Classificação e detecção automatizada do padrão de pré-excitação ventricular de Wolff-Parkinson-White em sinais de ECG usando redes neurais convolucionais 1D e explicação usando o Grad-CAM

Abstract

Cardiovascular diseases represent a diagnostic challenge, with Wolff-Parkinson-White (WPW) syndrome being a risky anomaly that requires early detection. This study aimed to develop and validate an artificial intelligence model to classify the WPW ventricular pre-excitation pattern in electrocardiogram (ECG) signals. The methodology used a 1D convolutional neural network, trained and evaluated on the MIT-BIH Arrhythmia Database. Preprocessing consisted of preparing the data with WPW ventricular pre-excitation samples, excluding records with atypical morphologies and rhythms, extracting and segmenting the beats, and normalizing the dataset. The model performed well, achieving 100% accuracy, precision, and recall in the internal test set. In addition to performance, interpretation by Grad-CAM validated the test set, showing that the model focuses on the delta wave, the clinical characteristic of WPW. It is concluded that the 1D convolutional neural network in ECG signals is an excellent tool for automated diagnosis, serving as a basis for clinical decision support and reinforcing the need for generalization studies on external databases.