Unsupervised Event Characterization and Detection in Multichannel Signals: An EEG application

Abstract

In this paper, we propose a new unsupervised method to automatically characterize and detect events in multichannel signals. This method is used to identify artifacts in electroencephalogram (EEG) recordings of brain activity. The proposed algorithm has been evaluated and compared with a supervised method. To this end an example of the performance of the algorithm to detect artifacts is shown. The results show that although both methods obtain similar classification, the proposed method allows detecting events without training data and can also be applied in signals whose events are unknown a priori. Furthermore, the proposed method provides an optimal window whereby an optimal detection and characterization of events is found. The detection of events can be applied in real-time.

An Unsupervised Method for Artefact Removal in EEG Signals

“…In this paper, the test signal is an 8-min EEG recording used in [15,16] to show the performance of DETECT toolbox and the UMED method. The data are sampled at 256 Hz using a 64-channel Biosemi Active Two System.…”

“…In Section 2.1 UMED’s basic principles have been presented (full details can be found in [16]). Using UMED on the AEEG described above ( L w = 155 samples and d = 32 samples) some clusters of intervals are found.…”

“…For example, statistical methods and/or threshold values can be used [2,12,13,14]. Classification methods are an alternative to detect artefacts, for example, DETECT [15] and the unsupervised method for event detection (UMED) [16]. These methods detect artefacts by means of an identification or characterization procedure.…”

“…The Unsupervised Method for Event Detection (UMED) method [16] is used to detect events in a multichannel signal in an unsupervised way. In this method, the multichannel signal is divided into a set of intervals created from a sliding window (with a sliding distance of d samples).…”

Automatic bad channel detection in intracranial electroencephalographic recordings using ensemble machine learning

“…Since bad channels can spoil the quantitative analysis and interpretation of iEEG signals, it is important, specifically for large datasets, to develop methodological approaches that automatically identify them. There are different approaches available in literature that address this challenging task for scalp electroencephalographic (EEG) signals ( Shoker et al, 2005 , Nolan et al, 2010 , Mognon et al, 2011 , Lawhern et al, 2013 , Alotaiby et al, 2015 , Mur et al, 2016 ). Most often, bad channel detection methods build upon the high spatial correlation of EEG signals and thus predict the value of each channel at each time point from the activity of all other channels at the corresponding time points.…”