Classification of magnetoencephalographic independent components in epilepsy by machine learning

Objective

Magnetoencephalography (MEG) provides valuable information for the pre-surgical assessment of patients with drug-resistant focal epilepsy, but analysis is time-consuming and subjective. Our objective was to combine Independent Component Analysis (ICA) and machine learning to ease interpretation of MEG signals.

Methods

We recorded 41 patients. Machine learning models were trained to classify independent components based on a set of 61 predefined features. In a first model, based on random forest (RF), we classified artifact components versus all others. In a second model, based on RF and logistic regression, we classified 4 classes (heart, noise, epileptic, physiological (i.e. normal brain activity)).

Results

With the first model 1, we obtained F1-score and balanced accuracy above 0.9. With the second model, balanced accuracy was above 0.8. Classification of epileptic component was above chance level, but with a moderate F1 score around 0.5 – with large variability across patients. Our analysis highlighted features based on spectrum, dipolarity, connectivity, kurtosis, regularity, as well as difficulties regarding spike detection.

Conclusion

Artifact classification can be performed efficiently with a combination of ICA and random forest. Distinguishing epileptic from physiological activity is more difficult, although some features show promise as biomarkers.

Significance

Our study demonstrates both the potential and the technical limitations of ICA classification of epileptic and artifactual components.