In five SEEG patients, random folds decoded staircase coherence from pre-stimulus activity, but blocked detrending reduced scores near zero

An adaptive staircase changed motion coherence over session time in OpenNeuro ds006253. We tested whether this drift could produce a false coherence decoder from 70–150 Hz SEEG activity. We used archived 50-ms stimulus-locked epochs and behavior tables for five patients (sub-02, 04, 12, 13, and 23), and a plan fixed before analysis. Ridge regression predicted z-scored coherence from mean channel activity in pre-stimulus (−0.5 to 0 s) and post-stimulus (0 to 0.5 s) windows. Random 10-fold cross-validation produced positive held-out correlations in all five patients for both windows: median r = 0.314 before the stimulus and 0.167 after it. Pre-stimulus activity also predicted trial number (median r = 0.525), consistent with slow signal drift. With ten contiguous folds and quadratic trend removal fitted within each training fold, median scores were 0.014 before and −0.027 after the stimulus. Exact one-sided sign-flip p-values were 0.031 for both random-fold tests and 0.281 and 0.438 for the blocked tests; after Holm correction across four tests, no result was significant (adjusted p = 0.125 and 0.563, respectively). In sub-12, pre-stimulus blocked r remained 0.52 because coherence drift was non-monotonic (r with trial number = 0.15), showing that quadratic detrending may not remove irregular drift. These results show that random-fold scores are not evidence of coherence coding separate from drift in this sample. Blocked detrending can also remove true coding because coherence covaries with trial number. The study includes five patients, one window per period, one model, clinical channels without anatomical selection, and offline analysis only. The smallest possible exact p-value with five patients is 1/32; no test was significant after Holm correction. The null blocked results do not prove that coherence coding is absent. BCI studies that decode staircase or session-varying variables should include blocked folds and a pre-stimulus control window.

SEEG · high-gamma activity · motion coherence · adaptive staircase · cross-validation · temporal drift · ridge regression · brain-computer interface · signal decoding