Abstract
This paper presents a method for the classification of short-time single-lead ECG recordings of variable size. These recordings are published as part of a challenge in 2017 by PhysioNet. The goal of the challenge is to classify the ECG recordings into four classes (Normal, atrial fibrillation, other abnormalities, and too noisy). The dataset is challenging because the high inter-class variability and because class sizes are unbalanced. The proposed method starts by denoising the ECG recordings using bandpass filtering, then detecting and correcting inverted signals using our own proposed algorithm. Since the recording have variable size, our proposed solution extracts a large set of features (188) that the literature has shown to be effective in characterizing ECG signals and detecting abnormalities. Then we present our own carefully designed residual convolutional neural network (CNN) with 5 hidden layers and use advanced and efficient training techniques to build a deep learning classifier for the solution. Finally the paper presents preliminary results of testing the proposed solution on the challenge dataset and shows its classification capabilities.