Preprocessing Selection for Deep Learning Classification of Arrhythmia Using ECG Time-Frequency Representations

The trend of using deep learning techniques to classify arbitrary tasks has grown significantly in the last decade. Such techniques in the background provide a stack of non-linear functions to solve tasks that cannot be solved in a linear manner. Naturally, deep learning models can always solve almost any problem with the right amount of functional parameters. However, with the right set of preprocessing techniques, these models might become much more accessible by negating the need for a large set of model parameters and the concomitant computational costs that accompany the need for many parameters. This paper studies the effects of such preprocessing techniques, and is focused, more specifically, on the resulting learning representations, so as to classify the arrhythmia task provided by the ECG MIT-BIH signal dataset. The types of noise we filter out from such signals are the Baseline Wander (BW) and the Powerline Interference (PLI). The learning representations we use as input to a Convolutional Neural Network (CNN) model are the spectrograms extracted by the Short-time Fourier Transform (STFT) and the scalograms extracted by the Continuous Wavelet Transform (CWT). These features are extracted using different parameter values, such as the window size of the Fourier Transform and the number of scales from the mother wavelet. We highlight that the noise with the most significant influence on a CNN's classification performance is the BW noise. The most accurate classification performance was achieved using the 64 wavelet scales scalogram with the Mexican Hat and with only the BW noise suppressed. The deployed CNN has less than 90k parameters and achieved an average F1-Score of 90.11%.