Interpreting human sleep activity through neural contrastive learning

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Spontaneous memory replay during sleep is vital for cognition.
• Existing electroencephalogram (EEG) decoders struggle with sleep rhythm variations, hindering memory replay analysis.
• Generalizing wake-trained decoders to sleep stages remains a significant challenge.

Purpose of the Study:

• To develop a robust method for decoding memory replay during sleep, overcoming limitations of current EEG decoders.
• To isolate shared semantic content from background sleep rhythms using neural contrastive learning.
• To create a generalized decoder applicable across different sleep stages and participants.

Main Methods:

• Developed the Sleep Interpreter (SI) using neural contrastive learning.
• Collected a large dataset (135 participants, ~1000h sleep, ~400h wake EEG) with targeted memory reactivation.
• Validated SI on independent datasets and real-time sleep staging.

Main Results:

• SI achieved high decoding accuracy (40.02% top-1) for semantic responses during NREM sleep, particularly during slow oscillation and spindle coupling.
• Decoded reactivations using SI correlated with post-sleep memory performance in independent experiments.
• Demonstrated SI's generalizability across participants and sleep conditions.

Conclusions:

• The Sleep Interpreter (SI) effectively decodes memory replay during sleep, even with varying sleep rhythms.
• SI shows promise for advancing memory research and has potential for clinical applications in sleep disorders.
• The shared dataset and codebase will facilitate future research in sleep and memory.