The role of the human hippocampus in decision-making under uncertainty

The role of the hippocampus in decision-making is beginning to be more understood. Because of its prospective and inferential functions, we hypothesized that it might be required specifically when decisions involve the evaluation of uncertain values. A group of individuals with autoimmune limbic encephalitis—a condition known to focally affect the hippocampus—were tested on how they evaluate reward against uncertainty compared to reward against another key attribute: physical effort. Across four experiments requiring participants to make trade-offs between reward, uncertainty and effort, patients with acute limbic encephalitis demonstrated blunted sensitivity to reward and effort whenever uncertainty was considered, despite demonstrating intact uncertainty sensitivity. By contrast, the valuation of these two attributes (reward and effort) was intact on uncertainty-free tasks. Reduced sensitivity to changes in reward under uncertainty correlated with the severity of hippocampal damage. Together, these findings provide evidence for a context-sensitive role of the hippocampus in value-based decision-making, apparent specifically under conditions of uncertainty.

Thus, for every possible circle placement, an expected error can be computed as:

Computational modelling of active information gathering
To further characterise active information sampling performance in Exp. 1, we analysed the behaviour using a well-validated computational model previously implemented in healthy and patient groups 1,2 .
The model calculates the expected utility of a sample (EU s ) accounting for economic and hidden cognitive effort costs to return five parameter estimates per participant.The first two parameters represent the weights participants assign to sample costs (w s ) and benefits (w e ).Two parameters describe the cognitive cost function η c (ISI, α) in terms of a penalty for sampling speed (w speed ) and efficiency (w α ).The fifth parameter represents an intercept per participant describing their baseline valuation of samples (w 0 ).
This was formalised quantitatively as follows: EU s (ISI, α, t max ) = EU s−1 + p(s|ISI, t max ).(5) where η e is the placement error penalty (1.2 credits/pixel) and t max is the allowed search time per trial (18 seconds).ÊE ∞ is the per-individual information sampling asymptotic limit estimated beforehand to take into consideration inter-participant variations in asymptotic information sampling performance.
Based on previous work 1,2 , we used quadratic cognitive cost function as follows: To obtain the likelihood function, softmax function was applied over the 3-dimensional space of EU (EU depends on ISI, α, s) for a given task condition as follows: For each individual, model fitting involved findings the parameters that achieved the lowest negative log-likelihood of observing the multivariate distribution of the number of samples acquired (s), inter-sampling interval (ISI) and sampling efficiency (α).
Optimisation of parameters was performed in MATLAB (The MathWorks inc., version 2019a) using Bayesian Adaptive Direct Search (BADS 3 ).Further information about this modelling framework is provided in 1,2 .
After the exclusion of potential outliers (1 patient with values > 3SD), comparing parameter estimates between two groups showed that ALE patients had lower weights assigned to sampling cost compared to controls (t 35 = −2.24,95%CI = [−0.077,−0.003], p = 0.0315, Cohen ′ s d = −0.72; Figure S1).There was no significant difference between the two groups in any of the other parameters.These results thus represent a computational formalisation of the findings from Exp. 1 suggesting that ALE patients have lower sensitivity to the cost of sampling.It is noteworthy that Exp. 4, as expected, demonstrated significantly increased decision time (β = 1.18, t 91 = 7.68, p < 0.0001), consistent with the more complex decision-making process involving the consideration of three attributes.This result further supports the argument against rapid responding discussed in the subsection of the Results.
While these findings shed light on task performance difficulty, it is important to acknowledge that reaction times may not entirely negate the presence of a complexity effect.To control for this effect, it might be necessary to conduct novel experiments with a redesigned task, requiring participants to infer uncertainty and effort levels using analogous cues (e.g., levels on a bar).This approach aims to eliminate any additional cognitive effort needed to infer these attributes.

Supplementary Figures
Figures S3 to  Inter-sampling Interval (sec)

Controls Patients
Figure S3: Active sampling (Exp. 1) -ALE patients commit to decisions at similar uncertainty levels as controls.a. Final uncertainty is the expected error (EE) in pixels (Px) that a participant is likely to obtain at the end of their search.In the experimental condition where ALE patients over-sampled more than controls, there was no significant difference between ALE patients and controls in this measure (z = −1.60,p = 0.108, Clif f ′ s δ = −0.30).b.Similarly, the actual error that participants obtained upon localising the circle (distance to hidden circle in pixels) was not significantly different between patients and controls (z = −0.81,p = 0.413, Clif f ′ s δ = −0.15).These two results indicate that ALE patients wasted monetary resources on samples with limited utility (i.e., over-sampled).c.There is no change in choice performance results when subjective estimates of uncertainty are used instead of expected error (EE) in the analysis.ALE patients (N = 19) demonstrate lower sensitivity to reward and intact sensitivity to uncertainty when compared to healthy controls (N = 19).Reward levels 1-4 correspond to the number of credits on display (R: 40, 65, 90, 115 credits).Subjective uncertainty levels were calculated by binning sign-flipped z-scored confidence ratings into five bins.level five describes the lowest level of subjective uncertainty estimate.Error bars and shading represent ±SEM.For statistical details see Table S18.

