Automatic Detection of Cardiac Pathologies from ECG Signals Using Variational Autoencoders (VAE)

Resumo

The electrocardiogram (ECG) is one of the most widely used tools for identifying cardiac pathologies (MINCHOLÉ et al., 2019). Although manual analysis of this exam can be effective, it is subject to human error and time constraints, which underscores the importance of automatic abnormality detection. In this context, the present study aims to develop an intelligent system based on Variational Autoencoder (VAE) networks for the automated detection of cardiac pathologies from ECG signals. VAEs are models that, in addition to reconstructing original inputs with high fidelity, have the ability to generate new data, which is particularly useful for analyzing complex signals such as those in ECGs. The use of this type of model can potentially improve the quality of automatic diagnoses by offering a more robust and reliable approach to identifying cardiac anomalies. For the implementation of this system, the MIT-BIH Arrhythmia Database was used, which is recognized for its diversity and richness of arrhythmia information. ECG signal preprocessing steps included noise filtering, data normalization, and categorization of pathology classes. These steps were crucial to ensure that the model was trained on high-quality and representative data. The model architecture consists of an encoder that captures the relevant features of the ECG signals, a latent space that allows for compact information representation, and a decoder that generates the reconstructions. The model evaluation focuses on analyzing the generated latent space, where samples are examined to understand how different pathologies are represented and to verify the model's ability to distinguish between them. The generation of new data from the model helped in creating balanced datasets, which are fundamental for training machine learning models in clinical scenarios. In the long term, this system can be integrated into telemedicine platforms, contributing to faster and more efficient diagnosis of cardiac conditions.