Sensor-Level EEG Residual Prediction of Temporal Reproduction Bias

Question

Incremental neural information beyond behaviour

The analysis reconstructed 19,419 verified target-reproduction trials from 27 participants. The behavioural model was treated as the baseline rather than a weak comparator, so EEG was evaluated only on the residual error remaining after task structure and recent experience were accounted for.

Methods

Validation-first residual prediction

• Signed logarithmic reproduction error was modelled from target duration, task context, range condition, trial order, and previous-trial covariates.
• Sensor-level theta, alpha, beta, and low-gamma features were evaluated for incremental prediction of behavioural residuals.
• Evaluation used leave-one-subject-out cross-validation, nested ridge regularization, feature-group ablations, shadow EEG controls, expanded raw-EEG-derived feature families, empirical null controls, and subject-level inference.

Key result

A small calibrated effect, not a general EEG biomarker

Strict train-only EEG standardization did not improve held-out prediction. Unsupervised within-subject EEG calibration reduced RMSE from 0.211012 to 0.210683 (ΔRMSE = 3.29 × 10⁻⁴; 0.156%; permutation p = 0.010) and exceeded the corresponding within-subject shadow-control model.

The improvement was calibration-dependent, heterogeneous across participants, and did not persist as a robust participant-generalizable signal in the expanded raw-EEG analysis. Across broader model families and empirical null controls, EEG added negligible predictive information beyond duration, context, and trial history.