Efficient ECG compression and reconstruction via deep compressive sensing and hardware-efficient models

This paper presents an efficient deep learning-based framework for ECG signal compression and reconstruction within the Compressed Sensing (CS) paradigm. Conventional CS techniques often suffer from computational inefficiency and strict sparsity requirements, limiting their suitability for real-time biomedical applications. To overcome these challenges, we propose a lightweight neural architecture that jointly learns an adaptive sensing mechanism and an accurate reconstruction process. The proposed framework introduces two key innovations: (1) a Data-Driven Sensing Matrix (DSM) that dynamically captures the intrinsic structure of ECG signals to achieve superior compression performance, and (2) a Binary Thresholding Matrix (BTM) that converts the learned sensing matrix into a hardware-efficient binary representation. Furthermore, an autoencoder-based end-to-end model is designed to seamlessly integrate sensing and reconstruction. Experimental validation on the MIT-BIH Arrhythmia Database demonstrates that the proposed methods achieve markedly higher reconstruction accuracy and compression efficiency compared to conventional CS approaches. Deployment on an STM32 microcontroller further verifies the potential of the proposed framework for real-time, low-power ECG monitoring in edge and wearable healthcare systems.