STAND-Net: A Spiking Temporal Attention autoeNcoDer Network for Efficient EEG Artifact Removal.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Electroencephalography (EEG)-based brain-computer interface (BCI) systems offer vast potential but are hindered by physiological artifacts that degrade signal quality.
• Current deep neural networks (DNNs) for artifact rejection are computationally expensive, limiting their use in wearable BCI devices.

Purpose of the Study:

• To develop an ultra-efficient, high-fidelity EEG artifact removal system for wearable BCI applications.
• To introduce STAND-Net, a neuromorphic architecture designed for low-power and effective artifact rejection.

Main Methods:

• Developed STAND-Net, a neuromorphic architecture utilizing spiking neurons, a spike-convolution encoder-decoder, and a spike-rate attention mechanism.
• Modeled spatiotemporal EEG dynamics and long-range dependencies using leaky integrate-and-fire neurons and a dilation-enhanced residual backbone.
• Employed a spike-rate attention mechanism for dynamic artifact localization based on neuronal firing patterns.

Main Results:

• Achieved >3.7 dB improvement in signal-to-distortion ratio compared to state-of-the-art methods across various artifacts.
• Demonstrated a 97.98% reduction in power consumption compared to comparable DNNs.
• Increased downstream BCI classification accuracy by 6.64% using STAND-Net processed signals.

Conclusions:

• STAND-Net provides a low-power, high-quality solution for EEG artifact removal in wearable BCI systems.
• This neuromorphic framework enables efficient and effective signal processing for improved BCI performance.
• The study establishes a new direction for developing efficient BCI systems through neuromorphic engineering.