Sleep Architecture and Memory Consolidation in Healthy Adults

Abstract: Sleep does not store the day’s experiences unchanged. It actively reshapes memory traces through a sequence of neurophysiological events whose between-individual variability is poorly characterized at the architectural level. This paper asks whether the within-night slow-wave sleep (SWS) to rapid eye movement (REM) temporal gradient is a reliable individual-level phenomenon or a statistical average that obscures substantial heterogeneity, and tests whether the oscillatory spectral signatures of waking affective valence processing persist into the sleep state theorized to consolidate emotional memories. A five-class EEG sleep stage classifier (SVM with RBF kernel, leave-one-subject-out cross-validation, macro-F1 = 0.593, Cohen’s κ = 0.486) transforms polysomnographic recordings from 69 healthy adults in the Sleep-EDF Expanded Cassette sub-study into per-epoch stage sequences. Three principal results emerge. First, first-half SWS proportion strongly predicts both the magnitude of SWS decline across the night (r= 0.770, p= 1.03 × 10−14) and subsequent REM expansion (r= 0.351, p= 0.003), demonstrating that the SWS-REM gradient is individually structured. Second, the spectral geometry of REM stage discrimination is inverted relative to the waking affective valence signature established by Khare (2026). Theta oscillations dominate REM discrimination while beta and gamma dominate waking valence classification, providing cross-domain evidence that sleep consolidation does not preserve the waking spectral architecture of affective encoding, supporting representational reorganization over reactivation-fidelity accounts. Third, eleven subjects (15.9%) show non-declining SWS with near-absent N3 throughout the recording window. Their mean first-half SWS proportion (0.063) falls far below the declining-subgroup mean (0.321), providing a clean constraint on the synaptic homeostasis hypothesis. The hypothesis requires meaningful early SWS accumulation to explain a flat trajectory, but these subjects show near-zero SWS from the outset. The directional correlation between first-half SWS and subsequent REM expansion (r= 0.351, p= 0.003) is separately consistent with two-stage model predictions.