Classification of Arrhythmia ECG Signals using Convolutional Neural Network.

The electrocardiogram (ECG) has been established as a reliable tool for monitoring cardiovascular health. Vast amount of ECG recordings can pose a challenge for its processing and analysis and seeking out experts to analyze such a large amount of ECG data can deplete valuable medical resources. Recently, there has been a growing interest in automatic methods for accurate heartbeat categorization. Our motivation behind this work is to improve efficiency and accuracy for arrhythmia classification based on convolutional neural networks (CNNs). We first pre-process the input signals by normalizing them using a z-score. Following the data segmentation, the under-represented classes are oversampled using random duplication. By randomly duplicating the samples of underrepresented categories and eliminating instances from the overrepresented categories, this approach minimizes the imbalance in the training data. We propose an efficient 12-layer 1D-CNN to classify cardiac arrhythmia into five classes. We demonstrate the oversampling technique’s impact on the proposed CNN’s performance by comparing its performance before and after its application.