Delayed, Reduced, and Redundant: Information Processing of Prediction Errors during Human Sleep.

Auditory prediction error processing continued throughout all sleep stages including N1, N2, N3, and REM sleep.

• Twenty-nine participants (15 women) underwent an auditory 'local-global' oddball paradigm during wakefulness and an 8-hour sleep opportunity monitored via polysomnography.
• The paradigm focused on 'local' mismatch responses to a deviating fifth tone after four standard tones.
• Event-related potential (ERP) analyses confirmed prediction error processing persisted across all sleep stages.
• ERP amplitudes actually increased with deeper NREM sleep, despite reduced information content.

Mutual information analyses revealed a substantial reduction in encoded prediction error information particularly during N3 and REM sleep.

• Despite increased ERP amplitudes during deeper NREM sleep, the amount of information encoded about prediction errors was substantially reduced.
• This dissociation between ERP amplitude and mutual information content highlights that larger neural responses do not necessarily reflect richer information encoding.
• Mutual information (MI) analyses were used to quantify the amount of prediction error information encoded in neural signals across sleep stages.
• The reduction in information content was most pronounced during N3 (slow-wave sleep) and REM sleep.

Information encoding of prediction errors was delayed during sleep compared to wakefulness.

• Mutual information analyses revealed not only reduced information content but also a temporal delay in when prediction error information was encoded during sleep.
• This delayed processing was observed across sleep stages.
• The finding is consistent with a transition to a 'sentinel processing mode' in which the brain continues to process stimuli but with altered temporal dynamics.
• The delay reflects how neural information evolves with variations in consciousness across the night.

Co-information analyses showed that neural dynamics became increasingly redundant with increasing sleep depth.

• Co-information (co-I) analyses were used alongside mutual information to characterize the structure of neural information encoding.
• Increasing sleep depth was associated with increasingly redundant neural information processing.
• Redundancy here means that different neural signals or time points carry overlapping rather than complementary information about prediction errors.
• This contrasts with the richer, less redundant information encoding observed during wakefulness.

Temporal generalization analyses revealed a largely shared neural code between N2 and N3, but a distinct neural code between wakefulness and sleep.

• Temporal generalization analyses (TGA) were used to assess whether neural representational geometries were shared across sleep stages.
• N2 and N3 shared a largely similar neural code for prediction error processing.
• The neural code during wakefulness differed substantially from that observed during sleep stages.
• This suggests that the transition from wakefulness to sleep involves a qualitative shift in the neural representation of prediction errors, not merely a quantitative reduction.

The study characterized how the 'sentinel processing mode' systematically changes from wakefulness through light sleep, deep sleep, and REM as consciousness wanes.

• Advanced information-theoretic analyses including mutual information and co-information were combined with ERPs and temporal generalization analyses.
• The sentinel processing mode is described as enabling continued processing of environmental stimuli despite the absence of consciousness.
• The overall pattern was: delayed processing, more redundant neural information, and less rich (lower mutual information) neural encoding with increasing sleep depth.
• The authors hypothesized a shared neural code across sleep stages with deeper sleep associated with reduced information content and increased redundancy, which was confirmed.