关键词: 心电图, 房颤检测, 高斯小波, 神经网络, 注意力机制

Abstract: Atrial fibrillation (AF) is a common and cryptogenic cardiac condition, the accurate diagnosis of which is essential for preventing severe complications. To improve the accuracy of atrial fibrillation diagnosis, a lightweight optimized aggregation network combining spectral features and spatial channel optimization is proposed. Firstly, in response to the weak and irregular nature of atrial fibrillation signals, Gaussian wavelet transform is used to convert ECG signals into spectrograms, effectively capturing their detailed features and energy in both frequency and time domains. Secondly, a spatial channel optimization module is proposed to enhance the model's ability to capture atrial fibrillation signals by separating and reconstructing feature information, while reducing the interference of redundant information. Finally, an attention enhancement single aggregation module is added to enhance the model's attention to key feature regions during the classification process by fusing convolutional features and 3D attention mechanisms three times, enabling the model to focus on atrial fibrillation feature regions.Experimental results on the MIT-BIH AF database and the 2017 PhysioNet/CinC Challenge dataset demonstrate that the proposed method achieves accuracy rates of 99.41% and 98.59%, respectively, with a model size of only 48.43% that of ResNet and a parameter count of only 3.45% that of the Vision Transformer. 

Key words: electrocardiogram, atrial fibrillation detection, Gaussian wavelet, neural net, attention mechanism