Real-Time Deep Compressed Sensing Reconstruction for Electrocardiogram Signals

The rapid development of wearable device technology provides an efficient way of data acquisition for remote ECG monitoring and identification. However, existing iteration based signal recovery methods have high latency, while the deep learning based method have a shortcoming that large increase in parameters makes training more difficult as the signal length increases. In this paper, we combine compressed sensing and generative adversarial networks to propose a signal recovery method based on dilated convolution. The proposed model can accept more prior information from compressed long signal without increasing parameters and achieve feature domain self-adaptation by fitting the distribution of reconstructed and original signals. Experiments result on MIT-BIH and PTB datasets demonstrate that the proposed method achieves comparable or better results in reconstruction accuracy and reconstruction time when compared to some existing iteration-based methods and some deep learning based methods. For example, reconstructing a 2s signal takes only 0.013s, which is a 50% improvement over other deep learning methods.