A leave-one-patient-out fsaverage map improves eight-channel SEEG movement decoding over random selection in 10 patients

We evaluated whether a spatial map made from nine other patients can select eight stereo-electroencephalography (SEEG) channels for decoding mouse movement before channel calibration. In 10 patients from OpenNeuro ds006253, we mapped the mean 70–150 Hz movement effect across fsaverage coordinates with a 10 mm Gaussian kernel, then selected eight channels without using the held-out patient’s labels. A shrinkage linear discriminant analysis trained on the first 60% of trials and tested on the last 40% gave mean held-out AUC 0.706 for map-selected channels and 0.642 for 1,000 random eight-channel sets (difference +0.065; 7/10 patients positive; one-sided exact sign-flip p = 0.021). The map value correlated with the held-out patient’s TRAIN channel effect (mean Spearman rho 0.157; 8/10 positive; Holm p = 0.012), but not significantly with its TEST channel effect (mean rho 0.084; 7/10 positive; Holm p = 0.077). Selection using the patient’s own training labels gave AUC 0.791. The result was stable at 5 mm (+0.064, p = 0.026) and 20 mm (+0.093, p = 0.003); distance to a hand-knob point did not predict channel effect (mean rho 0.075, p = 0.105). The gain closes about one third of the gap between random eight-channel selection and all channels (AUC 0.811), and does not replace calibration. The study includes only 10 patients and one dot-motion task. REST preceded the stimulus and MOVE occurred during mouse movement, so decoding may use visual and decision activity as well as motor activity. Hand and handedness were unknown; hemispheres were not mirrored. Coordinates are approximate, their stated unit is inconsistent with their values, and the map-selected decoder still uses the patient’s labels to train its LDA weights. Results are offline; closed-loop control and stimulation were not tested.

stereo-electroencephalography · channel selection · brain-computer interface · movement decoding · high-gamma power · leave-one-patient-out · linear discriminant analysis