A Deep Learning Method for Automatic Detection of Most Relevant ECG Features

The main cause of mortality in the world are cardiovascular diseases (CVDs). Each year, around 17.9 million people die from CVDs and this number represents the 32% of all global deaths. Early diagnosis and treatment are very important for people which present high cardiovascular risk. Irregularities in the heartbeat rhythm are defined "Arrhythmias" and they can rarely occur during human's life. This kind of problems can lead to some complications that may constitute an immediate risk for life and may cause potentially fatal events. The premature classification and detection of arrhythmias is a good starting point for cardiac disease diagnosis. The most important element to detect these events is the electrocardiogram (ECG), a non-expensive, non-invasive method, which gives a record of heart's electrical functionality. The interpretation process of ECG requires a high degree of training and is time-consuming. So, the first attempts to automate the interpretation of ECG are dated at the end of 1950s. The ECG is used for diagnosis of cardiovascular disease, and it mainly consists of three principal waves: P-wave, QRS complex and T-wave. The most relevant feature is the QRS complex because it represents the heartbeat, i.e., the ventricular contraction. It can also be the starting point for obtaining further useful medical information. Thus, for ECG-based evaluation, the QRS detection is critical. ECG signal, hence, contains a huge amount of information and it is difficult to analyse them just with a visual evaluation. For this reason, the design of an automated efficient system to detect the relevant features in an ECG signal is a critical task. A new interesting approach, developed in the last decade, is Deep Learning. The huge impact of Deep Learning has motivated the implementation of this methodology for automatic classification of ECG features. In particular, the application of Deep Neural Networks in electrocardiographic signals is gaining importance to explore the enormous quantity of information that these signals contain. Today, deep learning enlarges the vision, introducing new methods to achieve better accuracy and increase time management in ECG features detection. This thesis project focuses on a deep learning method to automatic detect the most relevant features in ECG signals. The learning-based approach is hybrid because it combines two different learning models. After signal pre-processing, using local regression, data are downsampled, then 1D ECG signals are converted into 2D Scalogram images to make easier the feature extraction. Finally, two methods are combined to construct the model: Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory Networks (BiLSTMs), so that a hybrid model, called CNN-BiLSTM, is constructed. For training purposes, a k-fold cross validation (with k = 10) was used to test the model's ability to predict data that was never seen before. At the beginning, for training and testing of the proposed method and for evaluating the performances of the CNN-BiLSTM approach, a publicly available dataset, named "QT Database", was used. With this large amount of data, the proposed method provides an accuracy of 97.5% in QRS complexes detection.