DeepSeparator

DeepSeparator separates EOG and EMG artifacts from EEG recordings using a deep learning framework that separates artifacts from neural signals in an embedding space and reconstructs denoised signals.

Key Features:

• Deep learning-based separation: Uses deep learning to distinguish artifacts from neural activity within an embedding space.
• Encoder component: Extracts and amplifies features from raw EEG to highlight characteristics relevant for artifact separation.
• Decomposer component: Identifies trends and isolates artifact components while aiming to preserve neural signal integrity.
• Decoder component: Reconstructs denoised EEG signals from cleaned feature representations.
• Artifact types targeted: Explicitly addresses ocular (EOG) and muscle (EMG) artifacts in EEG recordings.
• Interpretability: Enables explicit extraction of artifact components to support model interpretability.
• Multi-channel and variable-length support: Applicable to multi-channel EEG recordings and datasets of arbitrary length.

Scientific Applications:

• Artifact removal in EEG: Removes EOG and EMG contamination from EEG recordings to improve signal quality.
• Evaluation datasets: Evaluated on semi-synthetic and real task-related EEG datasets for performance assessment.
• Neural signal analysis: Provides denoised EEG suitable for downstream analysis in research and clinical studies.

Methodology:

The computational pipeline comprises an encoder that extracts and amplifies features from raw EEG, a decomposer that identifies trends and separates/suppresses artifact components while preserving neural signals, and a decoder that reconstructs the denoised EEG signal from cleaned features.