Altered predictive control during memory suppression in PTSD

The study addresses how disrupted predictive processing in PTSD affects the control of intrusive memories, a core symptom of the disorder. PTSD is characterized by avoidance of trauma reminders and altered expectations of future threat. Bayesian accounts suggest aberrant associations between safe cues and aversive outcomes, impairing prediction and reinforcing avoidance. Prior work showed impaired inhibitory control over intrusive memories in PTSD, with reduced top-down modulation from right DLPFC to memory regions such as hippocampus and precuneus during a Think/No-Think (TNT) suppression task. Cognitive control can operate proactively (predictive) and reactively (error-driven). The authors hypothesized that inhibitory control over memory relies on predictive inferences, and that PTSD involves an imbalance between predictive and reactive control: exaggerated reliance on beliefs about upcoming intrusions (predictive control) and/or reduced reactive control to prediction errors when intrusions occur. The study aims to track hidden computations (beliefs and prediction errors) during suppression and relate them to effective connectivity within control-memory circuits in PTSD+, PTSD−, and nonexposed individuals.

The paper situates PTSD within predictive processing frameworks, where aberrant threat predictions and hypersensitivity to prediction error influence behavior and cognition. Previous TNT studies identified a key role for right DLPFC (including anterior and posterior middle frontal gyrus) in suppressing unwanted memories via top-down inhibitory coupling to hippocampus and precuneus. PTSD shows impaired modulation of these networks during intrusions. Dual-mechanism models of cognitive control distinguish proactive (prediction-based) and reactive (prediction-error-based) processes, with Bayesian models capturing dynamic updating of beliefs and uncertainty. The authors reference evidence of increased PE-related attentional demand and structural/functional alterations in DLPFC and white matter in PTSD, and neurobiological models positing distinct pathways for proactive (rhinal gating) and reactive (thalamic reuniens to hippocampal interneurons) inhibition. They also discuss transdiagnostic symptom dimensions (anxiety- and affect-related) contrasted with PTSD-specific intrusion and avoidance.

Design and participants: Final sample included 101 trauma-exposed participants (55 PTSD+, 46 PTSD−) and 72 nonexposed controls. Exposed participants were recruited via the Program 13-Novembre; PTSD diagnosis (full or partial) was established using SCID for DSM-5. Symptoms were assessed with PCL-5; anxiety and depression with STAI and BDI. Participants were 18–60 years old, right-handed, French-speaking, MRI-eligible. Data collection occurred between June 2016 and June 2017. Task and procedure: Participants learned 54 neutral French word–object pairs to criterion (>90% correct), then performed a Think/No-Think (TNT) memory suppression task during fMRI across four sessions. On No-Think trials (red cue), they used direct suppression to prevent retrieval; after each trial they reported whether the associated object entered awareness (intrusion yes/no). Intrusion ratings were obtained trial-wise for No-Think items. MRI acquisition and preprocessing: 3T Philips Achieva. Structural T1 (1 mm3) and four functional EPI runs (TR=2050 ms, TE=30 ms, 32 slices, 3 mm thickness, 235 volumes/run). Preprocessing in SPM12: realignment, slice timing correction, coregistration, tissue segmentation, native-space ROI extraction using deformation fields, GLM with task regressors (think/no-think), motion parameters, session dummies, filler regressors; AR(1) noise model and whitening. Computational modeling of beliefs and prediction errors: Three perceptual models were fit to trial-by-trial No-Think intrusion ratings using TAPAS toolbox with variational Bayes: (1) Rescorla–Wagner (RW) with fixed learning rate α; (2) Kalman filter (KF) with dynamic learning via parameters π (reliability) and ω (uncertainty); (3) two-level Hierarchical Gaussian Filter (HGF2), modeling hierarchical beliefs with volatility parameter ω, uncertainty-weighted PE updates. For each perceptual model, three source models mapped belief formation to outcomes: state (based on overall trial history), item (based on item-specific repetition history), and combined (precision-weighted combination of state and item beliefs). Observation model used a beta density mapping with an inverse decision noise parameter to compute log-likelihood of binary outcomes (intrusion/nonintrusion). Model accuracy was the negative log-likelihood sum. Simulations validated generative performance (model falsification), trajectory recovery, model recovery, and parameter recovery. Dynamic causal modeling (DCM): Effective connectivity was modeled between right anterior MFG (aMFG), posterior MFG (pMFG), rostral hippocampus (rHIP), caudal hippocampus (cHIP), and precuneus (PC). ROIs were derived from the Brainnetome atlas and refined in native space by selecting the 30 most task-related voxels (No-Think>Think for MFG; Think>No-Think for memory regions). All models assumed bidirectional intrinsic connections. Modulatory inputs were boxcar functions for No-Think trials (parametrically modulated by beliefs) and for intrusive No-Think trials (parametrically modulated by positive prediction error, PE+ = outcome − belief when outcome=intrusion). Model space: 42 models across families—(i) computationally guided top-down modulation (beliefs and PE+ modulate top-down MFG→memory coupling), with subfamilies swapping aMFG/pMFG roles in predictive vs reactive control and seven target-pathway configurations; (ii) computationally guided bottom-up modulation; (iii) top-down modulation without parametric computations; plus a null family with no controlled modulation. Random-effects Bayesian model selection (BMS) compared families; Bayesian model averaging (BMA) across the 14 models of the winning family yielded participant-level coupling parameters for predictive (belief) and reactive (PE+) control. Statistics: one-tailed t-tests with FDR correction on four effects (Control×Group interactions; Control effect within groups; reactive negative coupling; predictive negative coupling) across rHIP, cHIP, PC, and whole hippocampus (wHIP=rHIP+cHIP). Posterior probabilities, 95% bootstrapped CIs, and Bayes factors (MCMC) reported. An imbalance analysis projected predictive and reactive coupling onto orthogonal axes to compute a resultant vector angle (imbalance index) with circular statistics for group comparisons. Ethics: Approved by Comité de Protection des Personnes Nord-Ouest III; informed consent obtained. Data/code availability provided.

