Time-dependent neural arbitration between cue associative and episodic fear memories

Traumatic events can differentially affect human memory systems: associative learning linking neutral cues to aversive outcomes is strengthened, while episodic components, including temporal order, may become incoherent in some individuals with PTSD. Despite parallel literatures on Pavlovian threat conditioning and episodic memory, a unified account of how the brain prioritises cue associations over episodic coding after trauma is lacking. The authors hypothesised that immediately after a threatening experience, the brain would rely on generalisable cue associations for rapid defense, then transition over time to selective episodic sequence coding. They developed an episodic threat conditioning paradigm where a simulated car crash (US) followed sequences of environmental sounds, enabling concurrent formation of cue-association and episodic sequence memories. They assessed physiological fear expression during acquisition, immediately after, and 24 h later, and used multivariate fMRI to test whether hippocampus (HPC) and dorsolateral prefrontal cortex (DLPFC) differentially transmit sequence representations to the ventromedial prefrontal cortex (VMPFC)–Amygdala fear circuit across time.

Prior work implicates HPC and DLPFC in sequence learning and temporal-order memory, with both projecting to VMPFC, a hub that regulates amygdala-driven fear expression. Although the amygdala’s role in human fear conditioning is debated, large-scale fMRI supports its contribution. PTSD involves exaggerated cue-based fear and impairments in episodic coherence/sequence coding. Acute stress can initially impair access to episodic details, suggesting a temporal shift from cue-based to episodic representations. The authors build on these literatures to propose that HPC and DLPFC regulate transmission of episodic sequences to VMPFC–Amygdala, arbitrating between associative and episodic fear memories over time.

Participants: 44 healthy adults (29 males, mean age 22.0±1.6 years) were recruited; 2 missed Day 2 and 1 had incompatible MRI parameters. Final N=41 for fMRI, N=42 for SCR. Ethics approved; informed consent obtained. Sample size planned via G*Power using pilot effect size (dz=0.57) targeting 95% power for SCR difference on Day 2 (target N≈38; recruited 44 to allow counterbalancing and dropouts). Design and task: A naturalistic audiovisual paradigm simulated a driving scene with three auditory cues (bicycle bell, traffic light melody, crow call) arranged in triplets forming three sequences: CS+sequence (a-b-c) paired with US (car crash) on 37.5% of its presentations during Acquisition; CS+element (b-a-c) never paired with US but sharing the last cue ‘c’ with CS+sequence; CS− (a-c-b) never paired and with unique last cue ‘b’. Participants underwent Acquisition and Immediate Test on Day 1 (no US in Immediate Test), and Long-term Test on Day 2. Day 2 included reinstatement (three unsignaled US-only clips) and a 10-min rest scan before testing. Trial timing: three sounds separated by ISIs (~2.5 s), truck appearance at 18±0.5 s; in US trials, crash at 1.36 s post-truck onset with 2 s crash sound; no-US trials showed truck passing. Behavioral measures: Skin conductance response (SCR) recorded via BIOPAC on the left hand. SCR phasic responses quantified 0.9–4 s after truck onset, log-transformed log(SCR+1), scaled to session’s max no-US SCR. Analyses focused on no-US trials to index anticipatory responses. ASCR computed as SCR(CS+)−SCR(CS−) per condition. fMRI acquisition: Siemens Prisma 3T, 20-channel head coil. EPI: TR=2000 ms, TE=27 ms, flip angle 70°, FOV 200 mm, 75 slices, 2×2×2 mm, AP phase-encode, multiband 3. T1 MPRAGE: TR=2300 ms, TE=2.98 ms, flip 9°, FOV 256 mm, 1 mm isotropic. Preprocessing: SPM12; motion correction, alignment to T1, no smoothing; BOLD shifted by 6 s for hemodynamic delay; detrended and z-scored per voxel per block. Epoch-averaged voxel patterns computed for sound-cues epoch (~12 s) and US anticipatory epoch (including ITI; ~6 s). ROIs: HPC and Amygdala from AAL; DLPFC defined as BA9/46; VMPFC spherical ROI (15 mm radius) at MNI [0, 40, −12]. MVPA decoding: Sparse logistic regression (SLR) classifiers trained per participant on sound-cues epoch to discriminate CS+sequence vs CS+element; validated on left-out sound-cues data and tested on US anticipatory epoch. 20-fold CV; iterative feature selection (10 iterations). Additional binary and ternary decoders (e.g., vs CS−) used in controls. Reported average selected voxels: HPC (sound-cues) 114±11/81±11/82±11; HPC (anticipatory) 57±24/35±17/36±20 across sessions; DLPFC analogous numbers (sound-cues) 111±12/83±11/86±15 and (anticipatory) 58±28/38±18/40±22. Information transmission analysis: For each trial, the decoder trained on sound-cues estimated CS+ likelihood during the US anticipatory epoch in a seed (HPC or DLPFC). Activity patterns in target (combined VMPFC–Amygdala, or VMPFC/Amygdala separately) were used to linearly predict seed likelihoods via leave-one-trial-out cross-validation. Prediction accuracy defined as correlation between predicted and actual likelihoods (Fisher z-transformed). This assesses communication of sequence information from seed to target. Control analyses examined other target/seed pairs and sensory/fear-related ROIs. Statistics: Wilcoxon signed-rank tests for paired or chance-level comparisons, with FDR correction within session unless between-session contrasts. Linear mixed-effects (LME) models with random intercepts and slopes per subject; categorical factors for session, CStype, ROI, epoch; decoding accuracy included as random effect in transmission models. Trait anxiety assessed via STAI; correlations tested with Pearson and robust regression.

