MVPA Analysis of Intertrial Phase Coherence of Neuromagnetic Responses to Words Reliably Classifies Multiple Levels of Language Processing in the Brain

Top, Heatmap of significant clusters across three linguistic contrasts, five frequency bands, and time. Lexical condition in blue colors, semantic in green, and syntax in red. Bottom, Surface topography of significant effects. For all conditions, colors go from lighter to darker as latency becomes longer.

A, Model patterns: interpreting coefficients of a machine learning model is not trivial and a high coefficient value does not necessitate a high signal value in the MEG data (for details, see Haufe et al., 2014). “Model patterns” are a way to highlight the signal in a neurophysiological sensible way that is directly interpretable compared to the raw coefficients (Haufe et al., 2014). We show top and bottom 5% of the patterns in the γ-low band from 222 to 238 ms. Blue colors are areas of activation able to predict real words and yellow/red are areas used to predict pseudo word. B, Average top and bottom 5% of ITPC difference; blue colors indicate higher ITPC for real words and yellow/red colors indicate higher ITPC for pseudo word γ-low band from 222 to 238 ms. C, Average ITPC over time; solid lines are the average of the selected features, dashed lines are the average of all vertices in the source space. Time 0 is the divergence point, when stimuli could be recognized from the available acoustic information.

Heatmap of ROC-AUC scores for all bands in lexical condition. Note that chance in this condition is 75%. Time is relative to DP.

A, Model patterns (see also Fig. 4 legend): top and bottom 5% of the patterns in the the α band from 80 to 122 ms. Blue colors are areas used to predict action verb and yellow/red are areas used to predict object noun. B, Average top and bottom 5% of ITPC difference, blue colors indicating higher ITPC for action verb and yellow/red indicating higher ITPC for object noun from 80 to 122 ms. C, Average ITPC over time, solid lines are the average of the selected features, dashed lines are the average of all vertices in the source space. Time 0 is the divergence point, when stimuli could be recognized from the available acoustic information.

Heatmap of ROC-AUC scores for all bands in semantic condition.

A, Model patterns: top and bottom 5% of the patterns in the γ-low band from 84 to 98 ms. Blue colors are areas used to predict correct syntax and yellow/red are areas used to predict incorrect syntax. B, Average top and bottom 5% of ITPC difference. Blue colors indicate higher ITPC for correct syntax and yellow/red colors indicate higher ITPC γ-low band from 84 to 98 ms. C, Average ITPC over time; solid lines are the average of the selected features, dashed lines are the average of all vertices in the source space. Time is relative to the divergence point.