Waveform detection by deep learning reveals multi-area spindles that are selectively modulated by memory load

Abstract

Sleep is generally considered to be a state of large-scale synchrony across thalamus and neocortex; however, recent work has challenged this idea by reporting isolated sleep rhythms such as slow-oscillations and spindles. What is the spatial scale of sleep rhythms? To answer this question, we adapted deep learning algorithms initially developed for detecting earthquakes and gravitational waves in high-noise settings for analysis of neural recordings in sleep. We then studied sleep spindles in non-human primate ECoG, human EEG, and clinical intracranial recordings (iEEG) in the human. We find a widespread extent of spindles, which has direct implications for the spatiotemporal dynamics we have previously studied in spindle oscillations (Muller et al., 2016) and the distribution of memory engrams in the primate.