SSDLabeler: realistic semi-synthetic data generation for multi-label artifact classification in EEG.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Electroencephalography (EEG) recordings are prone to artifacts (ocular, muscular, environmental noise) that obscure neural signals.
• Artifact classification offers a transparent alternative to Independent Component Analysis (ICA)-based methods but requires extensive manual labeling for training data.
• Existing semi-synthetic data (SSD) methods lack realism due to single artifact injection or reliance on separate artifact recordings.

Purpose of the Study:

• To develop a novel framework, SSDLabeler, for generating realistic, annotated semi-synthetic EEG data.
• To address limitations of previous SSD methods by enabling the injection of multiple artifact types.
• To improve the accuracy and scalability of EEG artifact classification.

Main Methods:

• Decomposition of real EEG data using ICA.
• Epoch-level artifact verification employing Root Mean Square (RMS) and Power Spectral Density (PSD) criteria.
• Reinjection of multiple artifact types into clean EEG data to create annotated SSDs.
• Training a multi-label artifact classifier using the generated SSDs.

Main Results:

• The SSDLabeler framework successfully generated realistic, annotated semi-synthetic EEG data.
• Training artifact classifiers with SSDLabeler-generated data significantly improved accuracy on raw EEG across diverse conditions.
• Performance surpassed classifiers trained on prior SSD methods or raw EEG data alone.

Conclusions:

• SSDLabeler provides a scalable foundation for robust EEG artifact handling.
• The method effectively captures the co-occurrence and complexity of real-world EEG artifacts.
• This approach enhances the reliability and applicability of automated artifact classification in EEG analysis.