Atrial Fibrillation Classification Using a Deep Spectral Autoencoder
Abstract
This study introduces a novel approach foratrial fibrillation classification using a deep autoencoderstacked with a fully connected softmax layer. Themodel is trained with spectral features extracted fromECG signals through spectral and signal analysis.The primary goal is to enhance existing algorithms inthe state-of-the-art by delivering superior results withreduced computational cost. The PhysioNet Challenge2017 database was used, which contains normal andatrial fibrillation rhythms. The signals were normalizedbefore spectral feature extraction. These features wereused to train the autoencoder, which performedadditional feature extraction and dimensionalityreduction. The resulting features were then used to trainthe fully connected layer responsible for classification.Performance was evaluated through quality metricsmentioned in the state-of-the-art using cross-validationto ensure robustness. The best median results obtainedwere: 99.7% Accuracy, 99.8% Precision, 99.5% Recall,99.8% Specificity, 99.7% F1-score, 99.3% MatthewsCorrelation, and 99.3% Kappa.
Keywords
Autoencoder, spectral features, ECG signals, signal analysis, feature extraction