Arrhythmia Classification with Continuous Wavelet Transform and Convolutional Neural Network on ECG

Abstract

Electrocardiogram (ECG) is an electric signal recorded from the heart which is used to monitor conditions in the cardiovascular systems of hospitalized patients. For many years, there have been efforts to accurately determine cardiovascular conditions represented by ECG. This paper proposes a four-step method integrating continuous wavelet transform (CWT) and convolutional neural network (CNN) to classify arrhythmia into five distinct classes. First, the ECG signals are preprocessed for improved data quality. Second, continuous wavelet transform with the Complex Morlet wavelet is applied on the data to return image representations of the signals over the time-frequency domains. Third, a lightweight CNN is trained to classify the five different arrhythmia classes based on the images obtained via wavelet transform. Finally, the network is used to predict the class of each ECG signal from the test dataset. The results derived using this method reveal an average F1 score of 90% and a weighted F1 score of 98%.