Human claustrum neurons encode uncertainty and prediction errors during aversive learning

This study demonstrates that human claustrum neurons encode higher-order cognitive variables like uncertainty and prediction errors during aversive learning, revealing distinct temporal dynamics in their tracking of latent task states compared to the anterior cingulate cortex and amygdala.

The Gist

Imagine your brain is a high-tech spaceship command center, and you are the captain navigating through a field of dangerous asteroids. Your goal is to fly through the safe gaps without crashing. This is exactly what the participants in this study did, but the researchers were secretly listening in on the brain's "control room" to see how it figured out the rules of the game.

The study focuses on a tiny, mysterious part of the brain called the Claustrum (let's call it the "Switchboard"). For a long time, scientists knew this Switchboard was connected to almost every other part of the brain, but they didn't know what it actually did. Was it just a passive relay station passing messages? Or was it a smart manager making decisions?

Here is the story of what they found, explained simply:

The Game: Flying Through Uncertainty

The participants played a game where asteroids flew at them. Sometimes there was a safe hole at the top, sometimes at the bottom, and sometimes the rules changed without warning.

• The Challenge: You couldn't just react when you saw an asteroid; you had to guess where the safe hole was before it appeared.
• The Brain's Job: Your brain had to constantly update its "map" of the world. It had to ask: "How sure am I that the top is safe?" (Uncertainty) and "Did I just guess wrong?" (Prediction Error).

The Three Key Players

The researchers put tiny microphones (electrodes) into three different areas of the brain to listen to the neurons (brain cells) talking:

1. The Claustrum (CLA): The mysterious Switchboard deep in the center.
2. The Anterior Cingulate Cortex (ACC): The "Chief Strategist" in the front of the brain, known for handling complex decisions.
3. The Amygdala (AMY): The "Alarm System" that usually handles fear and emotions.

The Big Discovery: The Switchboard is Smart!

The researchers found that the Claustrum wasn't just a passive wire. It was actively thinking.

1. It Tracks "How Sure Are We?" (Uncertainty)
Imagine you are guessing the weather. If you are 90% sure it will rain, you feel calm. If you are 50/50, you feel anxious.

• The Finding: The Claustrum neurons fired differently depending on how "anxious" the brain was about the next asteroid. When the brain was unsure, the Claustrum lit up. It wasn't just reacting to the asteroid; it was reacting to the feeling of not knowing.

2. It Tracks "Oops, I Was Wrong!" (Prediction Error)
When you guess the safe hole is at the top, but you crash because it was at the bottom, your brain registers a "Prediction Error."

• The Finding: The Claustrum neurons fired strongly when the brain realized it had made a mistake, especially when the brain was already feeling unsure. It's like the Switchboard shouting, "Hey! The rules changed! We need to update the map immediately!"

The Twist: How the Switchboard Differs from the Strategist

While both the Claustrum (Switchboard) and the ACC (Strategist) were doing this smart thinking, they did it at different times and in different ways:

• The Strategist (ACC): This area was like a calm planner. It kept track of uncertainty between turns, maintaining a steady sense of "how sure we are" before the next asteroid even appeared. It was the one holding the long-term strategy.
• The Switchboard (Claustrum): This area was like a rapid-response team. It didn't worry much about the time between turns. Instead, it exploded into action the moment the asteroid appeared and the outcome was revealed. It was the one screaming, "CRASH! We need to change our plan NOW!"

The Analogy:
Think of the ACC as the Project Manager who sits in a meeting room, reviewing the schedule and worrying about potential risks before the workday starts.
Think of the Claustrum as the Floor Manager on the factory floor. They don't worry about the schedule; they react instantly when a machine breaks or a part arrives late, coordinating the workers to fix it immediately.

What About the Alarm System?

The Amygdala (the fear center) was surprisingly quiet in this specific game. It reacted to crashes, but it didn't seem to care about the complex math of "uncertainty" or "prediction errors." It was more focused on the immediate "ouch" of a crash rather than the abstract logic of the game.

