Lead. A new study using Riemannian geometry-based machine learning on high-density EEG recordings suggests that declarative odor cueing during sleep produces more structured spindle-band covariance patterns than non-declarative odor cueing, particularly over central brain regions. However, the effects were modest and did not survive correction for multiple comparisons, tempering the strength of the conclusions.
Published June 18 in Frontiers in Neuroscience, the study by Jesyin Lai, Pankaj Pandey (co-first authors), David M. Baum, Jens G. Klinzing, Andrea Sanchez-Corzo, and Ranganatha Sitaram adds to a growing body of work using Targeted Memory Reactivation (TMR) to probe how the sleeping brain consolidates different types of memory.
What they found. The researchers analyzed high-density EEG recordings collected during NREM sleep from participants in a TMR paradigm. During learning before sleep, participants were exposed to two distinct odors: Odor D, associated with a declarative memory task (object-location associations), and Odor M, associated with a non-declarative motor sequence task. During subsequent sleep, the same odors were re-presented along with a vehicle control odor, and the team examined EEG epochs in two spindle frequency bands: fast spindles (12.5-16 Hz) and slow spindles (9-12.5 Hz).
Using within-participant machine learning classifiers based on Riemannian geometry, the team attempted to decode whether a given EEG epoch was elicited by Odor D versus vehicle, and by Odor M versus vehicle. This approach treats the covariance structure of multichannel EEG signals as points on a Riemannian manifold, capturing distributed patterns of neural activity that conventional univariate analyses might miss.
Decoding performance, assessed relative to permutation-derived chance levels, showed condition-dependent variation across frequency bands, time windows (0-2, 0-4, and 0-7 seconds post-cue), and channel subsets (all channels, frontal, central, and posterior). Across these analyses, decoding accuracy tended to be higher for the declarative condition (Odor D) than for the non-declarative condition (Odor M), with the strongest effects observed when using central channel data.
A channel-level contribution analysis further revealed that the covariance patterns driving classification in the declarative condition were spatially structured over central regions, suggesting coordinated spindle-band activity consistent with memory-related neural modulation. In contrast, contributions during the non-declarative condition were more diffuse and less consistent across participants.
Why it matters. Sleep spindles have long been implicated in memory consolidation, with evidence linking both fast and slow spindle oscillations to the reactivation and stabilization of newly acquired memories. TMR studies have shown that re-presenting learning-associated cues during sleep can boost memory performance, but the precise neural signatures differentiating declarative from non-declarative memory reactivation have remained elusive.
The present study’s use of covariance-based decoding, an approach derived from Riemannian geometry, represents a methodological advance. Rather than examining power changes at individual electrodes, this technique captures the coordinated structure of spindle-band activity across the electrode array, potentially offering a more sensitive window into distributed neural processes.
The finding that central-channel covariance patterns were more organized during declarative odor cueing aligns with the known involvement of centroparietal regions in sleep spindle generation and declarative memory processing. It suggests that different memory systems may engage spindle-band activity in qualitatively distinct ways, even within the same sleep session.
Limits. The most important caveat is that the reported effects, while descriptively consistent, were modest and did not survive correction for multiple comparisons across the many analyses performed. The authors are transparent about this limitation, noting that the results should be interpreted as preliminary and warrant further validation in larger samples.
Several other factors also constrain interpretation. The study relied on a single session of odor re-exposure, leaving open questions about dose-response relationships and consolidation time courses. The sample size, while typical for intensive EEG TMR studies, may have limited statistical power to detect small-to-moderate effects reliably. Additionally, the non-declarative motor task (Odor M) and declarative task (Odor D) differed in content, sensory associations, and learning contexts, making direct comparisons between conditions inherently multidimensional.
The Riemannian geometry approach, while promising, is relatively novel in sleep EEG research, and its sensitivity and specificity for detecting memory-related neural signatures require independent replication. The channel contribution maps provide descriptive evidence of spatially structured patterns, but formal statistical testing of these spatial maps was not reported.
Bottom line. This study provides preliminary evidence that covariance-based analyses of spindle-band EEG can detect differential neural responses during declarative versus non-declarative odor cueing in sleep, with declarative cueing associated with more structured central-channel patterns. The work demonstrates the potential of Riemannian geometry approaches for probing distributed sleep EEG dynamics. However, because the effects were modest and did not survive multiple comparison correction, these findings should be considered hypothesis-generating rather than confirmatory. Larger replication studies will be needed to determine whether these covariance patterns reliably index memory type during sleep.
Source. Lai J, Pandey P, Baum DM, Klinzing JG, Sanchez-Corzo A, Sitaram R. Covariance-based analysis of spindle-band EEG during declarative and non-declarative odor cueing in sleep. Front Neurosci. 2026;20:1810323. doi:10.3389/fnins.2026.1810323. PMID: 42395320.