FADLEC: feature extraction and arrhythmia classification using deep learning from electrocardiograph signals

Abstract Classifying arrhythmia is an essential step in the diagnosis and monitoring of cardiovascular illness. Deep learning (DL) models are trained on the electro-cardiogram recordings found in the ECG signal dataset to accurately classify arrhythmia into five groups: Normal (N), Fusion (F), Supraventricular (S), Ventricular (V), and Unknown (Q). In the proposed work, Ant Colony Optimization (ACO) to fine-tune the hyperparameter of two potent Deep Learning (DL) architectures, Bidirectional Long Short-Term Memory (Bi-LSTM) and Fully Convolutional Network (FCN) is utilised. Initially, ECG signals are pre-processed, where Multi-Resolution Wavelet-based techniques are applied for noise removal. Afterwards, the Stationary Wavelet-Hilbert transform (SW-HT) is applied for feature extraction. Next, training, validation, and testing sets are created from the extracted feature set. After performing data balancing using the SMOTE (Synthetic Minority Over-sampling Technique) algorithm, classification using optimized deep learning models is performed. With an overall accuracy of 98.9% (ACoBi-LSTM) and 99.1% (ACoFCN) on the 5-Class (N, S, V, F, Q) arrhythmia classification in the MIT-BIH dataset, the proposed model’s performance is compared and analyzed against the existing methods.