Log power did not significantly improve shrinkage-LDA AUC over linear power in eight SEEG subjects

We compared log power, linear power, amplitude, and rank-normal band-power features for offline cued key-press versus rest classification in eight SEEG subjects from OpenNeuro ds004473. A shrinkage LDA was trained on trials 1–100 and evaluated on trials 101 onward. For the primary 70–150 Hz plus 13–30 Hz feature set, log power did not significantly improve AUC over linear power (mean difference +0.0047, 6 of 8 subjects, one-sided exact sign-flip p = 0.137). For 70–150 Hz alone, log power was higher in 7 of 8 subjects (mean +0.0079, raw p = 0.023; Holm-adjusted p = 0.070). Amplitude and rank-normal features did not differ significantly from log power. Linear power was more right-skewed (median absolute skewness 1.43 versus 0.25 for log), but observed AUC changes were small: at most 0.020 for the combined high-gamma and beta features and 0.031 for high gamma alone. This offline analysis covers one dataset, one task, one session per subject, and eight subjects. No non-inferiority margin was pre-registered, so the results do not establish equivalence.

SEEG · band power · shrinkage LDA · AUC · key-press detection · log transform · rank-normal transform · OpenNeuro