Data fusion for paroxysmal events' classification from EEG.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Spatiotemporal analysis of electroencephalography (EEG) captures channel dependencies for event classification.
• High-dimensional feature vectors in EEG analysis can introduce noise, hindering model training with limited clinical data.

Purpose of the Study:

• To investigate the classification of epileptic and non-epileptic events using temporal and spectral EEG analysis.
• To compare early-integration (EI) with two late-integration (LI) fusion schemes for combining information across EEG channels.
• To evaluate dimensionality reduction techniques like feature selection and principal component analysis.

Main Methods:

• Developed and compared three fusion schemes: EI and two LI approaches (local and global spatial training models).
• Applied temporal and spectral analysis to EEG data.
• Implemented dimensionality reduction via feature selection or principal component analysis.
• Evaluated classification architectures on EEG epochs from 11 subjects.

Main Results:

• The framework was applied to generalized epileptic seizures, psychogenic non-epileptic seizures, and vasovagal syncope.
• Late-integration (LI) fusion schemes demonstrated superior recognition accuracy compared to existing literature.
• The best performing scheme, LI with a global model, achieved 97% classification accuracy.

Conclusions:

• Late-integration (LI) fusion schemes offer improved performance for EEG-based event classification.
• The LI with a global model is particularly effective, achieving high accuracy in distinguishing between epileptic and non-epileptic events.
• The developed framework provides a robust approach for analyzing complex EEG data in clinical settings.