A Data-Augmented Deep Hybrid Model for Atrial Fibrillation Detection

Atrial fibrillation (AF) is the most common sustained arrhythmia, often missed in short ECG recordings. This study proposes a hybrid approach for early AF detection using long-term Holter ECG data from the SHDB-AF dataset. The method combines data augmentation with signal decomposition via Empirical Mode Decomposition (EMD) and Continuous Wavelet Transform (CWT). A Convolutional Neural Network (CNN) extracts features from augmented signals, which are then classified using a Support Vector Machine (SVM). Using a fixed set of the first five IMFs, the model achieved 96.8% accuracy, 99% AUC, and 0.968 F1-score, demonstrating strong potential for reliable AF prediction in clinical settings.