STATE-AWARE EEG COHERENCE FEATURES FOR VIRTUAL INTERVENTION DURING SLEEP

Personalized non-invasive brain stimulation increasingly relies on model-based planning where EEG provides cost-effective read-outs of brain dynamics, enabling computational optimization of interventions before application¹. Sleep EEG's stereotyped architecture offers ideal dynamics for tuning models that produce meaningful brain states and transitions. However, which EEG features represent physiologically meaningful states relevant for predicting and modulating brain dynamics remains unclear².

We investigated whether complex coherence and cross-frequency coherence define structured regimes of sleep EEG dynamics that encode this stereotyped architecture and can be reproduced in neural mass models (NMMs). We analyzed full-night high-density EEG from 29 healthy adults in the ANPHY-Sleep dataset, constructing personalized feature bases combining spatially resolved complex and cross-frequency coherence³. Feature trajectories were examined across sleep stages using clustering and dimensionality reduction, aligned to state transitions to probe pre-transition dynamics².

Complex coherence differentiated canonical sleep stages while preserving spatial and phase information essential for reproducing archetypal brain states and predicting transitions in NMMs⁴. Cross-frequency coherence revealed NREM sub-states with distinct slow oscillation-spindle coupling patterns, markers of sleep-dependent plasticity and cognitive health⁵⁻⁶. Trajectories showed systematic drifts preceding transitions, with preliminary analysis revealing bifurcation-like behavior before REM onset.

We assessed model "expressivity" (the range of characteristics the model can generate) by simulating cortical NMMs in BraiNN, our high-performance framework for whole-brain simulations⁷. We evaluated whether noise-driven mean-field dynamics reproduce empirically observed features and stage-specific spectral regimes for N2/N3 sleep⁸⁻⁹. Results suggest coherence-based EEG features provide rich information for capturing brain states and health-relevant micro-architectural markers for model-based neuromodulation.