Per-shaft SEEG score fusion was non-inferior to central LDA in an offline eight-subject task and sent 13.6 times fewer values

In eight subjects performing a cued key-press-versus-rest task, per-shaft shrinkage linear discriminant analysis (LDA) followed by a learned second-stage LDA was non-inferior to one central LDA using all 70–150 Hz high-gamma and 13–30 Hz beta bipolar log-power features. Models trained on trials 1–100 and were tested on trials 101 and later. Mean test AUC was 0.840 for per-shaft fusion and 0.828 for central LDA (mean difference +0.012; six of eight subjects higher; non-inferiority margin 0.02; one-sided exact sign-flip p=0.0039). A two-sided test gave p=0.047, which is weak evidence because superiority was not the primary endpoint. Per-shaft fusion sent a mean 15.4 values per window instead of 208.5, a 13.6-fold reduction. A plain sum of shaft scores did not differ from central LDA (mean difference −0.004, p=0.58) and reduced AUC by 0.032 in one subject. Under simulated loss of one shaft after calibration, mean AUC was 0.013 higher for per-shaft fusion (p=0.039, uncorrected); for the worst lost shaft the difference was +0.023 (p=0.086). The drop from intact performance to worst-shaft loss was similar for both designs, so per-shaft fusion did not remove the risk from one key shaft. These results apply to this offline task, eight subjects, one session per subject, one train/test split, and these features; they do not establish closed-loop performance.

SEEG · brain-computer interface · linear discriminant analysis · score fusion · feature bandwidth · non-inferiority · shaft loss · offline decoding