeing with the standard of plausible clinical data by representing a range of classes of anomalies. The latter is a solution addressing the anonymization and privacy issues above mentioned, which are usually solved by de-identification methods. On the other hand, the parallel objective of this study is to interpret the latent semantics of rhythm anomalies learned by deep generative models in a strict comparison with baseline autoregressive architectures.