A multi-feature approach to the classification analysis of auditory event-related fields acquired with MEG

In auditory magnetoencephalography (MEG) studies, it is common practice to search for differences between event-related fields (ERFs) measured in different experimental conditions and between the two hemispheres of the human brain. This differentiation is usually based on single ERF parameters such as the most prominent peak of the auditory waveform, the M100.

Here, we introduce a multi-feature approach to differentiate such ERFs, where a multitude of parameters extracted from the ERFs served as input to machine-learning (ML) models.

We applied this approach to illustrative MEG data taken from a study in which the up-down categorisation of frequency-modulated (FM) sweeps was addressed in search for differences between the MEG signals of the left and the right hemisphere as well as between the two experimental conditions (passive listening vs active categorisation). We obtained results complementary to those of an earlier study using the same experimental paradigm, within the confines rendered by the new multivariate classification methodology and a limited sample size.

We argue that complex characteristics of evoked brain responses speak for multi-feature approaches with new measures of signal quantification, beyond those commonly used. However, we critically acknowledge the limitations of ML strategies in applications to MEG datasets of limited size.

We conclude that a sample size typical for neuroscience experimentation limits the usability of ML classification. We also point out some disadvantages of bootstrapping employed to generate surrogate data with the aim to increase dataset size.