ECG Beat Classification using Capsule Networks

Electrocardiogram (ECG) is a low-cost, non-invasive technology used to detect cardiac activity which leads to the detection of diseases related to the cardiovascular system. Due to the substantial volume of data generated in long-term ECG monitoring, the classification of ECG beats and reaching a justifiable diagnosis have been tedious tasks. Despite many machine learning solutions being provided for ECG classification, no study has specifically used the relationships between ECG waveform features as the basis for modeling such solutions and extracting the explanations based on those features. Therefore, based on this concept, a capsule network solution was developed for the classification of ECG beats. The improved accuracy in this classification consequently extended the solution to produce feature-wise explanations as well. The evaluation showed an overall 97.94% accuracy in beat classification. The accuracy for normal, left and right bundle branch block, atrial and ventricular beat classification yielded 98.89%, 99.44%, 99.56%, 98.83% and 99.17% respectively. For each of these classifications, feature-wise explanations were also generated. Hence it was concluded that this approach, based on capsule networks to classify ECG beats has recorded a higher accuracy than the existing approaches with feature-wise explanations leading to better diagnosis.