• Behavioral SCR shows time-dependent transformation from cue-based to sequence-based fear expression: • Acquisition (Day 1): ASCR averaged across CS+sequence and CS+element > 0 (z=2.35, P=0.009, r=0.36); no difference between CS+sequence and CS+element (z=−0.52, PFDR=0.60). • Immediate Test (Day 1): CS+sequence ≈ CS+element (z=−0.34, PFDR=0.78); overall ASCR vs CS− not significant (z=−0.65, P=0.74), though early trials show conditioning. • Long-term Test (Day 2): Selective sequence-based response: CS+sequence > CS− (z=1.92, PFDR=0.042, r=0.30); CS+element not > CS− (z=−0.41, PFDR=0.66); CS+sequence > CS+element (z=2.23, PFDR=0.042, r=0.34). LME interaction (CStype×Session Acquisition vs Long-term): β=−2.16, P=0.032.
• Decoding: Both HPC and DLPFC encode sequence representations across sessions. • Sound-cues epoch decoding above chance for both ROIs across sessions (e.g., DLPFC: PFDR≤0.0015; HPC: PFDR≤0.048). • US anticipatory epoch: Both maintain representations; DLPFC consistently superior to HPC (ROI×Epoch interaction P=0.026; main effect of ROI P=0.0005). • HPC shows a recency bias, maintaining separate CS representations better when last-cue identity differs (CS+ vs CS−) than when only sequence differs (CS+sequence vs CS+element; main effect in HPC P=0.023).
• Information transmission (seed→VMPFC–Amygdala) mirrors behavioral shift: • HPC→VMPFC–Amygdala transmission of CS+sequence vs CS+element decreases from Day 1 to Day 2 (A→L: z=3.16, PFDR=0.0024; I→L: z=3.45, PFDR=0.0017), becoming weaker than DLPFC on Day 2 (z=−3.66, PFDR=0.00038). DLPFC→VMPFC–Amygdala remains stable across sessions. • Similar patterns when targeting VMPFC or Amygdala separately (controls). • No time-dependent changes in communication with insula/thalamus or sensory cortices (controls).
• Trait anxiety moderates VMPFC communication balance on Day 2: • Greater DLPFC-over-HPC advantage in transmitting CS+sequence vs CS+element to VMPFC correlates with lower trait anxiety (r=−0.49, P=0.0012; robust β=−12.4, P=0.0003). No such relation for Amygdala (r=0.025, P=0.88); correlations differ (z=2.45, P=0.015).
• Component-wise transmission to VMPFC (vs CS−): • CS+element vs CS−: HPC→VMPFC decays by Day 2 to near chance; DLPFC→VMPFC remains constant, yielding DLPFC advantage (z=3.0, PFDR=0.0041; ROI×Session interaction P=0.018). Stronger DLPFC→VMPFC coupling on Day 2 correlates with greater suppression of SCR to CS+element (r=−0.37, P=0.018; robust β=−0.064, P=0.035), not observed on Day 1 (difference z=2.42, P=0.015). • CS+sequence vs CS−: Both HPC and DLPFC communicate stably; DLPFC consistently higher (main effect of ROI P=0.0013; between-ROI PFDR=0.024–0.005).

The study demonstrates a qualitative overnight transformation of human fear memory expression: from an overgeneralised cue-based response immediately after a threatening episode to a selective sequence-based response 24 hours later. This behavioral shift aligns with a neural rebalancing in communication to the fear-regulatory circuit: HPC, which preferentially conveys threat proximity via the last cue, withdraws its influence on VMPFC by Day 2, while DLPFC consistently communicates temporal sequence information. This arbitration allows rapid, generalisable defense shortly after trauma (cue-based) and subsequently promotes precise, episode-dependent control (sequence-based), potentially reducing false alarms. Heightened trait anxiety disrupts this rebalancing, with diminished DLPFC predominance in VMPFC coupling on Day 2, offering a mechanistic account linking exaggerated cue-based fear and incoherent episodic memory found in post-traumatic psychopathology. The findings situate fear memory within multiple competing neural representations whose expression is gated by dynamic HPC–DLPFC–VMPFC interactions over time, extending prior knowledge of HPC/DLPFC roles in temporal order memory to the domain of aversive learning and regulation.

This work introduces an episodic threat-conditioning paradigm revealing that fear memories transition from cue-association to episodic sequence-based expression over 24 hours. Multivariate fMRI shows that this transition is governed by a time-dependent shift in communication to VMPFC: HPC’s initial, last-cue–oriented signaling diminishes by Day 2, granting DLPFC’s sequence-based governance. Trait anxiety attenuates this shift, implicating altered HPC/DLPFC balance in anxiety-related fear memory dysfunctions. The study unifies strengthened cue associations and weakened episodic coding after trauma under a competitive, time-dependent framework and identifies an intervention window between Day 1 and Day 2. Future research should: (1) probe offline consolidation mechanisms (e.g., sequence replay, sleep) and their modulation to enhance DLPFC-mediated sequence integration; (2) establish causal directionality with intracranial or causal perturbation methods; (3) test generalisation to spatiotemporal contexts; and (4) explore therapeutic strategies that bolster DLPFC–VMPFC communication during consolidation.

• Directionality constraints of fMRI: while results suggest seed-to-target signaling (HPC/DLPFC→VMPFC), validating causal directionality requires intracranial or causal approaches; inverted analyses (VMPFC→HPC/DLPFC) did not show analogous effects.
• Contingency awareness: preliminary evidence indicates sequence-based fear emerged largely without explicit awareness; the relationship between sequence learning and consciousness remains unresolved.
• Reinforcement rate and stimulus complexity: the 37.5% US reinforcement was similar to prior work, but higher reinforcement or different complexities might alter contingency awareness and learning, requiring future tests.
• Anxiety effects: correlations with SCR were non-significant and evidence regarding anxiety-related behavioral expression was preliminary, warranting replication and extension.