Why Does This Matter?

This study is a huge deal because it proves that the Claustrum is a key player in human intelligence. It's not just a relay station; it's a dynamic coordinator that helps us adapt when the world gets confusing.

• Real-world impact: This helps us understand why people with conditions like schizophrenia or autism (where the brain struggles to filter information and update beliefs) might have trouble with the Claustrum. If the Switchboard is broken, the brain can't coordinate the "update the map" signal fast enough, leading to confusion or sensory overload.

In a nutshell: Your brain has a hidden Switchboard that acts like a rapid-response coordinator. It doesn't just pass messages; it actively monitors how unsure you are and how wrong your guesses were, helping you adapt instantly when life throws you a curveball.

Technical Summary

1. Problem Statement

The human brain's ability to adapt to changing environments relies on continuous updating of internal models based on sensory input and outcomes. While the anterior cingulate cortex (ACC) is well-established as a region encoding higher-order cognitive variables like uncertainty and prediction error (PE), the functional role of the claustrum (CLA) remains poorly understood.

• The Gap: The CLA is a thin, sheet-like structure reciprocally connected to nearly the entire neocortex, theoretically positioned to coordinate distributed cortical processing. However, direct evidence linking human CLA activity to latent cognitive variables (e.g., abstract task states, uncertainty) is scarce. Previous animal studies have yielded conflicting results, with some suggesting the CLA is a passive sensory relay and others suggesting a role in conscious integration.
• Objective: To determine if human CLA neurons encode latent task variables (specifically uncertainty and prediction error) during aversive learning, and to compare these dynamics with the ACC and the amygdala (AMY).

2. Methodology

Participants and Recording

• Cohort: Seven patients with medication-refractory epilepsy undergoing intracranial monitoring for seizure localization.
• Recording Technique: Robotic implantation of hybrid depth electrodes (Behnke-Fried) containing 8 microwires each.
• Targets: Single-unit recordings were obtained from:
  • Claustrum (CLA): 4 subjects, 6 microwire bundles (110 units).
  • Dorsal Anterior Cingulate Cortex (dACC): 6 subjects, 8 bundles (80 units).
  • Basolateral Amygdala (AMY): 3 subjects, 4 bundles (75 units).
• Localization: Verified via co-registration of preoperative MRI and postoperative CT, normalized to MNI space.

Behavioral Task

Participants played a computer-based "spaceship" game requiring aversive learning under uncertainty:

• Task: Navigate a spaceship to avoid asteroids. Two zones (Top/Bottom) had dynamically changing probabilities of containing a "safe hole" (90% vs. 10%).
• Constraints: Asteroids moved too fast for reactive correction; participants had to use anticipatory positioning based on learned probabilistic beliefs.
• Outcomes: Successful avoidance (gain) or crash (loss of integrity/health).
• Structure: Non-stationary blocks of 20–80 trials where safety probabilities switched, forcing continuous belief updating.

Computational Modeling

• Model: An Asymmetric Leaky Beta (ALB-sticky) model was used to infer trial-by-trial latent variables.
  • Uncertainty: Derived from the posterior variance of the beta distribution representing safety probability.
  • Prediction Error (PE): Calculated as the absolute mismatch between the observed outcome and the estimated safety probability.
  • Behavioral Validation: Model-inferred uncertainty correlated with behavioral metrics (higher uncertainty led to larger and more variable spaceship movements).

Neural Analysis

• Preprocessing: Spike sorting, stability assessment, and alignment to task events (asteroid appearance, outcome).
• Statistical Approaches:
  • Permutation Tests: To identify task-modulated neurons (appear-specific, outcome-specific).
  • Linear Mixed-Effects Models: To analyze region $\times$ outcome interactions.
  • Conditional Mutual Information (CMI): To quantify the dependence between neural activity and latent variables (uncertainty/PE) while controlling for confounds (spaceship position, outcome type).
  • Decoding: Support Vector Machine (SVM) classifiers to test if latent variables could be linearly decoded from single-neuron firing rates.

