A Multi-lead Automatic Arrhythmia Classification Based on Hybrid Feature Extraction and Machine Learning Classifiers

Electrocardiogram data is a powerful and important biomedical body signal utilized in clinical diagnosis for heart disease.It is highly exhausting for physicians to instantly analyse lengthy electrocardiogram (ECG) data records from more than one lead therefore, computer aided diagnostics systems are essential to diagnose heart diseases.This paper proposes an automatic arrhythmia classification method using combined features.After detecting the beats automatically in each electrocardiogram lead data, each beat in a single lead is converted into a red, green and blue (RGB) channel image.Features are extracted from these images using histogram of oriented gradients (HOG) algorithm, the dimension of the HOG feature descriptors depend on the block number and image dimension.Other features are also extracted using sixteen customized features from a second lead.The combination of the HOG descriptors and the customized features form a single vector and considered as total features.This procedure reduces the dimension of the features and eventually decreases the time required for classification using a multi-lead arrhythmia classification.Moreover, these two types of features are concatenated and used to classify six arrhythmia classes using neural network and histogram gradient boosting (HGB) classifiers.Results show that HOG feature descriptors achieved an accuracy of 98.8% for the conventional neural network classifier and an accuracy of 98.3% using the HGB classifier.When implementing the hybrid features of HOG and the sixteen customized morphological features, the algorithm achieved an accuracy of 99.8% for neural network classifier and an accuracy of 99.5% for histogram gradient boosting classifier.The experiments conducted in this research successfully achieved its goals of obtaining the desired outcomes of arrhythmia classification with high feature dimension reduction.The findings of the proposed approach exceeded prior studies of algorithms.The algorithm was tested on the St. Petersburg, INCART 12-lead and MIT-BIH ECG databases.