A feature enhanced EEG compression model using asymmetric encoding-decoding network.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Neuroscience

Background:

• Wearable devices increasingly utilize electroencephalography (EEG) for data acquisition.
• Existing EEG compression methods face challenges with parameter volume and low signal-to-noise ratio, limiting their use in resource-constrained wearable devices.
• Suboptimal compression leads to high reconstruction errors and loss of critical signal information.

Purpose of the Study:

• To develop a tailored EEG compression algorithm for wearable devices.
• To address limitations of current methods regarding computational constraints and signal fidelity.
• To improve the efficiency and effectiveness of EEG data transmission and analysis from wearable devices.

Main Methods:

• Proposed a feature-enhanced asymmetric encoding-decoding network for EEG compression.
• Employed a lightweight model for encoding EEG signals.
• Utilized a multi-level feature fusion network with a two-branch structure for decoding and signal reconstruction.
• Validated the method on public EEG datasets, including motor imagery and event-related potentials.

Main Results:

• Achieved state-of-the-art compression performance on public EEG datasets.
• Demonstrated that the method retains more task-related information as compression ratio increases.
• Neural representation analysis and classification performance confirmed the preservation of reliable discriminative information post-compression.
• The lightweight design is suitable for wearable devices with limited computing and storage.

Conclusions:

• The proposed asymmetric EEG compression method is tailored for wearable devices.
• Achieves superior compression performance while maintaining signal integrity and task-relevant information.
• Paves the way for enhanced application of EEG-based wearable technology.