Magnetoencephalography (MEG) is used as part of the presurgical evaluation of patients with drug-resistant focal epilepsy. Its analysis consists in visually searching for transient events: interictal epileptiform discharges (IEDs)methods.

The analysis can be simplified to a classification task by working on the independent component analysis (ICA) decomposition of the MEG data. Brain related independent components (ICs) are thus classified as epileptic (pathological) or physiological (non-pathological).

In this study, we propose a new classification method to retrieve the interictal epileptic network using machine learning (ML), and particularly clustering techniques on automatically detected peaks to find relevant events in each IC's time course. A set of 65 MEG recordings with expert annotations on ICA-derived components were used to train and test classical ML and deep learning (DL) models in a leave-one-subject-out cross-validation method. The mean of each cluster was used as input to three models. A classical ML model was trained on predefined features computed on the mean of each cluster. A DL model based on a convolutional neural network (CNN) was applied to the cluster mean. Finally, another DL model composed of dense layers was applied to the coefficients obtained through wavelet transformation.

The fully connected network combined with wavelet transform achieved the best performance, with a median F1-score was 57% (95% CI, 48–67%), recall 91% (95% CI, 75–100%), and precision 50% (95% CI, 33–58%). Furthermore, in a given IC, the proportion of clusters predicted as pathological by this model increased in the presence of clear epileptic spikes.

This approach opens new perspectives to delineate the interictal epileptic networks by proposing and alternative to regular IEDs detection.