Hippocampal damage disrupts the latent decision-making processes underlying approach-avoidance conflict processing in humans

Approach-avoidance conflict (AAC) occurs when potential reward and punishment are simultaneously associated with a stimulus, creating competing tendencies to engage or retreat. Animal work, especially in rodents, points to the ventral/anterior hippocampus (vHPC/aHPC) as central to AAC resolution, with lesions or inhibition increasing approach under conflict. Human fMRI studies have shown greater anterior hippocampal activity with higher AAC, and lesion evidence suggests hippocampal dysfunction can bias toward approach, though prior human studies often involved spatial navigation demands or single cases. Despite clear involvement of the hippocampus (HPC), prior research has emphasized overt behavior, leaving the latent decision processes (e.g., evidence accumulation, bias, thresholds) unclear and the degree to which effects reflect memory versus decision mechanisms unresolved. The present study tested individuals with focal hippocampal damage and matched controls on AAC tasks using learned object and scene stimuli, and applied hierarchical drift diffusion modeling (hDDM) to decompose latent processes (bias, drift rate, threshold, non-decision time). Additional Stroop and Go/No-Go tasks probed whether hippocampal involvement generalizes to response conflict beyond value-based decisions. The key questions were whether hippocampal damage selectively alters evidence accumulation and decision thresholds under conflict (beyond mnemonic deficits), and whether any effects are stimulus-class specific (objects vs scenes).

• Rodent studies implicate vHPC in AAC: gross vHPC damage/inhibition increases approach to conflicting stimuli; subfield-specific manipulations dissociate effects on approach vs avoidance. Nonhuman primate hippocampal lesions facilitate reward retrieval near threat.
• Human neuroimaging shows increased anterior hippocampal activity during high AAC; prior lesion evidence suggests increased approach with HPC dysfunction but often involves spatial tasks or single cases.
• Prior work focuses on behavior (approach rate, latency), providing limited insight into latent computational mechanisms (bias, drift, thresholds). Computational models (e.g., DDM/hDDM) are increasingly used in AAC but rarely applied to hippocampal lesions.
• Theoretical accounts suggest HPC may act as a comparator, detecting conflict and biasing toward negative outcome representations, fostering behavioral inhibition. However, findings across subregions (e.g., CA1 vs CA3) and tasks indicate HPC roles are nuanced and potentially subregion- and network-dependent.
• Regional stimulus-class preferences (HPC for scenes/context; PRC for objects) could predict stimulus-specific AAC effects, though results have been mixed across species and paradigms.

Design and participants:

• Participants: 8 individuals with focal hippocampal damage (7 male, 1 female; mean age 63.9±8.3; etiologies: 7 autoimmune limbic encephalitis with VGKC-complex antibodies and focal hippocampal atrophy; 1 medial temporal lobe epilepsy with hippocampal sclerosis) and 25 neurologically healthy controls (14 male, 11 female; mean age 68.5±9.2). Groups did not differ in age or IQ (WASI-II). Inclusion/exclusion procedures led to removal of some datasets for specific tasks due to failure to learn or instruction misunderstandings; remaining data retained across other tasks. Background neuropsychology confirmed mild amnestic profile in patients and average-to-superior performance in controls.
• Setting: Single session for patients (home or hospital); two sessions for controls (in-person or synchronous remote via Zoom). Ethical approvals: University of Toronto REB #26827; South Central Oxford REC #08/H0606/133; informed consent obtained.

Tasks and procedures:

1. Approach-Avoidance Conflict (AAC) tasks (Object and Scene versions):

• Stimuli: Object task used unfamiliar computer-generated objects; Scene task used unfamiliar real-world scenes. Each task had 4 stimuli (2 positive, 2 negative) to minimize memory load.
• Structure: Learning phase (3 blocks × 40 trials; 120 trials total): single image presented; choice: approach (key ‘1’) or avoid (key ‘2’). Contingencies: approach positive = +100 points; approach negative = −100; avoid = 0. Immediate feedback (outcome and cumulative points) after each trial. Decision phase (3 blocks × 36 trials; 108 trials total): pairs of images without feedback in three conditions—No-Conflict Positive (two positives), No-Conflict Negative (two negatives), Conflict (one positive + one negative). Instructions: approach boosts points on No-Conflict Positive and loses points on No-Conflict Negative; Conflict approach yields 50% ±100; avoid yields 0. Participants encouraged to approach some Conflict trials. No feedback during decision phase to prevent new learning.
• Timing (learning): self-paced stimulus; 500 ms to feedback; feedback 1,500 ms; ISI 1,000 ms. (Decision): self-paced; ISI 1,000 ms. Display sizes adapted for in-person vs remote.