3. Key Results

A. Task Engagement and Response Patterns

• Engagement: Both CLA and ACC showed high engagement (70% of neurons were task-responsive), whereas the AMY showed significantly lower engagement (30%).
• Temporal Dynamics: CLA neurons recruited earlier than ACC neurons following asteroid onset.
• Response Types:
  • CLA: Diverse patterns including "appear-specific" (sensitive to stimulus onset), "appear-and-outcome," and "outcome-specific." Notably, the CLA population showed a bias toward crash outcomes (higher firing after crashes than avoidance).
  • ACC: Also showed diverse patterns but lacked a consistent population-level bias toward crashes; subpopulations were heterogeneously tuned (some preferred crashes, others avoidance).

B. Encoding of Latent Variables

• Uncertainty & Prediction Error (PE):
  • CLA: ~28% of neurons encoded uncertainty, and ~23% encoded PE. Crucially, these signals were predominantly active during the asteroid-avoidance period (when outcomes occurred). CLA neurons showed enhanced responses to crashes specifically under high-uncertainty conditions.
  • ACC: ~24% encoded uncertainty and ~26% encoded PE. Unlike the CLA, ACC neurons carried uncertainty signals during the intertrial period (before the asteroid appeared). ACC neurons showed stronger responses under low-uncertainty conditions, suggesting a role in maintaining belief precision.
• Conditional Mutual Information (CMI): Confirmed that both regions encoded latent variables beyond what could be explained by external sensory variables (position, outcome).
  • CLA: Higher fraction of latent-variable encoding during the active avoidance phase.
  • ACC: Higher fraction of uncertainty encoding during the intertrial phase.

C. Decoding

• CLA & ACC: Latent variables (uncertainty and PE) could be linearly decoded from single-neuron firing rates during the asteroid-avoidance period.
• AMY: No significant decoding of latent variables was observed.
• Intertrial Period: Neither region could decode PE (as it requires an outcome), but ACC showed weak uncertainty decoding, while CLA did not.

4. Key Contributions

1. First Human Single-Unit Evidence: Provides the first direct evidence that human claustrum neurons encode abstract, model-derived latent variables (uncertainty and prediction error).
2. Dissociable Roles: Reveals a functional dissociation between the CLA and ACC:
   • ACC: Maintains and monitors internal model reliability (precision) during periods of inactivity (intertrial).
   • CLA: Signals when outcomes violate expectations, particularly when the internal model is imprecise (high uncertainty), facilitating rapid updating.
3. Hierarchical Inference: Supports the view that the CLA is part of a hierarchical inference system, potentially acting as a "gate" that releases bottom-up error signals when top-down precision (from ACC) is low.
4. Amygdala Contrast: Demonstrates that while the amygdala encodes valence and outcome, it does not appear to encode the specific latent belief-state variables derived from the Bayesian model in this context.

5. Significance

• Theoretical Impact: The findings challenge the view of the claustrum as a passive sensory relay. Instead, it positions the CLA as an active participant in predictive processing, specifically in coordinating cortical networks to adapt to dynamic environmental demands when confidence in internal models is low.
• Clinical Relevance: Given the CLA's structural and functional abnormalities in disorders like schizophrenia, autism, and during psychedelic states, these findings suggest that dysfunctions in hierarchical inference and precision-weighted error signaling may underlie the sensory-perceptual disturbances seen in these conditions.
• Mechanistic Insight: The study proposes a circuit mechanism where ACC inputs provide "precision signals" that tonically suppress CLA output; when uncertainty is high, this suppression weakens, allowing the CLA to propagate salient error signals to the cortex for rapid adaptation.

In summary, this study establishes the human claustrum as a critical node in the brain's inferential hierarchy, distinct from but complementary to the ACC, in tracking the hidden states of the environment to guide flexible behavior.