Edge-Deployable Deep Neural Network for Atrial Fibrillation Detection in Ambulatory ECG

Ambulatory electrocardiogram (ECG) monitoring is essential for detecting long-term atrial fibrillation (AFib), but it remains challenging due to severe noise, motion artefacts, and signal non-stationarity. This paper proposes a lightweight deep learning architecture for AFib classification from ambulatory ECG, designed for efficient deployment on resource-constrained hardware. The model combines one-dimensional convolutional encoding, residual feature refinement, and a compact self-attention mechanism to capture both local ECG morphology and long-range temporal rhythm irregularities. Global average pooling produces a compact representation for low-complexity classification. The proposed method is evaluated on the Contextualized Ambulatory Electrocardiography Arrhythmia Dataset (CACHET-CADB), a long-term real-world ambulatory ECG dataset with expert annotations and naturally occurring noise. Experimental results demonstrate an average classification accuracy of 97.31%, outperforming existing deep learning-based approaches while maintaining low computational complexity, with a model size of only 2.67 MB, making it suitable for edge and wearable deployment.