TFANet: A Time-Frequency Aware Network With Joint Entropy Coding for High-Ratio EEG Compression.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Large-scale electroencephalogram (EEG) data transmission and storage necessitate high-ratio compression.
• Existing EEG compression methods face challenges in balancing high compression efficiency with reconstruction quality due to statistical redundancy and high-frequency information loss.

Purpose of the Study:

• To introduce TFANet, a novel framework for high-ratio EEG compression.
• To address limitations in existing methods concerning redundancy and high-frequency information loss.

Main Methods:

• TFANet integrates autoencoder learning with entropy coding to optimize latent space distribution and reduce redundancy.
• The Frequency Attention Block (FAB) uses Fast Fourier Transform for frequency-aware compression.
• The Time-Frequency Enhancement Block (TFEB) employs Discrete Wavelet Transform and channel attention to preserve fine-grained time-frequency features.

Main Results:

• TFANet achieved an unprecedented 333x compression ratio on public EEG datasets.
• The method demonstrated superior reconstruction quality compared to existing techniques.
• TFANet effectively preserves critical EEG details even at extreme compression ratios.

Conclusions:

• TFANet offers a significant advancement in EEG data compression, enabling efficient storage and transmission.
• The framework holds potential for large-scale EEG applications, including medical diagnosis and remote monitoring.
• TFANet reduces storage and transmission costs for large EEG datasets, facilitating practical applications.