Implicit statistical learning in sleep: an overnight MEG/EEG study using a grammar learning auditory task

Introduction:
The ability to implicitly extract regularities and associations embedded in sensory inputs is fundamental for adapting behavior to environmental demands, and it is considered automatic: one that does not require a deliberate allocation of attentional resources. As attentional resources are naturally minimized in sleep, recent studies have sought to determine the extent to which the brain engages in the predictive processing of sensory inputs during sleep. Some studies report disruption of hierarchical predictive coding in sleep, whereas others argue for limited but preserved detection of violation of predictions, with profound implications for studies showing effective learning during sleep. Here, we look for predictive signals in the human brain during overnight sleep using an implicit grammar learning paradigm.
Method:
Participants (N=18) implicitly learned grammatical rules underlying a sequence of syllables and were subsequently re-exposed to the same syllables 2 days later before going to bed (passive listening) as well as during overnight sleep, while simultaneous Electroencephalography (EEG) and Magnetoencephalography (MEG) were recorded. We conducted Event-Related Fields, Time-Frequency, as well as multivariate pattern analyses during stimulus presentation in wakefulness and sleep comparing random to predictable syllables. We determined statistical differences using cluster-based permutations.
Results:
Although the event-related responses between amplitudes of the random and predictable syllables could not be differentiated, predictable syllables were characterized by higher peristimulus beta activity (15-30Hz) during learning, lower peristimulus beta activity (15-30Hz) during subsequent passive listening, as well as during light sleep. We furthermore show that we can decode which stimulus was played in all sleep stages; however, we found no evidence of predictions via preactivating the features of an expected stimulus using multivariate pattern analysis.
Conclusion:
Overall, in line with previous studies, our results suggest that although the processing of low-level stimulus properties persists in sleep, the brain is, however, restricted in engaging in the active prediction of sensory inputs.