A novel multichannel sparse convolutional autoencoder for electrocardiogram signal compression

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Continuous electrocardiogram (ECG) monitoring is crucial for cardiac patients.
• ECG data volume presents challenges for storage and transmission.
• Deep learning, specifically autoencoders, shows promise for ECG signal compression.

Purpose of the Study:

• To develop a novel multichannel convolutional autoencoder for efficient ECG signal compression.
• To explore the impact of sparsity constraints and quantization on compression performance.
• To evaluate the model's effectiveness in reducing data size while preserving signal integrity.

Main Methods:

• A multichannel convolutional autoencoder architecture was designed.
• ECG signals were encoded into a four-channel lower-dimensional space.
• Sparsity constraints were applied to specific channels, followed by channel-wise quantization.
• Huffman coding was used for final data encoding.

Main Results:

• The proposed model achieved an average compression ratio of 20.23:1.
• The average normalized percent root mean square difference (PRDN) error was 9.86%.
• The method demonstrated efficient compression with high reconstruction accuracy on a benchmark dataset.

Conclusions:

• The novel multichannel convolutional autoencoder offers an effective solution for ECG signal compression.
• The integration of sparsity constraints and quantization optimizes compression performance.
• This approach significantly reduces data storage and transmission costs for cardiac monitoring.