Replay-triggered brain-wide activation in humans

The study investigates how transient memory reactivation and its sequential replay engage brain-wide networks in humans, focusing on when replay occurs (temporal dynamics) and where it is expressed (spatial distribution). Replay—fast, sequential reactivation of past experiences—has been implicated in functions such as memory consolidation, planning, and flexible cognition. Yet, due to limitations of M/EEG (high temporal, low spatial resolution) and fMRI (high spatial, low temporal resolution), it remains unclear how replay relates to large-scale networks such as the default mode network (DMN), both during task (on-task mental simulation) and during wakeful rest (offline). The authors hypothesize that simultaneous EEG-fMRI can capture fast replay events via EEG and localize replay-triggered activations and connectivity via fMRI, revealing hippocampal and DMN engagement during on-task replay, and learning-induced reactivation and hippocampal–entorhinal interactions during rest.

Replay was first identified in the rodent hippocampus during sleep and later during wakefulness, supporting consolidation, planning, and flexible behavior. Noninvasive human M/EEG studies have detected rapid sequential reactivation (e.g., MEG sequences at ~40 ms lags), including reverse replay linked to value learning. fMRI studies localized sequential activations to regions such as visual cortex and hippocampus (e.g., Wittkuhn & Schuck; Schuck & Niv), but fMRI alone has limited sensitivity to direction and speed of replay. MEG source localization and animal work implicate the hippocampus and broader cortical networks, including the DMN. The interplay between replay dynamics and DMN activity, and differences between on-task and rest states, remain underexplored due to methodological constraints. This study situates itself at the intersection of these literatures by using simultaneous EEG-fMRI to bridge temporal and spatial gaps.

Participants: 40 healthy adults were recruited; after exclusions for head motion or incomplete participation, 33 (mean age ~22.9 years; 17 females) were included. The study was approved by the ethics committee; all provided informed consent.

Design and tasks: The experiment comprised resting-state scans (PRE and POST learning), a functional localizer, sequence learning, and cued mental simulation, all during simultaneous EEG-fMRI. The functional localizer presented four images (face, scissor, zebra, banana) followed by a semantic word; subjects judged congruence (accuracy ~94.6%). Sequence learning involved learning three pairwise associations (AB, BC, CD) to form a 4-item sequence A→B→C→D, followed by probe tests (final run accuracy criterion ≥90%, achieved by all; overall probe accuracy ~93.9%). In cued mental simulation, a cue (“1” or “4”) instructed forward or reverse simulation for 10 s, followed by a probe recognition (accuracy ~93.6%; vividness ratings high, mean ~3.35/4). Two 5-min resting-state scans were acquired, one before (PRE) and one after (POST) learning.

EEG-fMRI acquisition: EEG (64 channels, 1000 Hz) was recorded with MR-compatible equipment; preprocessing included MR artifact removal (AAS), filtering (1–40 Hz), downsampling to 100 Hz, ICA-based artifact rejection, and epoching around stimuli and cues; rest EEG underwent analogous preprocessing. fMRI was collected on a 3T Siemens Prisma (TR=1300 ms, TE=24 ms, 3 mm isotropic voxels, MB factor 2), with preprocessing via fMRIPrep (motion correction, slice-timing correction, normalization to MNI space, nuisance regression including motion and CompCor components). High-resolution T1-weighted anatomical images supported registration.

Decoding and replay detection: EEG decoding used lasso-regularized logistic regression (one-vs-rest classifiers) trained on functional localizer data using whole-head features, with time-resolved training/testing from −200 to 800 ms. Peak decoding occurred at 210 ms post-stimulus and classifiers at this latency were used to decode reactivations during mental simulation and rest. fMRI multivariate pattern analysis (Nilearn/scikit-learn) used feature selection combining anatomical masks (visual and MTL regions including hippocampus and entorhinal cortex) and functional t-maps; one-vs-rest logistic regression was cross-validated, then trained on full localizer data to decode reactivation probabilities during mental simulation (8 volumes/trial) and rest (230 volumes/session). Replay was quantified via Temporally Delayed Linear Modelling (TDLM) applied to decoded [time × state] probability matrices to estimate sequenceness across lags for forward and backward hypothesized transitions.

Event timing and fMRI modelling: EEG-identified replay probability time courses were derived by detecting two-state transitions at the peak lag (30 ms) and convolved with the HRF to form regressors in fMRI GLMs, enabling localization of replay-triggered activations (GLM 4) and psychophysiological interaction (PPI) analyses with hippocampal and mPFC seeds. For rest, single-item task reactivation (summed across stimuli) was similarly convolved and entered as a regressor (GLM 5); an analogous fMRI-based reactivation regressor (without additional HRF convolution) was tested (GLM 6). ROI analyses focused on anatomically defined hippocampus (and control M1) and entorhinal cortex (for PPI). Whole-brain analyses used cluster-wise FWE correction (cluster-forming threshold P_unc < 0.001 unless noted). Statistical testing used permutation methods for TDLM and appropriate t-tests or mixed models elsewhere, with multiple-comparisons corrections as specified.

