Deep Convolutional Neural Networks for Feature-Less Automatic Classification of Independent Components in...

Interpretation of the electroencephalographic (EEG) and magnetoencephalographic (MEG) signals requires off-line artifacts removal. Since artifacts share frequencies with brain activity, filtering is insufficient. Blind source separation, mainly through independent component analysis (ICA), is the gold-standard procedure for the identification of artifacts in multi-dimensional recordings. However, a classification of brain and artifactual independent components (ICs) is still required. Since ICs exhibit recognizable patterns, classification is usually performed by experts' visual inspection. This procedure is time consuming and prone to errors. Automatic ICs classification has been explored, often through complex ICs features extraction prior to classification. Relying on deep-learning ability of self-extracting the features of interest, we investigated the capabilities of convolutional neural networks (CNNs) for off-line, automatic artifact identification through ICs without feature selection.

A CNN was applied to spectrum and topography of a large dataset of few thousand samples of ICs obtained from multi-channel EEG and MEG recordings acquired during heterogeneous experimental settings and on different subjects. CNN performances, when applied to EEG, MEG, and combined EEG and MEG ICs, were explored and compared with state-of-the-art feature-based automatic classification.

Beyond state-of-the-art automatic classification accuracies were demonstrated through cross validation (92.4% EEG, 95.4% MEG, 95.6% EEG+MEG).

High CNN classification performances were achieved through heuristical selection of machinery hyperparameters and through the CNN self-selection of the features of interest.

Considering the large data availability of multi-channel EEG and MEG recordings, CNNs may be suited for classification of ICs of multi-channel brain electrophysiological recordings.