Attention Based Joint Learning for Supervised Premature Ventricular Contraction Differentiation with Unsupervised Abnormal Beat Segmentation

Deep learning has shown great promise in arrhythmia classification in electrocar-diogram (ECG). Existing works, when classifying an ECG segment with multiplebeats, do not identify the locations of the anomalies, which reduces clinical inter-pretability. On the other hand, segmenting abnormal beats by deep learning usu-ally requires annotation for a large number of regular and irregular beats, whichcan be laborious, sometimes even challenging, with strong inter-observer variabil-ity between experts. In this work, using Premature Ventricular Contraction (PVC)differentiation as an example of arrhythmia classification, we propose a methodcapable of not only differentiating the origin of PVC but also segmenting the as-sociated abnormal beats. Only the PVC origin labels are used in the training andno segmentation labels are needed. Imitating human’s perception of an ECG sig-nal, the framework consists of a segmenter and classifier. The segmenter outputsan attention map, which aims to highlight the abnormal sections in the ECG byelement-wise modulation. Afterwards, the signals are sent to a classifier for PVCorigin differentiation. Though the training data is only labeled to supervise theclassifier, the segmenter and the classifier are trained in an end-to-end manner sothat optimizing classification performance also adjusts how the abnormal beats aresegmented. Validation of our method is conducted on a dataset with 508 12-leadECG records of PVC patients. We observe that involving the unsupervised seg-mentation in fact boosts the classification performance. Meanwhile, a grade studyperformed by experts suggests that the segmenter also achieves satisfactory qual-ity in identifying abnormal beats, which significantly enhances the interpretabilityof the classification results.