Akademik Veri Yönetim Sistemi
Circuits, Systems, and Signal Processing, 2026 (SCI-Expanded, Scopus)
Electrocardiogram (ECG) compression reduces bandwidth and storage needs, enabling real-time transmission and long-term monitoring. However, many traditional methods have difficulty reconstructing signals at low sensing rates. Deep learning (DL) can improve reconstruction performance in this regime. In this study, a DL-based ECG compression and reconstruction method is proposed that combines compressed sensing (CS) with an unrolled architecture built on dilated convolutional neural network (CNN) blocks. The sensing operator is learned during training, and reconstruction proceeds through unrolled iterations that include a data-consistency update and a dilated CNN correction module. Adaptive thresholding is applied inside the convolutional block, and a sample-dependent scaling mechanism adjusts the correction term before it is added to the current estimate. In addition, a multi-SR training setting is considered, allowing a single set of parameters and a shared learned measurement matrix to be used across different SR values without retraining. The method is evaluated on the MIT-BIH Arrhythmia Database and the Non-Invasive Fetal ECG Arrhythmia Database (NIFEA) using PRD and SNR. Furthermore, it is compared with DL-based methods and classical CS algorithms. The proposed approach achieves lower PRD and higher SNR than competing methods across sensing rates. At SR = 0.50 on MIT-BIH, the proposed method achieves about 1.8 dB higher SNR and about 0.23 percentage points lower PRD than the best-performing competing method. At SR = 0.50 on NIFEA, the proposed method achieves about 3.1 dB higher SNR and about 0.49 percentage points lower PRD than the best-performing competing method.