EEGdenoiseNet

EEGdenoiseNet provides a benchmark dataset for training and evaluating deep learning models that denoise electroencephalography (EEG) signals by supplying paired clean and artifact segments for quantitative performance comparison.

Key Features:

• Comprehensive Data Composition: The dataset contains 4514 clean EEG segments, 3400 ocular artifact segments, and 5598 muscular artifact segments enabling synthesis of contaminated EEG signals with corresponding ground-truth clean data.
• Facilitation of Model Development: The paired clean and artifact segments support supervised training of deep learning models for EEG denoising and development of new denoising algorithms.
• Performance Evaluation: The dataset has been used to evaluate denoising performance of classical deep learning architectures including fully-connected networks, simple and complex convolutional networks, and recurrent neural networks under varying noise contamination levels.

Scientific Applications:

• EEG signal processing: Enables development and benchmarking of denoising methods to improve signal quality for downstream analysis.
• Neurology: Supports extraction of cleaner EEG signals for clinical and research studies of neurological conditions.
• Cognitive neuroscience: Facilitates removal of artifacts to enhance measurement of neural correlates of cognition.
• Brain-computer interfaces: Aids development of denoising approaches to improve reliability of EEG-based BCI systems.

Methodology:

Contaminated EEG signals are synthesized by combining artifact segments with clean EEG segments; supervised training and evaluation of deep learning models—including fully-connected, simple and complex convolutional networks, and recurrent neural networks—are performed to learn to distinguish noise artifacts from genuine neural signals.