2. Stroop task:

• Conditions: Control (colored rectangle), Congruent (color word matches ink), Incongruent (color word mismatches ink). Responses via keys corresponding to ink color. 28 trials per condition (84 total). Feedback “X” for errors (400 ms). Emphasis on speed and accuracy.

3. Cued Go/No-Go task:

• Trial: fixation (500 ms), orientation cue (white rectangle; horizontal or vertical), SOA (100 or 500 ms), then target color (green = Go, press; blue = No-Go, withhold). Orientation cues predicted Go/No-Go ratio: vertical = 4:1 Go/No-Go; horizontal = 1:4 Go/No-Go. 250 trials. ITI 400 ms; trial ended after 700 ms if no response.

Data collection platforms:

• In-person: AAC in E-Prime 2; Stroop and Go/No-Go in Inquisit. Remote: PsychoPy tasks delivered via Pavlovia; neuropsychology adapted for remote administration.

Statistical analyses:

• Linear Mixed Models (LMMs): Implemented in R (lme4). Generalized LMMs for categorical choices; continuous LMMs for RT. Deviation coding; random effects structures optimized iteratively. Multi-parameter likelihood ratio tests used for factors with 3 levels (e.g., condition). Post hoc estimated marginal means (EMMs) with Tukey HSD; Bonferroni correction across 7 LMMs (Pcorr = p × 7).
• Hierarchical Drift Diffusion Modeling (hDDM): Python HDDM (PyMC). Single model across groups collapsed over Object/Scene (no stimulus-type effects in LMMs). Parameters estimated per condition for drift rate (v), threshold (a), non-decision time (t), with bias (z) estimated at group level and collapsed across conditions (best fit: DIC = 9848.59; all alternatives DIC > 9862). Participant-wise estimates for all except bias (group-level only). Convergence assessed via trace plots, Gelman–Rubin R < 1.01, and MC error relative to posterior SD; 55,000 samples (5,000 burn-in), thinning to 10,000. Hypothesis tests based on posterior overlaps (probability a sample from one posterior exceeds another).

Learning phase (AAC):

• Accuracy improved from Block 1 to Block 10 (Pcorr < 0.001). By Block 10, hippocampal damage and control groups showed comparable accuracy for both Scene and Object tasks (both Pcorr = 1.000). No valence differences at Block 10 for either task (both Pcorr = 1.000). RTs decreased from Block 1 to Block 10 (Pcorr < 0.001), with no group difference at Block 10 (Pcorr = 1.000). This indicates comparable acquisition and retention of stimulus-valence associations.

Decision phase (AAC choices and RTs):

• Main effect of condition on approach rates (Pcorr < 0.001): approach ~0 for No-Conflict Negative, ~1 for No-Conflict Positive, intermediate for Conflict.
• Critical group-by-condition interaction (Pcorr < 0.001):
  • No-Conflict Positive: controls vs hippocampal damage, no difference (Pcorr = 1.000).
  • No-Conflict Negative: no difference (Pcorr = 1.000).
  • Conflict: hippocampal damage group approached more often than controls (EMM contrast estimate −0.19±0.04; t = −4.68; Pcorr < 0.001; Cohen’s d = 0.70).
• RTs: Conflict trials slower than No-Conflict Positive (difference ~753.5 ms; Pcorr < 0.001) and No-Conflict Negative (~685.4 ms; Pcorr < 0.001), confirming elicited conflict. No group differences in RTs across conditions after correction (all Pcorr = 1.000).
• No interactions with stimulus type (objects vs scenes), indicating effects generalized across stimulus classes.

Hierarchical drift diffusion modeling (hDDM):

• Within groups across conditions:
  • Non-decision time (t): similar for No-Conflict Positive vs Negative (patients PPos>Neg = 0.493; controls PPos>Neg = 0.378), longer for Conflict vs No-Conflict (patients PConflict>Pos = 0.984; PConflict>Neg = 0.982; controls both >0.999).
  • Drift rate (v): as expected, Positive > Conflict > Negative directionality (e.g., PPos>Neg > 0.999; PPos>Conflict > 0.999; PNeg 0.999).
  • Threshold (a): weak evidence for lower thresholds on Conflict vs No-Conflict (patients PPos>Conflict = 0.956; PNeg>Conflict = 0.927; controls PPos>Conflict = 0.915; PNeg<Conflict = 0.856).
• Between groups:
  • No-Conflict trials: similar non-decision time, drift rate, and threshold (e.g., Ppatients<controls for v and a across No-Conflict conditions ≥ 0.723).
  • Conflict trials: strong differences—hippocampal damage group showed more positive drift rate (less rapid accumulation toward avoidance) relative to controls (Ppatients>controls = 0.995) and lower decision threshold (Ppatients<controls = 0.982); non-decision time similar (Ppatients<controls = 0.391). Overall starting bias (z) was more positive in hippocampal damage (Ppatients>controls = 0.992), indicating greater baseline approach propensity.

