Deep Compressive Sensing on ECG Signals with Modified Inception Block and LSTM

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Electrocardiogram (ECG) monitoring faces challenges in managing data volume and energy consumption.
• Compressed sensing (CS) offers a solution by enabling simultaneous data under-sampling and reconstruction.
• Deep learning (DL) has enhanced traditional CS methods, overcoming prior limitations.

Purpose of the Study:

• To propose a novel deep compressed-sensing (CS) scheme for ECG signals.
• To improve the efficiency and accuracy of ECG data acquisition and reconstruction.
• To address the data burden and energy costs associated with ECG monitoring.

Main Methods:

• A deep learning framework combining a modified Inception block and long short-term memory (LSTM) was developed.
• The scheme includes preprocessing, adaptive compression using convolutional layers, and signal reconstruction.
• Experiments were conducted on the MIT-BIH Arrhythmia and Non-Invasive Fetal ECG Arrhythmia databases.

Main Results:

• The proposed scheme achieved the lowest percentage Root-mean-square Difference (PRD) and highest Signal-to-Noise Ratio (SNR) across all tested sensing rates.
• When the sensing rate exceeded 0.5, the PRD dropped below 2%, indicating substantial reconstruction improvement.
• The model demonstrated robust performance and good signal recovery quality even with noisy ECG data.

Conclusions:

• The deep CS scheme effectively reduces data burden and energy costs in ECG monitoring.
• The integration of modified Inception blocks and LSTM significantly enhances ECG signal reconstruction performance.
• This approach offers a promising solution for efficient and accurate remote or wearable ECG monitoring systems.