In 15 SEEG patients, shaft-mean pooling reduced movement-detection AUC by 0.094

In an offline analysis of 15 patients with stereo-electroencephalography (SEEG), we compared movement-versus-rest decoding from all bipolar pairs, the mean feature per shaft, and the best training-selected pair per shaft. Features were log10 mean high-gamma activity (70–150 Hz); the first 60% of trials trained a shrinkage linear discriminant and the last 40% tested it. Mean held-out AUC was 0.833 for all pairs, 0.739 for shaft means, and 0.797 for selected pairs. Shaft means were 0.094 AUC below all pairs (95% CI −0.127 to −0.063; 14 of 15 patients lower) and failed the 0.02 non-inferiority test (p = 0.9997). They were also 0.058 below selected pairs (95% CI −0.085 to −0.033; p = 0.0005). Selected pairs were 0.036 below all pairs in an exploratory comparison (95% CI −0.051 to −0.021; p = 0.0006). These offline, single-session results indicate that averaging contacts along each shaft reduced movement information in this dataset. For feature-limited decoding in this setting, selecting a pair per shaft retained more information than averaging. The analysis used one event-locked 0.5 s window per class, one frequency band, a linear decoder, clinically placed electrodes, and a shaft definition based on contact-name prefixes; pair selection used training data from one session.

stereo-electroencephalography · SEEG · brain-computer interface · movement decoding · high-gamma activity · feature pooling · linear discriminant analysis · AUC