Source-linked AI summary
Decoding Decision Correctness from EEG Under High Cognitive Workload in Virtual Reality: Implications for Collaborative Brain-Computer Interface Teams
Christopher Baker, Stephen Hinton, Tom Reed, Stephen Fairclough
TL;DR
Existing cBCI aggregation approaches mainly use post-hoc behavioural evidence, motivating a pre-emptive neural signal of decision correctness. The study evaluates motor-safe spatial-covariance EEG decoding in a continuous VR task under different workloads and tests its value for team aggregation. The signal improved contested-trial accuracy under High Workload but was unreliable and detrimental under Low Workload, defining a workload-conditional deployment boundary.
Problem
Existing team-aggregation approaches rely on post-hoc behavioural evidence, whereas cBCIs need signals that can inform decisions before responses are committed.
Method
The study used a continuous VR target-detection task with within-subject workload manipulation and decoded correctness from motor-safe EEG spatial-covariance features for simulated team aggregation.
Results
Under High Workload, the classifier decoded decision correctness above chance at 61.4%, and neural weighting increased contested-trial team accuracy from 57.2% at team size 2 to 88.2% at team size 16.
Takeaways & Limitations
EEG-based decision-reliability signals support team augmentation selectively under high cognitive workload rather than uniformly across conditions.
Takeaways & Limitations
The Low Workload result is constrained by too few error trials, so absence of a decodable signal there cannot distinguish missing neural evidence from insufficient data.
Abstract
from arXiv · showhide
Collaborative Brain-Computer Interfaces (cBCIs) offer a promising mechanism to augment team decision-making, but existing approaches rely exclusively on evidence available only after a decision has been made and reported, such as reaction time or stated confidence. This limits their use to explaining or discounting a decision after the fact, rather than informing a team's response before it is finalised. We tested whether spatial-covariance EEG features could instead provide a genuinely pre-emptive signal of an operator's decision correctness, available within the response window itself, and whether such a signal depends on cognitive workload. Using a continuous virtual reality target-detection task, participants (N = 23) completed a within-subject workload manipulation (High vs. Low). At the team level, weighting votes by this pre-emptive neural signal, available before a response is committed, produced substantial accuracy gains on contested (evenly-split) trials under High Workload (57% to 88% as team size increased from 2 to 16), but was actively detrimental under Low Workload. Critically, this advantage held even against post-hoc behavioural signals: confidence was the strongest single team-level signal overall, but by definition cannot inform a decision still in progress, whereas the neural signal can. These findings indicate that EEG-based decision-reliability signals are not a general-purpose team augmentation tool, but a workload-conditional one, with clear implications for when and how cBCI systems should be deployed in operational teams.
Introduction
Operational teams must make rapid perceptual decisions under demanding conditions, yet workload can degrade individual reliability and undermine conventional aggregation based on behavioural outputs. This study tests whether a pre-emptive EEG signal of decision correctness can support workload-conditional team aggregation.
- High cognitive workload can degrade perceptual reliability, challenging team aggregation methods that assume behavioural outputs remain trustworthy proxies for perceptual state.
- Neural markers of decision correctness can be detected independently of, and sometimes before, operators’ explicit awareness or report.
- cBCIs could give team aggregation access to neural information that behavioural channels may not fully reveal.
- The study tests a motor-safe Riemannian EEG classifier, workload dependence, likely neural origin, and team-level value against confidence and reaction time.
Methods
Participants performed a continuous VR target-detection task under within-subject workload manipulation while EEG was recorded and processed using motor-safe spatial-covariance features. Simulated teams compared unweighted and evidence-weighted aggregation using neural, confidence, and response-time signals.
- Study Design and VR Target-Detection Task: Twenty-three participants completed a within-subject repeated-measures VR target-detection task using a drone-view simulation and joystick responses.
- Study Design and VR Target-Detection Task: The task presented continuously appearing 3D Targets and Non-Targets, with a reticle locking onto each stimulus for 2500 ms.
- Environmental Workload Manipulations: High Workload reduced ambient lighting by 50% and lowered the solar angle relative to standard-daylight Low Workload sessions.
- Data Acquisition and EEG Signal Processing: EEG was recorded at 500 Hz using a 32-channel LiveAmp system, synchronised with behavioural and VR events via LabStreamingLayer.
- Motor-Safe Feature Extraction and Classification: Features were extracted from 300 ms after reticle onset until 100 ms before each trial’s response, excluding windows shorter than 150 ms.
- Team Simulation and Aggregation: Team simulations used bootstrap resampling across team sizes 2, 4, 8, and 16, comparing unweighted majority votes with neural, confidence, and response-time weighting.
