Fusion of Multi-stage CNN Features for ECG Classification

Abstract

ABSTRACT: Detecting and classifying cardiac arrhythmias is critical to the diagnosis of patients with cardiac abnormalities. Identification and classification of abnormalities are time consuming because it often requires analysing each heartbeat of the ECG recording. Moreover, computerized ECG classification can also be very useful in shortening hospital waiting lists and saving the life by discovering heart diseases at early stages. Therefore, automatic classification of the arrhythmias using machine-learning technologies can bring various benefits. In this thesis, novel and high-performance approaches based on deep learning techniques are proposed for the automatic classification of electrocardiogram (ECG) signals. In this research work, two fully automatic systems have been presented which are shown to have high efficiency and low computational cost. In one of the proposed systems, a novel decision-level fusion of features is presented by three different approaches; the first one uses normalized feature-level fusion of handcrafted global statistical and local temporal features by uniting these features into one set, the second one uses the morphological feature subset, and the third one combines features extracted from multiple layers of a Convolutional Neural Network (CNN) through using a score-level based refinement procedure. The second proposed system utilized a new architecture of deep neural networks, Directed Acyclic Graph Convolutional Neural Networks (DAG-CNNs). DAG-CNNs fuse the feature extraction and classification stages of the ECG classification into a single automated learning procedure and utilize the multi-scale features and perform the score-level fusion of multiple classifiers automatically. The results over the MIT-BIH arrhythmia benchmark database exhibited that the proposed systems achieve superior classification accuracy compared to all of the state-of-the-art ECG classification methods. Keywords: electrocardiogram, convolutional neural networks, directed acyclic graph CNN, morphological feature, statistical feature, temporal features, multi-stage CNN-based features, feature-level fusion, score-level fusion, decision-level fusion