Interpretable ECG Beat Embedding using Disentangled Variational Auto-Encoders

Electrocardiogram signals are often used in medicine. An important aspect of analyzing this data is identifying and classifying the type of beat. This classification is often done through an automated algorithm. Recent advancements in neural networks and deep learning have led to high classification accuracy. However, adoption of neural network models into clinical practice is limited due to the black-box nature of the classification method. In this work, the use of variational auto encoders to learn human-interpretable encodings for the beat types is analyzed. It is demonstrated that using this method, an interpretable and explainable representation of normal and paced beats can be achieved with neural networks.