Optimizing magnetometers arrays and pre-processing pipelines for multivariate pattern analysis

Abstract

AbstractMultivariate pattern analysis (MVPA) has proven an excellent tool in cognitive neuroscience for identifying representational-specific neuronal patterns using EEG, MEG, and MRI. Likewise, it also holds a strong promise when applied to optically-pumped magnetometer-based magnetoencephalography (OPM-MEG) data. To optimize OPM-MEG systems for MVPA experiments this study examines data from conventional MEG magnetometer arrays, focusing on appropriate noise reduction techniques for magnetometers and determining the optimal number of sensors for effective MVPA. We found that the use of signal space separation (SSS) aimed at projecting out the noise contributions not generated by the brain, significantly lowered the classification accuracy considering the 102 magnetometers. Therefore, we advise against SSS filters for magnetometers when performing MVPA. Instead, we recommend employing noise reduction techniques like signal-space projection, independent component analysis (ICA), or third-order gradient noise reduction based on reference sensors to enhance MVPA performance. We also tested how many magnetometers were required for multivariate pattern analysis. We found that classification accuracy did not improve when going beyond 25 sensors. In conclusion, when designing an MEG system based on SQUID or OPM magnetometers which is optimized for multivariate analysis, about 25 magnetometers are sufficient possibly augmented with reference sensors for noise reduction.HighlightsUsing signal space separation on magnetometer data prior to multivariate pattern analysis might reduce classification accuracy as it reduces the rank of the data resulting in loss of information.When performing multivariate data analysis, other noise reduction approaches that reduce the rank of the data less are advisable such as signal space projection or 3rd-order gradient noise reduction.A sensor array of about 25 magnetometers is sufficient for multivariate pattern analysis.