- Statistical Analysis: Classifier accuracy was tested against chance, reticle responses with cluster-based permutation testing, and channel importance through covariance ablation.
Results
The workload manipulation reduced accuracy and slowed responses, while EEG decoding of decision correctness was robust only under High Workload. Pre-emptive neural weighting improved accuracy on contested High-Workload trials but harmed performance under Low Workload.
- Workload Manipulation and Behavioural Asymmetry: 77.55% vs. 93.84% accuracy and 1.40s vs. 0.81s response time confirmed lower performance under High than Low Workload.Both workload differences were significant at p < .0001.
- Workload Manipulation and Behavioural Asymmetry: Target detection was significantly harder than Non-Target rejection across workload conditions (p = .007).This asymmetry also recurred in the analysis of trial difficulty.
- EEG Decoding: 61.4% classifier accuracy decoded Correct from Incorrect decisions above chance under High Workload, with TPR 64.8% and TNR 46.3%.Leave-one-subject-out validation across N = 23 participants yielded t(22) = 3.65, p = .0014.
- EEG Decoding: Occipital and parietal channels, especially Oz, O2, and P3, produced the largest accuracy drops when ablated.The posterior-weighted profile was characterized as consistent with a perceptual rather than motor origin.
- Trial Difficulty: Target trials had significantly lower vote margins than Non-Target trials under both High Workload (t = -5.63, p < .0001) and Low Workload (t = -4.78, p < .0001).Lower margins indicate more contested trials.
- Team-Level Aggregation: 57.2% to 88.2%: pre-emptive neural weighting increased High-Workload team accuracy from team size 2 to 16 on true-tie trials.Unweighted voting and average individual accuracy remained approximately 50%; under Low Workload, the same weighting collapsed accuracy toward 10% by team size 8.
- Post-Hoc Behavioural Signals: 99.4% team accuracy at size 16 under High Workload made confidence-weighting the strongest single tested signal, but it was post-hoc.Under Low Workload, confidence-weighting reached 94.3% at team size 8; response-time weighting underperformed and declined with team size.
Discussion
The study identifies a pre-emptive EEG signal of decision correctness whose team-level value depends on workload. It improves contested-trial accuracy under High Workload but is unreliable or harmful under Low Workload, with posterior channels and trial difficulty helping delimit its operational relevance.
- Limitations: Low Workload decoding was unreliable because near-ceiling performance produced too few genuine errors for learning, leaving the absence of a signal unresolved.A more balanced low-workload task would be needed to distinguish a genuine absence from insufficient data.
- The Value of a Pre-emptive Signal: Confidence outperformed the neural signal in absolute team-level accuracy, but only the neural signal could inform a decision before response finalisation.Confidence reached near-ceiling contested-trial accuracy in both workload conditions, whereas the neural signal was actionable during the response window.
- Mechanistic Origin and Its Implications: Posterior occipito-parietal channels—Oz, O2, and P3—carried most decoded information, supporting a perceptual or evidence-accumulation interpretation rather than residual motor preparation.Features came from a motor-safe window ending 100 ms before each trial’s response, but the channel-level interpretation remains correlational and not source-localised.
- Trial Difficulty and the Boundary Conditions of the Signal: Contested trials concentrated the practical value of neural weighting because team disagreement occurred where Target detection was intrinsically harder than Non-Target rejection.This bounds the signal’s operational relevance to specific decision classes rather than all team decisions.
- Limitations: Team-level findings came from bootstrap-resampled simulated teams using one task and one session per participant, limiting generalisation beyond the study setting.The authors caution that broader evidence is needed before generalising these team-level results.
Data Availability
The full study dataset is too large for standard manuscript-linked repositories and is maintained in a version-controlled GitLab repository, with access available on reasonable request.
- Data Availability: The complete dataset exceeds 20 GB, making standard repositories such as Figshare and OSF impracticable for hosting it.The dataset includes raw and preprocessed EEG recordings, behavioural logs, and analysis code.
- Data Availability: The complete dataset is maintained in a version-controlled GitLab repository.
- Data Availability: Qualified researchers may obtain access from the corresponding author upon reasonable request for verification purposes.
Code Availability
The study’s custom analysis code is organised in Marimo notebooks covering EEG feature extraction, Riemannian classification, and team aggregation, and is available from a sister GitLab repository on request.
- Code Availability: Custom Python code implements motor-safe spatial-covariance feature extraction, Riemannian tangent-space classification, and team aggregation simulation.
- Code Availability: The code is organised as Marimo notebooks spanning all three studies in the research programme.
- Code Availability: The notebooks are hosted in a sister GitLab repository and are available on reasonable request, subject to Dstl approval.