Predictability of Sleep Slow Oscillation Emergence and Spatial Extent from Pre-Onset Neural Dynamics

Abstract

Slow oscillations (SOs; [~]0.5-1.5 Hz) are a hallmark of non-rapid eye movement (NREM) sleep and are known to support memory consolidation and large-scale cortical communication. Although their instantaneous dynamics are well characterized, the neural processes that precede SO initiation--and whether they predict the spatial extent of the upcoming event--remain largely unknown, despite their potential utility for anticipatory closed-loop intervention. Using high-density EEG from 29 healthy adults, we examined neural activity in the 2-s interval preceding the SO trough. Analyses focused on distinct SO subtypes defined by their spatiotemporal properties: Global, Frontal, and Local. We specifically focused on Global and Frontal events, both of which originate frontally but differ in their propagation. We quantified instantaneous spectral power, time-frequency dynamics, phase-amplitude coupling (PAC), and amplitude-amplitude coupling. Across all analyses, theta-band power (4-8 Hz) emerged as the earliest and most robust predictor of SO initiation, remaining informative after restricting analyses to temporally isolated SOs to reduce potential residual-aftereffect confounds from preceding SOs. Theta power exhibited a sustained rise beginning nearly two seconds before the SO trough ([~]1.6 s before onset). Theta power increases were absent in surrogate epochs and reliably differentiated Global from Frontal SOs with moderate-to-strong effect sizes (Cohens d = 0.45-0.77), demonstrating the robustness of theta power as a physiological signature. Mechanistically, Global SOs were preceded by enhanced delta-theta PAC and broad low-frequency synchronization, whereas Frontal SOs were preceded by elevated theta/alpha-to-beta/low-gamma coupling, reflecting a state of locally enhanced coupling at frequencies higher than the SO range, which appears to restrict propagation. A simple logistic regression classifier using only pre-onset theta power achieved >95% accuracy in distinguishing SOs from surrogate events and differentiated Global from Frontal SOs with [~]83% multiclass accuracy, showing further sensitivity improvements when delta power was included. These findings demonstrate that isolated SOs are preceded by structured network dynamics that are tied to their spatial extent. Thus, theta activity can be used to predict SO occurrence and might be leveraged in closed-loop neuromodulation.