Clustering strategies for optimal trial selection in multisensor environments. An eigenvector based approach.

Background: Quite often, magnetoencephalography (MEG) measurements are contaminated by a series of artifacts that degrade the quality of the various source localization methods applied to them. In particular, eye blinking, minor head movement and related activities are a constant source of measurement contamination. In order to solve this problem, trial selection and rejection is applied, a task that is usually performed manually.

New Method: The present work shows an automatic trial selection and rejection algorithm based on clustering techniques. These techniques employ a measurement of the dissimilarity of the items belonging to a set. This measure, based on the projection of the eigenvector corresponding to the largest eigenvalue of the covariance matrix, is provided and its rationale is explained. Subsequently, covariance matrices belonging to the selected cluster are averaged and used in the well-known Linearly Constrained Minimum Variance (LCMV) Beamformer.

Results: The results show a marked improvement of the specificity of the localization algorithm compared to the application of the LCMV without clustering.

Comparison With Existing Method(s): The method shows a marked reduction in computational cost compared with other data cleaning procedure widely used: Independent Component Analysis (ICA).

Conclusions: Thus, we propose clustering techniques to be used in brain localization activity algorithms.