Cardiac Arrhythmias Detection Using a Hybrid 2D CNN-SVM Model

Time-domain analysis of ECG signals waveform is the most common technique for detecting cardiac arrhythmia. The analysis aims to detect relevant patterns in amplitude, duration, or shape variations in ECG signal weaves. Handcrafting feature extraction from raw ECG signals is time-consuming, computationally expensive, and requires access to powerful hardware resources. Feature selection is a process of removing irrelevant and/or redundant extracted features and retaining only relevant features. Handcrafting feature extraction and selection can be bypassed using deep learning models since the models can learn, extract, and select the features automatically in their hidden layers. This paper proposes a hybrid 2D CNN-SVM model that combines a deep learning model as a feature extractor and selector, and a machine learning model as a shallow classifier. Raw ECG signals are filtered, normalized, resampled, and segmented into single heartbeats lasting 1.5 s. The motivation behind converting ECG signals in the form of 2D images is to extract time-domain and deep features contained in the grid-like data. However, the computational complexity and resource consumption of the CNN are increasing over time. To overcome these problems an SVM classifier is used as classifier. Experimental results indicate that our proposed model achieves an average accuracy of 98.46% on different dataset sizes, which has better accuracy with lower complexity compared to the existing machine and deep learning models.