Decoding Motor Action Timing and Metacognitive Signals from Single-Trial EEG Using a Transformer-Based Model (EEG-ViT)

This study demonstrates that metacognitive evaluation of motor timing is a computationally distinct process requiring the simultaneous integration of theta, alpha, and beta band EEG signals, rather than a direct readout of first-order performance signals.

The Gist

The Big Question: Are We Just "Reading the Dashboard" or "Running a Second Computer"?

Imagine you are driving a car. You press the gas pedal, and the car speeds up.

• The "First-Order" task: Actually driving the car (hitting the gas, steering).
• The "Metacognitive" task: Looking at the speedometer and thinking, "Hmm, I think I'm going a bit too fast. I need to slow down."

For a long time, scientists debated how the brain does this second part.

1. The "Single-Process" Theory: This suggests that your brain just looks at the same data it uses to drive the car. It's like reading the dashboard. If the car is fast, the brain says, "Oh, we are fast." The "feeling" of speed is just a direct readout of the driving data.
2. The "Higher-Order Representation" (HOR) Theory: This suggests the brain has a second, separate computer running in the background. This second computer doesn't just read the dashboard; it actively analyzes the driving data, combines it with other information, and creates a new thought: "I am driving too fast."

This paper asks: Is the brain just reading the dashboard (Theory 1), or is it running a separate, complex analysis program (Theory 2)?

---

The Experiment: The "2-Second Timer" Game

To test this, the researchers asked people to play a game with their brains and a computer mouse.

1. The Task: Participants had to click a mouse button to start a timer, wait exactly 2 seconds in their head, and click again to stop it.
2. The Twist: They had to do this without looking at a clock. They had to rely entirely on their internal sense of time.
3. The Metacognition: Immediately after stopping the timer, they had to answer two questions:
   • "How far off were you?" (Did you stop too early or too late?)
   • "How sure are you about that answer?"

While they did this, the researchers put a cap with 64 sensors on their heads to record their brain waves (EEG). This is like having a high-tech microphone listening to the electrical conversations inside the brain.

---

The Secret Weapon: The "Brain Transformer" (EEG-ViT)

The brain is messy. It's like trying to hear a single conversation in a crowded, noisy stadium. Traditional methods often just look at the "average" noise, which misses the specific details of what happened on one specific try.

The researchers used a new, fancy AI tool called a Vision Transformer (ViT).

• The Analogy: Imagine you have a stack of 1,000 photos of a storm. A normal camera just blurs them all together to show "rain." The Transformer is like a super-smart detective that looks at every single drop of rain in every single photo, notices the pattern of how they fall, and figures out exactly how strong the wind was on that specific day.

They used this AI to look at the brain waves from every single trial (every single time the person clicked the mouse) to see if the brain was "thinking" about the time or the error before the person even finished the task.

---

The Big Discoveries

Here is what they found, broken down simply:

1. The "Driving" Signal (First-Order)

When the researchers tried to predict if the person would stop the timer too early, too late, or just right, they found the answer in any single type of brain wave (Theta, Alpha, or Beta).

• Analogy: It's like checking the car's speed. You can tell how fast it's going just by looking at the engine noise, or just by looking at the wheels, or just by the wind noise. You don't need all three to know the speed.
• Result: The brain's "driving" signal is simple and can be found in just one frequency band.

2. The "Self-Check" Signal (Metacognition)

This is where it gets interesting. When the researchers tried to predict if the person would accurately realize they made a mistake, they found that no single brain wave worked.

• The Finding: The AI could only predict the "self-check" if it looked at all three brain waves (Theta + Alpha + Beta) at the same time and combined them.
• Analogy: This is like trying to diagnose a complex car problem. You can't just listen to the engine. You can't just look at the tires. You need to combine the engine sound, the tire grip, and the fuel gauge all at once to figure out, "Ah, the transmission is slipping."
• Result: The brain needs to mix different types of signals together to create a "metacognitive" thought. This proves it's a complex, separate process, not just a simple readout.

