Decoding auditory working memory content from intracranial high frequency activity in humans

How the human brain maintains information in working memory (WM), a process critical for all our goal-directed function, has been debated for decades. Classic neurophysiological models, which argue that WM is maintained via persistent content-specific “delay activity,” have been challenged by alternative ideas suggesting a combination of dynamic activity patterns and activity-silent mechanisms. Here, utilizing human intracranial stereo-EEG (sEEG) recordings and machine learning techniques, we tested understudied auditory WM in multiple cortical and subcortical brain areas. Neuronal activity was quantified as broadband high frequency activity (HFA, 70-190 Hz) which has been shown to be highly correlated with multiunit activity of neuron populations. Our multivariate pattern analysis (MVPA) results, validated via robust non-parametric permutation testing, show that information can be decoded from multiple brain regions, including prefrontal regions, superior temporal auditory cortices, and the hippocampus. However, the recording sites with high WM decoding accuracies were not accompanied by statistically significant increases in HFA power. In contrast, HFA power was reduced relative to the period preceding WM encoding in many frontal, superior temporal, and hippocampal sEEG recording sites. These results are in line with the hypothesis that WM maintenance can be supported by highly dynamic, “activity silent” processes rather than via persistent activity only.