Across 15 SEEG subjects, SPREAD gains tracked the chance of sampling a movement-informative pair

We tested whether the per-subject AUC gain of K bipolar pairs on K distinct shafts (SPREAD) over K adjacent pairs on one shaft (DENSE) followed the chance that a set included a pair with movement signal or the noise correlation of adjacent pairs. We analyzed archived 70–150 Hz features from 15 subjects in two OpenNeuro tasks, using pre-registered random set draws and shrinkage LDA. At the primary K = 4, the gain correlated with the SPREAD–DENSE hit-probability gap (Spearman rho = 0.70, Holm p = 0.0053), while adjacent-pair noise correlation did not predict the gain (rho = 0.22, Holm p = 0.21). At K = 8, the gain also correlated with the hit gap (rho = 0.66, p = 0.005). Sets containing an informative pair had higher AUC than sets without one (+0.102 at K = 4; 13 of 13 subjects), but when both set types contained one, SPREAD was not better (-0.009 AUC at K = 4; -0.054 at K = 8, p = 0.07, uncorrected). The result suggests that spreading pairs helps when the informative shaft is unknown by increasing the chance of sampling it; it did not improve a set that already included an informative pair. The study was offline, used two tasks and 15 subjects, and evaluated random placement on clinical implants. The hit-gap and AUC gain used the same set choices, so their association is not an independent prediction. Dataset release identity was not established for the archived features.

stereo-EEG · brain-computer interface · high-gamma activity · feature selection · movement decoding · shrinkage LDA · signal sampling