3. The "Who is Better at Self-Checking?" Test

The researchers split the participants into two groups:

• Group A: People whose brain waves were easy for the AI to decode (the AI could clearly see what they were thinking).
• Group B: People whose brain waves were messy and hard to decode.

The Surprise:

• If you looked at the "Driving" group (Group A vs. B), there was no difference in how well they realized their mistakes. Being good at "driving" didn't make you good at "self-checking."
• BUT, if you looked at the "Self-Check" group, the people whose brains showed a clear "metacognitive signal" were significantly better at realizing their mistakes.

What this means: The ability to know you made a mistake isn't just a byproduct of doing the task well. It comes from a separate, higher-level system in the brain that actively combines different signals to evaluate performance.

---

The Conclusion: The Brain Has a "Second Computer"

This paper provides strong evidence for the Higher-Order Representation Theory.

• Old Idea: Metacognition is just the brain looking at its own work and saying, "Oh, that was fast."
• New Evidence: Metacognition is the brain running a complex, multi-step analysis program. It takes signals from different parts of the brain (like different instruments in an orchestra), mixes them together, and creates a brand new "thought" about the performance.

The Takeaway:
Your brain doesn't just "read" your actions. It has a dedicated, sophisticated manager that constantly analyzes your performance by combining different types of data. This manager is what allows you to say, "I messed up that timing, and I know exactly why," even without anyone telling you. It's a separate, powerful layer of intelligence that sits on top of your basic actions.

Technical Summary

1. Problem Statement

The central question addressed is whether metacognitive evaluation (the ability to monitor one's own performance) is a direct readout of first-order performance signals (Single-Process Theory) or a computationally independent process requiring distinct neural mechanisms (Higher-Order Representation or HOR Theory).

• Context: While previous studies have linked specific brain oscillations (e.g., beta-band) to both timing accuracy and metacognition, these findings were based on condition-averaged data. It remains unknown whether these signals are shared or dissociable at the single-trial level.
• Gap: There is a lack of direct neural evidence distinguishing whether the brain forms a second-order representation of motor actions separate from the first-order signals controlling the action itself.
• Hypothesis: The authors test if metacognitive inference (MI) requires the integration of multiple frequency bands and distinct neural dynamics compared to first-order time production (TP).

2. Methodology

Experimental Design

• Task: Participants performed a Time Production (TP) task, reproducing a 2-second interval using a mouse. Immediately after, they engaged in Temporal Error Monitoring (TEM), estimating the direction/magnitude of their error and rating their confidence, all without external feedback.
• Participants: 29 healthy adults (28 included in final analysis after excluding one for excessive alpha activity).
• Data Acquisition: 64-channel EEG recorded at 500 Hz. Data was drawn exclusively from the "EEG-only" block of a Sham session to avoid tDCS aftereffects.
• Behavioral Metrics:
  • TP: Categorized into "Short," "Correct," or "Long" based on tertiles.
  • Metacognitive Inference (MI): Quantified as the alignment between the subjective error estimate and the actual TP error.
  • Confidence: Rated on a continuous scale.

Data Preprocessing & Feature Engineering

• Denoising: Standard pipeline including re-referencing, notch filtering (50 Hz), bandpass filtering (1–120 Hz), and ICA for artifact removal (EOG, EMG, ECG).
• PCA Optimization:
  • Signals were frequency-averaged across $\theta$ (4–7 Hz), $\alpha$ (8–12 Hz), and $\beta$ (13–30 Hz) bands.
  • Principal Component Analysis (PCA) was applied to extract dominant temporal patterns.
  • Trial/Channel Selection: Only the top 30% of trials and top 10% of channels (based on correlation with the accuracy-maximizing principal component, which resembled $\beta$-rebound dynamics) were retained to maximize signal-to-noise ratio.

The EEG-ViT Model (Vision Transformer)

The authors adapted a Vision Transformer architecture for 1D EEG time-series data:

• Input Representation: Time-series data segmented into non-overlapping temporal patches ($P_t = 30$ points, $N=4$ patches per trial).
• Embedding: Patches were linearly projected into 64-dimensional embeddings. A learnable [CLS] token was prepended to aggregate global trial information.
• Architecture:
  • 6 Transformer Encoder layers.
  • Multi-Head Self-Attention (6 heads) to capture long-range temporal dependencies and interactions between patches.
  • Feed-forward networks (MLP) with residual connections and layer normalization.
• Training: Stratified 5-fold cross-validation within participants to prevent data leakage and handle class imbalance.

