Anobeat: Anomaly detection for electrocardiography beat signals

Electrocardiography signals are composed of variform heartbeats which could indicate the condition of the heart and reveal the risk of heart attacks. Many existing classification-based works for abnormal beats detection are limited by the class-imbalanced data or labor-intensive manual annotation bias. A promising trend to address the issue is to identify the abnormal data that differs from the normal data by utilizing normal (oneclass) data to learn the manifold and detect the anomaly to the unseen and unlabeled data in an/aunsupervised/semi-supervised manner. In this paper, we propose Anobeat, a semi-supervised approach, to perform the abnormal beat detection by facilitating adversarial regularized autoencoders constrained with multifeature and reconstruction error. In order to obtain a robust and reasonable latent coding, we deploy two discriminators in the latent space and visual space to distinguish real and fake features and minimize the distance between two features to train the visual discriminator in alternate steps. Meanwhile, we minimize the reconstruction error and maximum distance between input and noise features to improve the decoder. The adversarial multi-feature constraints enable the generator to learn the latent representations of the target normal data and reconstruct the beats properly. Experiments showed that Anobeat achieved ROC-AUC of 0.960 and 0.894 in the MIT-BIH intrapatient and inter-patient dataset respectively, which outperforms the most competitive baseline by 1.61% and 0.62% respectively. Anobeat also performs comparative robustness and shows good interpretability in the European ST-T and MIT-BIH Arrhythmia Database.