Fig. 6: Sensitivity analysis links fatigue state to frequency-specific perceptual weighting.

A Analysis pipeline. A feed-forward neural network (25 inputs = 5 bands × 5 electrodes; one hidden layer; single VCT output) was trained separately on low- and high-fatigue trials from the sham condition. From the trained weights we formed the sensitivity matrix S and obtained band-wise contributions by eigen-decomposition. A permutation test (10000 shuffles of fatigue labels) assessed whether the observed Low–High difference for each band exceeded chance. B Learning curves for the two networks (8:2 train/validation split, early stopping). C Eigen spectrum of S (left) and cumulative variance (right). The first seven eigen-vectors (vertical line) explain ≥ 90% of the variance and were used to compute contributions. D Normalized contributions of the five bands. Blue = Low-fatigue, red = High-fatigue. A dashed box highlights that the Low–High gap for Alpha (ΔAlpha) greatly exceeds that for Beta (ΔBeta). E Permutation distributions of Δ = Low–High for each band. Red line = observed Δ. Two-tailed p-values (FDR-corrected) appear above each panel. ΔDelta and ΔAlpha are significant; smaller but significant effects appear for Beta and Gamma, whereas Theta shows no state dependence.