ECG beat classification using capsule networks

Abstract

Electrocardiogram is a low cost non-invasive technology used to detect cardiac activity which leads to detection of diseases related to cardio vascular system. Because of the high prevalence of cardiovascular disease among people all over the world, electrocardiogram is used frequently to diagnose these diseases which has reduced the mortality rate significantly. Due to recent developments of telemetry devices and popularity of outpatient monitoring, long term ECG monitoring has become a norm in most clinical practices to detect underlying cardiac conditions. However, due to extreme amount of data generated, the automatic ECG classification systems has become an essential requirement. These automated systems can detect and report anomalies in ECG signal without requiring a cardiologist or technician to manual check the whole ECG signal. Although these systems provide significant assistance to healthcare workers, the accuracy and explainability of these systems are still not addressed completely. In this thesis, I have presented an ECG beat classification system using Capsule networks. The goal here is to address the aforementioned issues of existing ECG beat classification methods and technologies. Therefore, for this purpose, the approach I have taken is first and foremost based on the hypothesis that the ECG beat classification can be modelled using the capsule network approach. According to this approach, once the raw ECG signal is given as the input, the accurately classified ECG beats should be obtained as the output. For this, the design of the system was drafted to contain several key sections which work together to produce the expected outcome. These sections include, ECG signal acquisition, preprocessing and classification. The signal acquisition was included to acquire ECG raw signal from sources with different standards. Then the preprocessing stage was introduced to remove noise, baseline wander like artifacts and also to detect QRS complexes which is essential for beat classification. Then the deep learning model training was carried out using capsule network as the base technology. Finally, in addition to the aforementioned sections, the evaluation was carried out to ensure the performance and the explainability of the trained model. Unlike many existing researches on ECG classifications, we have implemented a model that can both perform better and provide feature-wise explanations for the classification. Sensitivity, Specificity, Positive predictivity like performance parameters are used to evaluate the model in terms of classification performance. The accuracy values obtained for this model was 98.89%, 99.44%, 99.56%, 98.83% and 99.17% for N, L, R, A and V beat classification and 97.94% for overall classification accuracy. As the conclusion, we identified that this ECG classification model based on Capsule networks has comparable performance to the other existing beat classification systems. In addition to that, the feature-wise explainability given by the capsule network model shows the superiority of this method in terms of reliability and practical usability.