Diagnóstico automatizado de fibrilação atrial em ECG de dispositivos vestíveis com Deep Learning

Abstract

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia in the adult population and is associated with an increased risk of stroke, heart failure, and mortality. With the widespread adoption of wearable devices, continuous ambulatory monitoring of the electrocardiogram (ECG) has become feasible, enabling early detection of arrhythmias. In this work, a system based on a smartwatch was developed to perform automatic classification of cardiac rhythms from 30-second ECG windows using an optimized convolutional neural network (DMKS). The methodology employed the PhysioNet/Computing in Cardiology Challenge 2017 dataset for training, together with a signal processing pipeline including length standardization, Butterworth filtering using filtfilt, and Min–Max normalization. Controlled generation of noisy examples and additional data augmentation techniques were applied to increase data diversity while preserving the natural class imbalance observed in real-world scenarios. The selected architecture was converted to TensorFlow Lite for on-device deployment. The results demonstrate robust classifier performance, with an overall F1-score of approximately 0.93 and high sensitivity in the detection of atrial fibrillation episodes (recall ≈ 0.89; F1-score ≈ 0.89). Post-training quantization and conversion to TensorFlow Lite enabled efficient on-device inference while maintaining performance comparable to the non-quantized model. It is concluded that the proposed solution is promising for automatic AF screening from smartwatch-acquired signals, achieving high sensitivity for AF and a high overall F1-score.