CSNet: A deep learning approach for ECG compressed sensing

Abstract

Remote electrocardiogram (ECG) monitoring plays a very important role in the prevention and treatment of cardiovascular diseases. However, the current long-term ECG monitoring generates a large amount of data, which puts great pressure on the bandwidth and transmission systems. Compressed sensing (CS) has great attraction for resource-limited wearable devices used in remote ECG monitoring. Traditional CS reconstruction algorithms often require complex signal processing and prior knowledge, and the reconstruction process is time-consuming, which limits the application of CS in remote ECG monitoring systems. This paper proposes a non-iterative fast reconstruction algorithm based on CS and deep learning, which combines convolutional neural network (CNN) and long short-term memory (LSTM) to directly learn the mapping relationship between the rising dimension signal of the measurements and the original signal, and is validated in the MIT-BIH Arrhythmia Database (MITDB). The experimental results show that the reconstruction error of this method is lower than that of traditional algorithms including basis pursuit (BP), orthogonal matching pursuit (OMP), bound-optimization-based block sparse Bayesian learning (BSBL-BO) and rotate-singular value decomposition + basis pursuit (R-SVD+BP) at high compression ratios (CRs, when CR⩾50%). At the same time, for a 30-min ECG signal, only about 0.12 s is needed to complete the reconstruction, which is at least 45 times faster than the traditional algorithms, which is enough to support real-time applications. In addition, the proposed method is also validated in MIT-BIH Normal Sinus Rhythm Database (NSRDB), MIT-BIH Atrial Fibrillation Database (AFDB) and European ST-T Database (EDB), and the reconstructed signals of NSRDB and AFDB meet the clinical requirements at CR⩽70%, and the reconstructed signals of EDB meet the clinical requirements at CR⩽90%.