Can EEG decode decision correctness before a response under high cognitive workload?

A signal available before the vote is cast

Collaborative brain-computer interfaces aim to improve team decisions by weighting each operator's contribution. Existing weighting schemes, however, depend on evidence that exists only after a decision has been reported: reaction time or stated confidence. That makes them useful for discounting a bad vote after the fact, not for informing the team while the decision is still forming.

This study asked whether EEG could supply something earlier. The target was not what an operator saw, but whether that operator's judgment was correct, decoded from neural activity inside the response window itself. The authors also asked whether such a signal depends on cognitive workload, since workload is known to degrade perceptual reliability and could break the assumption that behavioural output reflects the underlying perceptual state.

Motor-safe windows and spatial covariance

Twenty-three participants performed a continuous virtual reality target-detection task, discriminating targets from non-targets presented inside a reticle that stayed locked for up to 2500 ms. They pressed a joystick button before the reticle disappeared. Workload was manipulated within subject: high-workload sessions reduced ambient lighting by 50% and lowered the solar angle relative to standard-daylight low-workload sessions.

To avoid decoding movement rather than judgment, features came from a strictly motor-safe window running from 300 ms after reticle onset to 100 ms before each trial's own response time, with a 150 ms minimum; shorter trials were excluded. Within that window the study estimated a spatial covariance matrix per trial, projected it into a Euclidean tangent space via the Riemannian mean, and classified with logistic regression using leave-one-subject-out cross-validation, separately for each workload condition. This approach captures the relational structure of oscillatory activity across the electrode array rather than depending on precise temporal alignment, which matters in self-paced virtual reality where latency jitter is large.

Above-chance decoding only under high workload

Under high workload, the classifier separated correct from incorrect decisions with 61.4% accuracy, a true positive rate of 64.8% and a true negative rate of 46.3% (t(22) = 3.65, p = .0014, N = 23). The asymmetry matters: the classifier was considerably better at recognising correct trials than at catching errors.

Low workload produced a numerically higher raw accuracy of 79.4%, but the authors explicitly do not treat this as a comparable finding. The true negative rate collapsed to 12.3%, meaning the classifier almost never correctly identified an incorrect trial. The cause is a behavioural ceiling: low-workload accuracy reached 93.84%, leaving too few genuine error trials to learn a reliable decision boundary. The low-workload classifier is reported but not mechanistically interpreted, and all later analyses of classifier structure and team value focus on high workload.

A posterior, perceptual rather than motor profile

To see which channels drove classification, the authors removed each channel's contribution to the spatial covariance matrix in turn and measured the drop in leave-one-subject-out accuracy. The largest drops occurred at occipital and parietal sites (Oz, O2, P3), with central and temporal channels (C4, T8) contributing more modestly further down the ranking.

The authors read this posterior-weighted profile as consistent with a perceptual rather than motor origin for the decoded signal. They also note that this ablation ranking is more conservative than raw classifier coefficient magnitude, which was diffusely distributed across channels because of L2 regularisation and did not reliably separate informative channels from correlated, redundant ones. This is the authors' interpretation of the channel pattern, not a direct measurement of the underlying neural generator.

Weighting votes helps only under high workload

Team performance was evaluated by bootstrap resampling with 500 draws per team size and session, across team sizes of 2, 4, 8 and 16. Votes were aggregated by unweighted majority and separately by evidence-weighted vote using the pre-emptive neural signal, stated confidence, or response time.

On contested trials, where votes were evenly split and weighting could plausibly change the outcome, weighting by the neural signal raised team accuracy from 57% to 88% as team size grew from 2 to 16 under high workload. Under low workload the same weighting was actively detrimental. The signal is therefore workload-conditional rather than a general-purpose team augmentation tool.

The comparison against behavioural signals is the sharper point. Confidence was the strongest single team-level signal overall, but by definition it cannot inform a decision still in progress, whereas the neural signal is available within the response window. Response-time weighting failed, and the authors trace that failure to a workload- and trial-difficulty-dependent reversal of the speed-accuracy relationship in this task.

Where the workload-conditional claim stops

The low-workload result is the clearest boundary. Because the classifier's unreliability follows from small error-trial counts in that condition, the study cannot establish whether the absence of a decodable signal under low workload reflects a genuine absence of the underlying neural process or simply insufficient data to detect it. A task producing a more even balance of correct and incorrect trials under low workload would be needed to separate those explanations.

The response-time explanation is likewise scoped to this task's response dynamics and should not be read as a general claim that response time is uninformative elsewhere. Team-level findings come from simulated teams built by resampling recorded individual responses, not from live collaborating teams, so the 57% to 88% gain describes a simulation under the high-workload condition rather than an observed operational deployment.

Deploy the neural weight when errors are frequent

The decision-relevant split in this paper is not neural versus behavioural signals in general, but high versus low workload. Where operators are error-prone and behavioural accuracy has room to fall, the pre-emptive EEG weight improved contested-trial aggregation and offered something confidence cannot: information before the vote is committed. Where performance is near ceiling, the same weight hurt team accuracy and the classifier had too few errors to learn from.

That makes error prevalence, not workload label alone, the practical threshold to check before deploying such a system. The paper's own framing follows: EEG-based decision-reliability signals should be deployed conditionally, in settings where genuine errors are frequent enough for the signal to carry information and for the classifier to be trained on them.

Sources used in this answer