Anatomy-predicted SEEG channel weights did not detect cued key presses without labels or improve few-shot calibration in 8 subjects

We tested whether channel weights predicted from anatomy in other subjects could detect cued key presses from stereo-EEG (SEEG) without labeled calibration trials, or improve decoding with 3–20 labeled trials. We used archived 70–150 Hz bipolar power features from 8 subjects in OpenNeuro ds004473 v1.0.2, with 60 trials per subject. In leave-one-subject-out tests, a ridge model (alpha 10) predicted each channel’s move-versus-rest Cohen’s d from atlas and distance attributes. Zero-shot median AUC was 0.452 (mean 0.454); 3 of 8 subjects exceeded 0.5, and the one-sided exact sign-flip p value was 0.94. Predicted and observed channel effects had median correlation r = -0.12. The anatomy-assigned weights were not better than weights shuffled across channels (mean AUC difference -0.029, p = 0.88), and were worse than equal weights (mean difference -0.047; one-sided p for zero-shot greater than equal-weight AUC = 0.99). As a shrinkage prior, they reduced AUC relative to mean-difference weights by 0.011, 0.008, 0.004, and 0.002 for 3, 5, 10, and 20 labeled trials, respectively (two-sided p = 0.13–0.36). Post hoc variants also showed no significant zero-shot detection. Within this offline, cued task, channel effects varied across these 8 subjects, and the anatomy prior did not provide label-free detection or reduce the need for subject-specific calibration. The study is limited to 8 subjects, one session, region-level atlas features, one linear prior model, and a cued task in which the movement window included the visual cue.

stereo-EEG · brain-computer interface · motor decoding · zero-shot learning · few-shot learning · anatomical prior · high-gamma power · cross-subject transfer