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