Detection of Ventricular Fibrillation from ECG Signal Using Hybrid Scalogram-Based Convolutional Neural Network

Abstract

AbstractVentricular fibrillation (VF) is a shockable ventricular cardiac arrhythmia that causes sudden cardiac death. Numerous studies around the world either utilized hand-engineered features or automatic feature extraction techniques together with different classification algorithms to detect VF from electrocardiogram (ECG) signals. This study introduces a novel hybrid scalogram-based convolutional neural network (CNN) that deploys empirical mode decomposition (EMD) on ECG signals by taking the advantage of the intrinsic mode functions (IMFs). ECG recordings that contained VF and Non-VF episodes were collected from the CU Ventricular Tachyarrhythmia Database. After preprocessing, each ECG recording was segmented at definite intervals (2, 3, and 4 s) and each segment was named either ‘VF’ or ‘Non-VF’ episode, based on the information given in the database. Using EMD, the IMFs of these segments were obtained. IMFs were then converted into hybrid scalogram images using continuous wavelet transform. A CNN model was trained by these hybrid scalogram images for detecting Non-VF and VF segments. To classify VF and Non-VF episodes, this study applied Visual Geometry Group (VGG)-19 that is a novel object-recognition system consisting of 19 layers of deep CNN. The accuracies of VGG-19 for the detection of VF using the segment interval of 2, 3, and 4 s were 95.91, 96.21, and 97.98%, respectively. This method can be employed to identify and categorize VF and Non-VF events from ECG signals where the feature extraction dilemma was mitigated.KeywordsEmpirical mode decompositionContinuous wavelet transformConvolutional neural network