Automatic Arrhythmia Detection Using Deep Learning

Saeka Rahman, Miad Faezipour

Abstract: Arrhythmia is a common cardiac disease that can be a potential life threat. Arrhythmias are detected using an Electrocardiogram (ECG) signal record. ECGs are usually performed with 12 lead channels. However, this arrangement requires extensive equipment and expertise to handle. The doctors also need much time to find problematic instances when there are many records. A single-channel system is necessary to make a portable, wearable embedded system to detect arrhythmias. Such a system could assist the patients for frequent monitoring and the caregivers to follow up treatment plans and/or immediate actions. Therefore, an automated arrhythmia-detecting mechanism using single-channel information is required to improve cardiac treatment. In so doing, a Convolutional Neural Network (CNN) based lightweight deep learning model is developed to detect arrhythmia. The dataset is prepared from the MIT-BIH arrhythmia database of ECG records to validate the model. 2-D images are taken from the snapshots of R-R peaks, including regular, ventricular premature beats, and atrial premature beats of the MLII lead. The proposed classifier's overall performance on test data to detect arrhythmia and the normal rhythm is 0.9167, with 8,658 parameters.