feature parameters

We evaluated seven eye‐movement features before and after fatigue using paired t‐tests on mean ± SD values. Sample Entropy on the X-axis decreased from 0.095 ± 0.054 to 0.080 ± 0.048 (p = 0.025) and on the Y-axis from 0.106 ± 0.046 to 0.090 ± 0.043 (p = 0.025), whereas Approximate Entropy and Cross Entropy showed no significant change. Pearson correlation fell sharply from 0.049 ± 0.227 to –0.062 ± 0.201 (p = 0.002), highlighting a pronounced breakdown of linear coupling across axes. Coherence exhibited no clear trend, but NLID(X→Y) declined from 0.736 ± 0.185 to 0.654 ± 0.182 (p = 0.013) and NLID(Y→X) from 0.747 ± 0.151 to 0.698 ± 0.155 (p = 0.030), both statistically significant. Overall, bi‐axial features (Correlation and NLID) proved more sensitive to visual fatigue than single‐axis entropy, markedly enhancing fatigue detection sensitivity; Appendix A provides detailed statistics across all embedding dimensions and delays.