A sparse Bayesian learning based network for energy-efficient ECG compressed sensing

Abstract

In recent years, wireless body-area networks (WBANs) have become prevalent for remote electrocardiogram (ECG) monitoring. However, the long-term operation of these systems demands significant energy from sensors. To address this, it is essential to streamline signal acquisition and reduce signal dimensionality, thereby decreasing communication bandwidth and on-chip power usage. Compressed sensing (CS), an emerging sampling technique, has been increasingly adopted for remote ECG monitoring. While traditional CS methods enhance reconstruction precision by using signal features as prior knowledge, they do not fully exploit the potential of these priors. This paper introduces a hybrid approach, PC-BCSNet, which combines the CS-based framework of pattern-coupled sparse Bayesian learning (PC-SBL) with a data-driven deep learning method. This dual-driven architecture develops a generalized prior model for post-sparsification ECG signals, employing the generalized approximate message passing (GAMP) algorithm for rapid reconstruction. Furthermore, an interpretable deep iterative neural network is designed to execute the full iterative Bayesian inference process. The scale parameters of the prior model serve as trainable weights, capturing features specific to ECG signals. Experiments demonstrate that PC-BCSNet significantly outperforms other state-of-the-art algorithms in reconstruction accuracy and speed, as evaluated on the European ST-T and MIT-BIT Arrhythmia databases. Notably, our network design adapts readily to changes in measurement matrices, providing enhanced flexibility and robustness for practical applications.