CAF-Net: A Convolution and Attention Fusion Network with RR Interval Context for Five-Class Arrhythmia Detection from Single-Lead ECG

Abstract

Automatic arrhythmia detection from the electrocardiogram (ECG) is a central component of modern cardiac screening, yet accurate discrimination of rare beat types from a single lead remains difficult. This paper proposes CAF-Net, a convolution and attention fusion network that combines a residual convolutional encoder for local beat morphology, a two-layer multi-head self-attention encoder for global intra-beat context, and a parallel fully connected branch that embeds four RR interval features describing the local rhythm. The fused representation is classified into the five beat classes recommended by the Association for the Advancement of Medical Instrumentation, namely normal (N), supraventricular ectopic (S), ventricular ectopic (V), fusion (F), and paced or unknown (Q). On 109,398 heartbeats extracted from all 48 records of the MIT-BIH Arrhythmia Database, CAF-Net attains a mean overall accuracy of 99.35 percent and a mean macro F1 score of 0.952 across three independent runs, with a mean area under the receiver operating characteristic curve of 0.9981, without any synthetic oversampling. The network contains only 446,309 parameters and classifies one beat in 4.0 ms on a single CPU thread, which makes it suitable for wearable and point-of-care deployment. Learned attention weights concentrate on diagnostically relevant waveform regions, providing a degree of interpretability.