Content area
Full text
Received Nov 9, 2017; Revised Mar 8, 2018; Accepted Mar 29, 2018
This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
1. Introduction
In magnetoencephalography (MEG) recordings, electrophysiological brain activity is measured noninvasively by means of detecting small magnetic field changes of the electrical activity within the human brain. In addition to environmental noise, in MEG recordings the neuromagnetic brain signals interfere with strong magnetic field components from ocular and cardiac activities. As the signal strength of these biological artifacts is relatively large compared to the brain signal, the artificial signals need to be separated and removed prior to analysis.
For successful isolation of ocular and cardiac artifacts from MEG and electroencephalography (EEG) recordings independent component analysis (ICA) has been extensively and successfully used over the last ten years [1–8]. Semi- and fully automatic ICA-based approaches use different strategies to identify ocular and cardiac activities from a set of independent components. To tackle this problem, a variety of time-series-, frequency-, and topographical-based models have been proposed [1, 3, 7, 9, 10]. In time-series based models, a statistical evaluation on the decomposed signals [3, 11, 12] is typically performed, while spatial-based approaches often perform a linear transformation to the sensor space [9, 13] to localize or estimate the origin of the source [14]. In many of these approaches a reference signal is needed, such as recordings from electrocardiography (ECG) and electrooculography (EOG), which is used to identify the features of interest. The signal quality of such recordings strongly influences the reliability of the artifact rejection method applied. In particular, poor electrode conductivities, positioning of the electrodes, and even slight arm or chest movements will have a large impact on the ECG and EOG signal quality. In only a few studies, the artifact rejection methods applied have not relied on additional reference signals [3, 15–17].
In recent years, statistical machine learning methods have been introduced into the field of neuroimaging [18, 19]. These methods have been increasingly used for feature extraction, classification, and decoding in MEG and EEG [20–22]. The different approaches can be organized into supervised, unsupervised, semisupervised, or reinforcement learning based on the desired...