Clock-time difference did not reliably predict SEEG decoder AUC across sessions

We tested whether the clock-time difference between training and test sessions predicted cross-session SEEG decoder AUC or a label-free feature shift beyond elapsed time. We analyzed archived late-maintenance log band-power features from 10 subjects in OpenNeuro ds004752 (verbal Sternberg working-memory task); 8 subjects had other-day pairs with at least a 2 h range in clock difference. The planned primary analysis used 208 ordered other-day pairs (16–72 h apart), per-session z-scored features, and per-subject regression of AUC on log2(1 + elapsed hours) and circular clock-time difference (0–12 h). The mean clock-difference slope was 0.147 AUC/h (95% bootstrap CI 0.001 to 0.432; exact sign-flip p = 0.10). Clock-time difference and elapsed time were strongly coupled (median within-subject |r| = 0.77); in one subject r = −1.00, making its primary slope of 1.14 AUC/h meaningless and driving the large mean. Without an elapsed-time covariate, the z-scored mean slope was −0.0066 AUC/h (95% CI −0.0185 to 0.0043; p = 0.36). Raw features with the log2(1 + elapsed hours) covariate produced positive slopes (p = 0.008), opposite to the hypothesis; the result was absent without that covariate (p = 0.52) and with a day-gap covariate (p = 0.27), and is treated as a likely model suppression effect rather than a clock-time effect. Label-free feature-shift tests were not significant (p = 0.06–0.89). We did not correct 14 summaries for multiple testing. The no-covariate confidence interval permits a loss of about 0.11 AUC or gain of about 0.03 AUC for a 6 h difference, so a moderate effect cannot be excluded. This dataset does not show that clock-time difference reliably predicts AUC or feature shift beyond elapsed time. Its session schedule couples clock time and elapsed time; testing circadian effects requires sessions at planned clock times. The package preserves the analysis code, archived inputs, provenance hashes, per-pair results, and figure.

SEEG · brain-computer interface · decoder transfer · clock time · circadian effects · cross-session generalization · working memory · AUC · feature shift