Fragment-level classification of ECG arrhythmia using wavelet scattering transform

Cardiovascular disease detection and its prevention are among the most demanding tasks in the healthcare system nowadays, as around 50 million people worldwide are at risk of being affected by heart disease. The heart’s electrical activity recorded by an electrocardiogram (ECG) provides vital pathological information about cardiac abnormalities such as arrhythmia. However, the complexity and non-linearity observed in ECG signals make disease anticipation difficult. In this work, we proposed a new approach to classify 17-classes of cardiac arrhythmia using wavelet scattering transform (WST). The WST can provide translation-invariant and deformation-stable representations of ECG by using a series of wavelet convolutions with non-linear modulus and averaging operators. Scattering coefficients from four-time windows of WST for fixed-duration ECG fragments are taken as input features to the SVM classifier. We achieved an overall classification accuracy of 98.90% in categorizing 17 arrhythmia classes taken from the MIT-BIH arrhythmia database, having 1000 ECG fragments of 45 subjects. The proposed method categorizes a 10-second ECG fragment with an average classification time of 0.007 s on a computing platform of a 2.5 GHz processor with 8 GB RAM. Our results outperform existing state-of-the-art solutions and can be deployed in real-world applications.