3. Key Contributions

1. Novel Architecture: Introduction of EEG-ViT, a Vision Transformer adapted for single-trial EEG decoding, capable of capturing both local temporal dynamics and global integration without fixed spatial topologies.
2. Single-Trial Decoding: Successful decoding of first-order (TP) and second-order (MI) states from single trials before the action was completed (pre-termination).
3. Theoretical Adjudication: Provided direct single-trial evidence supporting the Higher-Order Representation (HOR) theory over Single-Process accounts.
4. Frequency Integration Discovery: Demonstrated that while first-order timing can be decoded from single frequency bands, metacognitive inference strictly requires cross-frequency integration ($\theta + \alpha + \beta$).

4. Key Results

A. Decoding First-Order Timing (TP)

• Performance: TP categories (Short/Correct/Long) were decodable significantly above chance (33%) using individual frequency bands ($\theta, \alpha, \beta$) and the broadband combination.
• Frequency Specificity: $\beta$-band activity showed the strongest predictive power, followed by $\theta$ and $\alpha$. Combining bands did not significantly improve accuracy over $\beta$ alone, suggesting redundant or sufficient information within single bands for first-order tasks.
• Timing: Decodable signals emerged well before interval termination (within 1.2s post-onset).

B. Decoding Metacognitive Inference (MI)

• Performance: Reliable decoding of MI (alignment between subjective error and objective performance) was only achieved when integrating all three frequency bands ($\theta + \alpha + \beta$).
• Frequency Specificity: No single frequency band ($\theta, \alpha,$ or $\beta$ alone) yielded significant decoding accuracy for MI. This indicates that metacognitive inference is not localized to a single oscillatory mechanism but relies on the simultaneous integration of multi-band signals.
• Timing: MI signals were detectable pre-termination but peaked during the explicit error evaluation phase.

C. Behavioral Correlates (The "Dissociation")

• TP Decoding vs. Behavior: Participants with high neural decoding accuracy for TP did not show stronger behavioral error monitoring. This contradicts Single-Process theories which predict a correlation between first-order signal fidelity and metacognitive precision.
• MI Decoding vs. Behavior: Participants with high neural decoding accuracy for MI showed significantly stronger behavioral error monitoring (higher correlation between actual and reported errors) and reduced Metacognitive Distance (MD).
• Conclusion: The precision of metacognitive evaluation is predicted by the brain's ability to decode metacognitive states (second-order), not by the fidelity of the performance signals (first-order).

D. Spatial Topography

• TP: Localized to centro-parietal and frontal-central regions (consistent with motor execution and $\beta$-rebound).
• MI: Shifted to a distributed parieto-frontal network (POz, P6, AFz, etc.), suggesting the integration of temporal representations with prefrontal evaluative processes.

5. Significance and Implications

• Computational Independence: The study provides robust evidence that metacognition is a computationally distinct process requiring higher-order multi-band neural integration, rather than a simple readout of first-order decision variables.
• Continuous Monitoring: The detection of MI signals before action termination challenges post-decisional monitoring models, supporting continuous monitoring frameworks where self-evaluation accumulates in parallel with action execution.
• Methodological Advance: The use of Transformers (ViT) on PCA-optimized single-trial EEG demonstrates a powerful new approach for decoding complex cognitive states that traditional linear classifiers or averaged ERP methods might miss.
• Clinical Relevance: Understanding the distinct neural architecture of metacognition could inform interventions for neuropsychiatric conditions characterized by metacognitive deficits (e.g., schizophrenia, anxiety), suggesting targets for disrupting or enhancing specific cross-frequency coupling mechanisms.

In summary, the paper establishes that the brain constructs a "second-order" representation of its own actions through the dynamic integration of $\theta, \alpha,$ and $\beta$ oscillations, a process that is functionally and computationally separate from the mechanisms driving the actions themselves.