Signal Reconstruction Method Based CNN for Atrial Fibrillation Detection

Wearable devices can monitor patients' heart activity in real-time, supporting early detection and treatment. However, ensuring high performance while maintaining low computational requirements is crucial for wearable devices. To enhance disease identification by providing more dimensions and features, and to improve accuracy and interpretability of data analysis, this study explores an innovative signal reconstruction method. By converting one-dimensional (I-D) time series into two-dimensional (2-D) signal matrices, this method better captures important information in the time series. Subsequently, by employing multi-scale convolutional kernels, the network structure is deepened, enabling the capture of spatial information across different ranges. This approach helps learn more complex features while effectively reducing the number of parameters. On the MIT-BIH Atrial Fibrillation Database, this method achieved an accuracy of 99.30%. Furthermore, validation on a clinical dataset yielded an accuracy of 99.04%. Compared to one-dimensional time series, this method improved accuracy by 1.17% in atrial fibrillation identification tasks. Additionally, compared to other common types of 2-D images, the new input format significantly reduces the computational burden during model training, making it more suitable for practical applications in wearable devices.