CONVOLUTIONAL NEURAL NETWORK BASED CLASSIFICATION OF THREE CARDIAC CONDITIONS

Abstract

Clinicians and healthcare professionals use Electrocardiogram (ECG) as a heart monitoring test for identifying unusual cardiovascular activity. The abnormal cardiac behavior is dependent on the heart‟s electrical activity. Detection of real-time and precise ECG anomalies will offer helpful knowledge for providing possible treatment for patients. On a fundamental level, ECG is a time-series signal generated because of the heart's electrical activity. Different techniques have been formulated to apply Machine learning algorithms for the classification of these time-dependent signals i.e. ECG. Machine learning techniques involve manual extraction of the features thus, leading to issues such as irregularity in the extracted features along with irregularity found in the ECG features. Deep learning methodologies make use of the capabilities of Convolution Neural Networks (CNNs) i.e., they provide the benefits of automatic extraction and identification of complex and intricate features from images. Along with that, pre-trained networks which are trained using some different datasets are successful in learning and extracting features associated with new data. Using and modifying such pre-trained networks can produce desired results. AlexNet, GoogLeNet, and SqueezeNet are some of the popular CNN models, used in various classification tasks using transfer learning. In this scope, performance evaluation and comparison of these three CNN models for multi-class classification of ECG signals is presented in this thesis. Three distinct classes of ECG signal i.e. Arrhythmia (ARR), Congestive Heart failure (CHF), and Normal Sinus Rhythm (NSR) are considered in this work, which is representative of an individual‟s heart conditions. The classification methodology adopted in this thesis focuses on Continuous Wavelet Transform (CWT) which produces images having time-frequency information of available 1-dimensional samples of ECG signals. ECG dataset was collected from different PhysioBank databases. Images with time-frequency representation of ECG signals were applied as input to three CNN models. Using transfer learning approach and modification in certain layers of three models, ECG classification was performed and the performance of three DNN architectures was studied. The results revealed promising performance by three models with different internal architectures. Classification accuracies in the range of 97.22% to 97.78% were obtained. AlexNet outperformed GoogLeNet and SqueezeNet models in terms of accuracy as well as training time.