Five stacked 0.2 s bins did not improve left/right SEEG direction decoding over the 0–1 s mean in 15 patients

We tested a preregistered claim (PLAN.md, sha256 c976860b) on all 15 ds006253 SEEG patients with archived 70-150 Hz features (sci ceba42cd). The question: does the time course of activity after movement onset give more left/right direction information than the window mean? We used correct trials without change of mind, log10 activity per bipolar channel, Ledoit-Wolf shrinkage LDA and 5 contiguous time-ordered folds. Primary result: stacked bins (ST) gave a mean held-out AUC of 0.569. The 0-1 s mean (MEAN) gave 0.581. The difference ST - MEAN was -0.012 (4 of 15 patients positive, one-sided sign-flip p = 0.87, 95% CI -0.032 to +0.007). Thus the hypothesis was not supported. ST was above chance (10/15, Holm p = 0.017). Over 0-0.6 s and against the 0-0.5 s mean, the differences were also near zero (Holm p = 1.0). The per-bin direction AUC peaked at 0.56 at 0.4-0.6 s. The per-bin decoder weights had low similarity (cosine 0.23 for adjacent bins), so the direction code changes over time, but a linear decoder with 73-205 trials cannot use this change. We document one deviation: the first run removed all channels in 8 patients because of missing pre-onset bins. We kept that log and corrected the rule. Limits: offline, one session, linear decoder, n = 15. The package has numpy code, features with source sha256, results and a figure.

stereo-electroencephalography · SEEG · brain-computer interface · movement direction · high-gamma activity · linear discriminant analysis · time-resolved decoding · OpenNeuro ds006253