Evaluating the Impact of EMD-Derived Intrinsic Mode Functions on Atrial Fibrillation Detection

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia and is a major contributor to increased morbidity, mortality, and stroke risk worldwide. Early and reliable detection of AF is essential to guide timely clinical interventions. Although electrocardiography (ECG) remains the standard diagnostic tool, conventional short-term recordings often fail to capture paroxysmal AF episodes, while long-term Holter monitoring offers improved detection capability. This study explores a novel AF classification framework that integrates empirical mode decomposition (EMD), wavelet transform (WT), and 2D convolutional neural networks (CNNs). Unlike prior work that selects IMFs based on statistical criteria, we evaluated the predictive performance of combining the first five IMFs (IMF0-IMF4) without additional selection steps. The continuous wavelet transform (CWT) is applied to generate 2D representations of these components, enabling CNN to use rich time-frequency features. The model is trained and validated using SHDB-AF, a newly introduced Japanese Holter ECG database annotated for AF detection. Our results show that IMF0 alone provides highly discriminative features, outperforming models trained on the full signal and highlighting its importance in capturing rhythm-related dynamics essential for accurate AF detection. To our knowledge, this is the first study to apply deep learning on SHDB-AF for AF classification, offering insights into fixed IMF combination strategies for robust arrhythmia prediction.