Classification of magnetoencephalographic independent components in epilepsy by machine learning

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Imaging

Background:

• Magnetoencephalography (MEG) is crucial for pre-surgical epilepsy assessment.
• Current MEG analysis is time-consuming and subjective.
• Automating MEG interpretation can improve clinical workflows.

Purpose of the Study:

• To combine Independent Component Analysis (ICA) with machine learning (ML) for automated MEG signal interpretation.
• To develop ML models for classifying independent components in MEG data.
• To differentiate between artifactual, epileptic, and physiological brain activity.

Main Methods:

• Recorded MEG data from 41 patients with drug-resistant focal epilepsy.
• Trained Random Forest (RF) and logistic regression models on 61 features.
• Classified components into artifact, heart, noise, epileptic, and physiological categories.

Main Results:

• Achieved high accuracy (F1-score > 0.9) in classifying artifact components using RF.
• Attained balanced accuracy > 0.8 for classifying four component types.
• Epileptic component classification showed moderate performance (F1 ~0.5) with significant inter-patient variability.

Conclusions:

• ICA combined with RF offers an efficient method for artifact classification in MEG.
• Differentiating epileptic from physiological activity is challenging, but specific features show potential as biomarkers.
• The study highlights the potential and limitations of ICA for classifying epileptic and artifactual components.