*Figure S1 :
FigureS1: Computational modelling of active information sampling (Exp.1).Compared to healthy matched controls, ALE patients assigned lower economic costs (w s ) to sample acquisition (t 35 = −2.24,95%CI = [−0.077,−0.003], p uncorr = 0.031, Cohen ′ s d = −0.72,18 patients and 19 controls).All other model parameters including weights assigned to sample benefit (w e ), efficiency w α , and speed (w speed ) were not significantly different between patients and controls.w 0 captures a subjective fixed cost of sampling that is not explicitly specified in the task (e.g., cost of the motor action).This was not significantly different between the two groups.Error bars show ± SEM.

Figure S2 :
Figure S2: Decision times measured in seconds (sec) in Exps.2-4.Across passive decision tasks (Exps.2-4), no significant difference was found between ALE patients (N = 19 in Exp. 2 & 3, N = 8 in Exp. 4) and controls (19 in Exp. 2 & 3, N = 12 in Exp. 4).ALE patients made faster decisions in Exp. 2 compared to Exp. 3, indicating less deliberation when making decisions under uncertainty compared to effort-based decision making.Exp. 4 had significantly slower decisions, reflecting the more complex task structure with three decision attributes to consider.Error bars and shading represent ±SEM.For full statistical details see TableS24 Figures S3 to S8

Figure S5 :Figure S6 :
Figure S5: Amygdala as control region.No significant correlation was detected between amygdala volume and sensitivity to reward or uncertainty (Robust regression p > 0.20 for all correlations).

Figure S7 :
Figure S7: Passive choices as a function of reward and subjective uncertainty estimates.There is no change in choice performance results when subjective estimates of uncertainty are used instead of expected error (EE) in the analysis.ALE patients (N = 19) demonstrate lower sensitivity to reward and intact sensitivity to uncertainty when compared to healthy controls (N = 19).Reward levels 1-4 correspond to the number of credits on display (R: 40, 65, 90, 115 credits).Subjective uncertainty levels were calculated by binning sign-flipped z-scored confidence ratings into five bins.level five describes the lowest level of subjective uncertainty estimate.Error bars and shading represent ±SEM.For statistical details see TableS18.
All other model parameters including weights assigned to sample benefit (w e ), efficiency w α , and speed (w speed ) were not significantly different between patients and controls.w 0 captures a subjective fixed cost of sampling that is not explicitly specified in the task (e.g., cost of the motor action).This was not significantly different between the two groups.Error bars show ± SEM.

Table S1 :
Minimal effect of cognitive deficit and memory decay on performance.a.There was no significant correlation between cognitive scores indexed by ACE-III scores and sensitivity to either reward or uncertainty in Exp. 2, indicating that the difference between ALE patients and controls is likely related to cognitive dysfunction.b.Whether trials were played in the second half of the experiment compared to the first half did not have a significant effect on reward sensitivity, suggesting minimal presence of memory decay that could influence behaviour or result in random responding.c.Catch trials in Exp. 4 show that ALE patients had intact sensitivity to uncertainty (right panel) and blunted sensitivity to effort (middle panel), Demographics.ACE-III: Addenbrooke's Cognitive Examination.DS: Digit Span.
pointing against random responding during the task, and replicating results from the main task trails.Reward sensitivity (left panel) is intact in these trials but this should be interpreted with caution as reward represented in a balanced design in these catch trials.Shaded area around the lines indicate ±SEM .For statistical details see Tables S20, S22 & S23.AMI: Apathy Motivation Index.FSS: Fatigue Severity Scale.BDI-II: Beck Depression Inventory.SHAPS: Snaith-Hamilton Pleasure Scale.Hipp: Adjusted Hippocampal Volumes.* 15 patients and 17 controls.Statistical testing was performed with a two-sample t-test if data fulfilled parametric assumptions or a Wilcoxon rank-sum test if assumptions were violated.

Table S2 :
Patients Characteristics.Abs: Autoantibodies.Lt. Hipp.: Left Hippocampus.Rt.Hipp.: Right Hippocampus.Hippocampal volumes were adjusted for intra-cranial volumes.Percentile is determined by plotting raw hippocampal volumes against normative brain volumes from UK biobank data 4 .c : Describes percentiles outside the age range of the UK biobank nomograms.Percentiles according to the closest age value within the UK biobank range was used instead.

Table S7 :
Active Search (Exp. 1) -Generalised mixed-effects model investigating the effect of ALE on efficiency (α) in the condition with high sampling cost and high initial reward reserve, i.e., the condition where ALE patients over-sampled more than controls.Model was specified as follows.: α ∼ 1 + group*ISI + (1 |trial) + (1 + ISI |participant).

Table S20 :
Robust regression model investigating the correlation between the subdomains of Addenbrooke's Cognitive Examination (ACE III) and sensitivity to reward and uncertainty in ALE patients.Behavioural data is from Exp. 2. Models were specified as follows: Sensitivity ∼ 1 + Attention + Memory + Fluency + Language + VisuoSpatial.