Decoding imagined speech from EEG signals using hybrid-scale spatial-temporal dilated convolution network.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Brain-computer interfaces (BCIs) offer communication for individuals with severe motor impairments.
• Decoding imagined speech from electroencephalogram (EEG) signals is crucial for intuitive BCIs.
• Existing methods struggle to capture complex spatial-temporal dependencies in EEG data.

Purpose of the Study:

• To develop a novel model for enhanced EEG-based imagined speech recognition.
• To effectively integrate spatial and temporal feature learning for improved decoding accuracy.
• To address limitations in capturing long-range contextual cues within EEG signals.

Main Methods:

• Proposed a hybrid-scale spatial-temporal dilated convolution network (HS-STDCN).
• Employed hybrid-scale temporal convolution for multi-level temporal feature extraction.
• Utilized depthwise spatial convolution to model electrode relationships and dilated convolutions for long-range feature learning.

Main Results:

• The HS-STDCN achieved an average classification accuracy of 54.31% for decoding eight imagined words.
• This performance was significantly superior to existing methods (p < 0.05).
• Analysis revealed insights into word semantics, important brain regions, and electrode efficiency.

Conclusions:

• The HS-STDCN model effectively leverages both temporal and spatial EEG signal dependencies.
• This approach offers a promising advancement for imagined speech recognition in BCIs.
• The study highlights the potential for efficient BCI design using fewer electrodes.