Sistema integrado de aquisição e classificação automática de arritmias cardíacas usando redes neurais convolucionais

Abstract

Cardiovascular diseases have been the leading cause of global mortality over the past decades, with cardiac arrhythmias being among the most prevalent disorders in this group. These abnormalities affect the heart's electrical impulses and can be identified through the electrocardiogram (ECG), a widely used diagnostic tool that remains susceptible to human interpretation errors. In this context, this study proposes the development of an automated system for the acquisition and classification of cardiac arrhythmias into three categories: normal beats, ventricular ectopic beats, and supraventricular ectopic beats. The methodology encompasses the acquisition of real ECG signals using the AD8232 sensor connected to an ESP32 microcontroller, followed by the training of one-dimensional convolutional neural networks (1D CNNs) developed in PyTorch. Signal acquisition was performed with 12-bit resolution and a sampling rate of 977 Hz. The preprocessing stage included digital filtering, normalization, and segmentation based on R-peaks. The models were trained using the publicly available Leipzig Heart Center ECG-Database, which contains over 113,000 beats annotated by specialists. The proposed CNN architecture achieved an accuracy of approximately 92.8% during validation, demonstrating robustness in detecting the three arrhythmia classes even when tested with real signals acquired by the embedded system. The results confirm the feasibility of integrating deep learning techniques into portable monitoring devices, suggesting potential applications in clinical or home environments. [resumo fornecido pelo autor]