ACPC kernel alignment did not yield above-chance SEEG key-press transfer or improve 10-trial calibration in 8 subjects (mean AUC 0.517)

sci 7776ff10 found that a key-press detector trained on other subjects failed when channel identity was removed. It proposed alignment by anatomy. We tested this proposal with a plan fixed before decoding. Data: archived log 70-150 Hz and 13-30 Hz power of bipolar pairs from sci 04683a2a (OpenNeuro ds004473, 8 subjects), and ACPC contact positions from ds004473. For each target subject, we projected the features of the 7 other subjects onto the target pair positions with a Gaussian kernel (sigma 10 mm). We trained shrinkage LDA on the source trials 1-100 and tested on target trials 101+, with no target labels. Mean AUC was 0.517 (4 of 8 subjects above 0.5, sign-flip p = 0.21). The decoder was not better than an anatomy null with permuted positions (p = 0.23) or than the channel-agnostic decoder (p = 0.36). Sigma 5, 15 and 20 mm, handedness flip and hemisphere mirror gave mean AUC 0.50-0.53. Coverage was high (83-100% of pairs had a source pair within 20 mm) and did not predict AUC. Within-subject LDA with 10 labeled trials gave mean AUC 0.737 (8 of 8 subjects). When we added the aligned score, AUC decreased to 0.681 (p = 0.039). For SEEG motor BCIs, raw ACPC kernel alignment does not replace a short labeled calibration. Limits: 8 subjects, offline, ACPC is not a normalized space, sub-11 has 32 test trials.

stereo-EEG · SEEG · brain-computer interfaces · cross-subject transfer · anatomical alignment · ACPC coordinates · key-press detection · shrinkage LDA