- Model validation and selection: Only the HGF produced valid, reliable trajectories that simulated behavioral intrusion reductions, supported trajectory/model/parameter recovery. Bayesian model selection across source models favored the combined source model (PXP=0.999; BOR=0; P(H_py)=0.996), indicating that beliefs integrate state and item histories with uncertainty weighting. Power analyses indicated sufficient power to detect between-group differences in HGF-item ω. - Belief updating parameter: PTSD+ showed significantly slower updating for item-derived beliefs compared to nonexposed (t(122) = −2.10, p = 0.037, bootstrapped 95% CI [−0.59, 0.06]) and a trend vs PTSD− (t(99) = 1.82, p = 0.072; bootstrapped 95% CI [−0.58, −0.04]; significant with bootstrapped mean). No difference between PTSD− and nonexposed (t(113) = 0.15, p = 0.880). - DCM family comparison: Strong evidence that beliefs and PE+ modulated top-down MFG→memory coupling (PXP=0.886), with bottom-up, no-computation, and null families not supported (PXP≈0; fBOR=0). No reliable differentiation between aMFG and pMFG roles in predictive vs reactive control (PXP=0.343 vs 0.657; fBOR=0.677). The preferred family architecture generalized across groups (P(H_py)=0.968). - Predictive vs reactive control imbalance in PTSD+: Significant Control×Group interactions showed disproportionate negative coupling (stronger inhibition) during predictive relative to reactive control in PTSD+ vs nonexposed for rHIP (t(125) = −2.81, PFDR=0.007, PP=0.999, 95% CI [−0.99, −0.17]) and vs PTSD− (t(99)=−2.17, PFDR=0.009, PP=0.999, 95% CI [−1.01, −0.01]). For cHIP: PTSD+ vs PTSD− (t(99)=−3.23, PFDR=0.006, PP=1, 95% CI [−1.20, −0.29]); trend vs nonexposed (t(125)=−1.62, PFDR=0.071, PP=0.995). Whole hippocampus (wHIP): PTSD+ greater imbalance vs nonexposed (t(125)=−2.49, PFDR=0.014, PP=0.992, 95% CI [−0.81, −0.09]) and vs PTSD− (t(99)=−2.91, PFDR=0.007, PP=0.998, 95% CI [−1.06, −0.19]). No PC differences between groups. - Within-group coupling patterns (BMA): Reactive negative coupling (inhibitory control to PE+) over hippocampus was significant in nonexposed and PTSD− but absent in PTSD+. Predictive inhibitory control (belief-driven) over hippocampus was present in all groups. Only PTSD+ showed significantly stronger predictive than reactive inhibition within rHIP, cHIP, and wHIP. PC was proactively controlled (predictive inhibition) but not reactively controlled in PC across all groups. - Symptom correlations (PTSD+): Greater predictive-over-reactive imbalance (more predictive inhibition) over wHIP correlated with higher avoidance severity (Spearman R=−0.32; 95% CI [−0.52, −0.09]; Z=2.27; PFDR=0.047) and marginally with intrusions after correction (R=−0.26; 95% CI [−0.47, −0.03]; Z=1.84; PFDR=0.065). No significant associations with anxiety-related (R=0.04; 95% CI [−0.08, 0.16]; PFDR=0.30) or affect-related (R=0.09; 95% CI [−0.04, 0.23]; PFDR=0.18) dimensions. Predictive-control correlations with re-experiencing and avoidance were significantly stronger than with anxiety- or affect-related dimensions (PFDR≤0.018). The predictive-control–avoidance correlation was stronger in PTSD+ than PTSD− (Z=2.12; p=0.034). - Independence of control components: Predictive and reactive coupling magnitudes were not negatively related in PTSD+ (R=0.01; 95% CI [−0.26, 0.30]), supporting partially independent processes. - Imbalance angle (circular analysis): Mean imbalance toward predictive control in hippocampus was significant in PTSD+ (M=33.35°, 95% CI [20.2°, 46.2°]) and nonexposed (M=15.33°, 95% CI [4.55°, 26.51°]), not in PTSD− (M=6.86°, 95% CI [−9.17°, 23.8°]). PTSD+ showed greater predictive bias than PTSD− (t(99)=2.10, p=0.018, 95% CI [−46.8°, −4.2°]) and nonexposed (t(125)=1.74, p=0.042, 95% CI [−35.77°, −0.86°]); no difference between nonexposed and PTSD−.