• Decoding performance:
  • EEG classifiers peaked at 210 ms post-stimulus with 46.25 ± 0.95% accuracy (chance 25%; t(32)=22.41, P<0.001). Classifier probabilities were selective for their trained images.
  • fMRI classifiers achieved 83.39 ± 1.77% accuracy at the 4th TR post-onset (chance 25%; t(32)=38.00, P<0.001). Across subjects, EEG and fMRI decoding accuracies correlated (r=0.49, P=0.004).
• On-task replay during mental simulation (EEG TDLM):
  • Significant forward replay at short lags in both cue conditions: cue “1” (forward) trials showed forward replay at 30–50 ms (peak 30 ms; β≈0.021), and cue “4” (backward) trials at 20–40 ms (peak 30 ms; β≈0.023).
  • At the 30 ms peak, forward sequenceness was significant in both forward and backward cue conditions (paired t-tests reported: forward condition t(32)=2.80, P=0.009; backward condition t(32)=3.09, P=0.004). Replay strength increased with task experience (F(32)=4.18, P<0.001) and was unrelated to vividness ratings.
  • No significant replay was detected using fMRI-alone methods (TDLM or regression slope) during mental simulation (all Pcorr ≥ 0.06).
• Replay-triggered brain-wide activations and connectivity (fMRI):
  • EEG-derived replay probability predicted increased BOLD activation in hippocampus and medial prefrontal cortex (mPFC) during mental simulation (whole-brain cluster-level FWE P<0.05).
  • Hippocampal-seed PPI showed increased connectivity with DMN regions (mPFC, PCC) and visual cortex as replay probability increased.
  • mPFC-seed PPI revealed increased connectivity with DMN regions and visual cortex, but not with hippocampus.
• Rest reactivation after learning:
  • Mean EEG-based task reactivation was higher POST vs PRE Rest (t(31)=2.75, P=0.010).
  • EEG-based reactivation modulated hippocampal activity during POST Rest (ROI t(32)=3.83, Pcorr<0.001) but not PRE Rest (t(32)=1.08, Pcorr=0.287); POST>PRE difference significant (paired t(32)=2.44, Pcorr=0.030). No effects were observed in control M1.
  • Aligning hippocampal BOLD to EEG reactivation onsets showed a peak at the 2nd TR (t(32)=3.02, P=0.005) during POST Rest.
  • Hippocampal PPI during POST Rest revealed increased connectivity with entorhinal cortex (EEG-based reactivation: t(32)=2.75, P=0.010); not significant during PRE Rest.
  • No robust evidence for sequential replay during rest (EEG or fMRI-based metrics) in PRE or POST sessions.
  • During POST Rest, EEG- and fMRI-based reactivation measures positively correlated in their ability to explain hippocampal activity (r=0.38, P=0.029), but not during PRE Rest.
• Behavior: High accuracy across tasks (sequence learning probes ~93.9% overall; cued simulation ~93.6%; final learning run ~99.5%), with no forward vs backward performance differences and high vividness ratings; no correlations between rest reactivation strength and behavioral measures, likely due to ceiling effects.

By integrating EEG and fMRI, the study delineates when and where human memory replay and reactivation engage the brain. Fast forward replay (~30 ms lags) occurs during cued mental simulation but is independent of the instructed direction, suggesting a spontaneous, instruction-invariant process. These replay events are accompanied by increased activation in hippocampus and mPFC, and enhanced hippocampal functional connectivity with DMN nodes and visual cortex, consistent with hippocampal-initiated interactions with a cognitive map maintained in the DMN. During rest, learning induced stronger task-related reactivation without clear evidence of sequential replay. Post-learning reactivation selectively engaged hippocampus and increased hippocampal–entorhinal connectivity, aligning with consolidation-related mechanisms and suggesting distinct network dynamics for on-task replay (hippocampus–DMN–visual) versus offline reactivation (hippocampus–entorhinal). The absence of rest replay may reflect simplified task structure, ceiling-level learning, and limitations of simultaneous EEG-fMRI sensitivity for dispersed, burst-like replay. The findings support a cross-modal framework where EEG timestamps transient cognitive events enabling fMRI to map brain-wide activations and connectivity. While EEG- and fMRI-based reactivation time series are not temporally aligned, their convergent prediction of hippocampal activity during POST Rest indicates complementary sensitivity to reactivation processes. Overall, the results refine understanding of how replay interacts with large-scale networks to support cognition, and highlight differences between on-task and offline states.

This work establishes a simultaneous EEG-fMRI pipeline that captures transient replay events with high temporal precision and maps their brain-wide correlates with high spatial resolution. Key contributions include demonstrating instruction-independent, fast forward replay during mental simulation, localizing replay-triggered activation to hippocampus and mPFC, and showing replay-strength-dependent increases in hippocampal connectivity with DMN and visual cortex. After learning, task-related reactivation strengthens, selectively engages hippocampus, and enhances hippocampal–entorhinal connectivity during rest, suggesting a shift toward consolidation-related interactions. Future directions include applying this framework to sleep to probe consolidation, examining hippocampal replay alongside grid-like coding during cognitive map computations, testing causal relationships and directionality with more advanced analyses or invasive recordings, and exploring how task complexity and learning strength modulate replay versus reactivation across brain networks.

• No significant evidence of sequential replay was detected during rest using either EEG or fMRI approaches; this may reflect reduced EEG signal quality in the scanner, temporally dispersed replay during rest, or the simplicity of the single-sequence design with ceiling-level learning reducing replay demands.
• fMRI-only replay metrics (TDLM or regression-based) did not yield significant replay during mental simulation, highlighting limited sensitivity of fMRI to direction/speed of replay in this paradigm.
• PPI analyses reveal condition-dependent connectivity changes but cannot infer causal directionality.
• Behavioral measures showed ceiling effects, limiting correlations with neural reactivation/replay.
• Human replay/rehactivation is defined at the representational level and may not map directly onto neuronal-level replay observed in animals, constraining direct cross-species inferences.