An electrocardiogram (ECG) is one of the most powerful tools to inform about the functionality of the cardiovascular system. It could be defined as the graphical representation of the heart activity during a period of time and is used to determine different cardiovascular diseases and metabolic alterations. The way is a EGC is studied can be viewed in two approaches, the pattern recognition, as it is usually done, and the understanding of the exact electrical vectors recorded by an ECG as they relate to cardiac electrophysiology, which is more difficult due to the quantitative aspect of the process.

The focus for studying this at large scales is on the analysis of huge datasets of ECG from different patients to classify the different health issues found in them. Commonly, this classification is done by humans, which convert the task in a really time consuming labour and makes it really prone to errors. Categorizing and detecting different waveforms and morphologies in the signal is one of the main task to succeed in the classification. So at the end of the day it is up to subjectivity of the doctor to classify a ECG signal into one category or another.

In order to automatize this process and avoid the errors provided by the human biases, some computational approaches are done. In the last years, more specifically, many machine learning approaches have been suggested. The usual way to face this classification problem it to analyze a one, or more, massive datasets of ECG signals labeled by pri,or.

Most of these approaches involve a preprocessing step for preparing the signal. And then, these handcrafted features, which are mostly statistical summarizations of signal windows, are extracted from these signals and used in further analysis for the final classification task. For this final classification step, conventional machine learning approaches for ECG analysis have been developed in terms of support vector machines, multi-layer perceptrons and decision trees.

These handcrafted features provide an acceptable representation of the signal, but based on recent machine learning studies, automated feature extraction and representation methods are proven to be more scalable and are capable of making more accurate predictions, this allow the model to learn the features that are best suited to the specific task that it is dedicated to carry out. This approach provides a more accurate representation of ECG signal, and using them, the models can compete with a human cardiologist in analyzing the signal.

This work is organized as follows: after this introduction, in the section 2 we make a presentation and exploration of the dataset we are considering for our study; in the section 3 we are applying different supervised methods in order to discover hidden patters of the data without the use of labels; then in section 4 we are using the labels to implement many supervised methods to get a taste of the accuracy that we can obtain and to establish a benchmark for more complicated models; in section 5 we are applying some deep neuronal network architectures in order to study how the accuracy of the simpler models can be improved; finally in section 6 we are doing a recap and a comparison between all the presented methods. All the models exposed are developed by ourselves in Python using libraries such as sklearn and keras.