The findings demonstrate that during intentional memory suppression, trial-wise predictions about the likelihood of intrusions (beliefs) and ensuing positive prediction errors jointly modulate top-down inhibitory coupling from right DLPFC (aMFG/pMFG) to hippocampus and precuneus. Healthy and resilient trauma-exposed individuals balance predictive and reactive control to mitigate hippocampal activity, whereas PTSD+ individuals exhibit excessive belief-driven (predictive) inhibition with diminished error-driven (reactive) inhibition when intrusions occur. This imbalance specifically relates to PTSD’s trauma-linked symptoms (avoidance, re-experiencing), not to transdiagnostic anxiety or affect dimensions, supporting predictive-processing accounts of PTSD. The absence of a trade-off between predictive and reactive coupling magnitudes argues for partially independent control computations within a shared inhibitory system. Potential mechanisms include limited executive resources (DLPFC gray/white matter alterations) that sustain background predictive control but constrain transient reactive regulation during intrusions, and/or alterations in inhibitory receptor/interneuron pathways (e.g., rhinal vs thalamic reuniens to hippocampal interneurons) affecting PE-driven suppression. The approach refines interpretations of prior TNT findings by disentangling predictive from reactive components and clarifying why PC but not reactive control effects appeared in PC across groups. Clinically, the results suggest that maladaptive predictive avoidance generalizes to memory control, emphasizing the need to restore balance between predictive and error-driven inhibitory processes.

This study integrates computational modeling with effective connectivity to show that memory suppression is governed by predictive (belief-based) and reactive (PE-based) inhibitory control from right DLPFC to hippocampus, and that PTSD is characterized by an imbalance favoring predictive over reactive control. This imbalance is linked to core PTSD symptoms (avoidance, re-experiencing), providing a mechanistic bridge between predictive-processing disturbances and impaired inhibitory control accounts of PTSD. The work highlights predictive control disorder as a potential contributor to the maintenance of traumatic memories. Future research should: (1) test causal links between control imbalance and formation/persistence of traumatic engrams; (2) probe neurobiological substrates (e.g., GABAergic interneuron pathways, DLPFC structural integrity); (3) assess whether balancing predictive and reactive control via interventions (e.g., targeted training, neuromodulation) reduces avoidance and intrusions; and (4) examine generalization to threat-related materials and longitudinal changes with therapy.

- Causality and memory trace formation remain untested: the study cannot determine whether the identified control imbalance contributes to the formation or persistence of traumatic engrams. - Stimuli were neutral and not trauma-related, potentially limiting ecological validity for PTSD-specific triggers. - Model parameter recovery: HGF parameter recovery was modest; RW/KF recovery was poor, and only HGF provided valid belief trajectories under this design. - Power to detect between-group differences in the HGF ω parameter was adequate for item-derived beliefs but low for state-derived beliefs. - No differentiation between aMFG and pMFG roles in predictive vs reactive control was observed, which may reflect limited sensitivity or genuinely shared functions. - Data access constraints (restricted raw data) may limit independent replication at the raw-data level, though derived DCM parameters and code are shared.