SEEG decoder margins showed lag-1 dependence, but fixed-threshold errors did not cluster

In archived features from eight subjects performing a cued SEEG key-press task, we tested whether outputs of a shrinkage linear discriminant decoder depended on trial order. The decoder was trained on trials 1–100 and evaluated on trials 101 onward (32–198 trials per subject). The primary measure was the lag-1 autocorrelation of each trial's move-minus-rest score margin, less the mean from 10,000 trial-order permutations. Mean excess autocorrelation was +0.118 (median +0.155), positive in 7 of 8 subjects (exact one-sided sign-flip p = 0.035; Stouffer z = 3.56). Mean margin autocorrelation was 0.108 at lag 1, 0.020 at lag 2, and near the permutation mean at lags 3–10. Linear detrending and five-fold time-blocked cross-validation gave similar positive results; an 11-trial running-median detrend did not (mean excess +0.051, 6 of 8, p = 0.16). At the fixed threshold, miss and false-alarm errors did not cluster. Accuracy for a three-trial majority vote over consecutive trials (0.877) was close to random triplets (0.885) and the independence prediction (0.883); all vote comparisons had p > 0.25. Thus margin scores had weak, short-lag serial dependence, but this did not produce error bursts or reduce majority-vote accuracy in this task. The study was offline and used one cued task and session, eight epilepsy patients with clinical electrode placement, trial-locked windows about 7 s apart, and one linear decoder. Subject 12 showed the opposite sign, and secondary tests were not corrected.

SEEG · brain-computer interface · serial dependence · decoder margin · shrinkage LDA · error bursts · majority vote · OpenNeuro ds004473