Response conflict tasks:

• Stroop: Accuracy—main effect of condition (Incongruent < Congruent/Control; Pcorr < 0.001), but no group or interaction effects (all Pcorr = 1.000). RT—Incongruent slower than Congruent (Pcorr = 0.014) and Control (Pcorr < 0.001); no group differences (all Pcorr = 1.000).
• Go/No-Go: Trend to higher inhibition errors on No-Go trials preceded by Go-predictive cues (p = 0.016; Pcorr = 0.112), but no significant group or group-by-cue effects (Pcorr ≥ 0.882).

Summary: Individuals with hippocampal damage specifically increased approach under motivational conflict with intact learning and non-conflict performance. hDDM revealed this reflects a more approach-biased starting point, reduced evidence thresholds, and reduced negative drift (slower accumulation toward avoidance) during conflict, while non-decision times were unaffected. No deficits were observed on classic response conflict tasks.

The study demonstrates a causal role for the human hippocampus in resolving approach-avoidance conflict by shaping latent decision processes rather than simply affecting memory for stimulus valence. Despite equivalent learning and non-conflict behavior, hippocampal damage increased approach choices under conflict. hDDM decompositions indicate that this shift arises from three complementary alterations: a stronger baseline approach bias, reduced decision thresholds (less caution) under conflict, and a failure to rapidly accumulate evidence favoring avoidance (drift rates near zero vs strongly negative in controls). These findings align with theoretical models positing a hippocampal role in behavioral inhibition and conflict detection via comparison of predicted negative and positive outcomes—normal HPC may preferentially weight potential aversive outcomes under conflict, fostering avoidance, whereas structural damage blunts this process.

Notably, stimulus class (objects vs scenes) did not modulate the hippocampal effect, contrary to expectations that scene/context demands would preferentially engage hippocampus and that object-based conflict might rely more on perirhinal cortex. This suggests broader hippocampal involvement in value-based conflict decision-making irrespective of visual category or that network-level dysfunction accompanying hippocampal atrophy (e.g., altered connectivity along anterior-posterior axes or with extra-hippocampal regions) may drive generalization across stimulus classes. The absence of group differences on Stroop and Go/No-Go indicates that hippocampal contributions are not simply explained by generic response conflict or inhibition deficits; rather, they appear specific to value-based motivational conflict and its evidence accumulation dynamics. Overall, these data provide mechanistic insight linking hippocampal integrity to cautious, avoidance-leaning evidence accumulation when faced with competing reward and punishment.

This study provides robust causal evidence that hippocampal integrity is critical for arbitrating approach-avoidance conflict by modulating latent decision-making processes. Individuals with hippocampal damage showed increased approach under conflict, driven by a more approach-biased starting point, reduced decision thresholds, and diminished accumulation of avoidance-supporting evidence. Learning and non-conflict performance were intact, and response conflict tasks showed no group differences, indicating specificity to motivational conflict.

Main contributions:

• Establishes hippocampal involvement in evidence accumulation mechanisms (bias, drift, threshold) during value-based conflict in humans.
• Demonstrates specificity to AAC rather than generic response conflict.
• Shows effects generalize across object and scene stimulus classes.

Future directions:

• Map hippocampal subregion and network-level contributions (e.g., anterior vs posterior, CA1 vs CA3; connectivity with amygdala, prefrontal, and perirhinal regions).
• Test generalizability across AAC paradigms using innate reinforcers/threats versus learned secondary reinforcers and across different schedules of outcomes.
• Develop paradigms minimizing mnemonic demands further to include larger lesion cohorts and refine links to psychiatric conditions (e.g., anxiety, depression) using computational phenotyping.
• Integrate multimodal imaging and lesion-symptom mapping to relate structural and functional connectivity to hDDM parameters under conflict.

• Small lesion sample (n=8) limits generalizability and precluded detailed analyses of hippocampal subregions and network connectivity; selection emphasized relatively circumscribed HPC atrophy, potentially limiting etiological diversity.
• AAC used secondary reinforcers (points) and learned associations; findings may differ for tasks with innate threat/reward or different reinforcement schedules.
• Although mnemonic demands were minimized, tasks still required learning/recall of stimulus-valence associations; subtle memory differences cannot be completely excluded despite matched performance by end of learning.
• Mixed in-person and remote testing necessitated task adaptations; while standardized, modality could introduce variability.
• No direct functional connectivity measures; potential network-level abnormalities accompanying hippocampal damage were not assessed.
• Stimulus-class null effects may reflect limited power to detect modest object vs scene differences or broader network disruptions beyond regional specialization.