Three to four SEEG channels on different shafts improved recall decoding over channels on one shaft in 19 patients

This offline replication tested whether, for prediction of later word recall from bipolar SEEG, selecting channels on different shafts improves decoding over selecting the same number of adjacent channels on one shaft. Archived FR1 features from 19 patients in OpenNeuro ds004789 were evaluated with Ledoit–Wolf shrinkage LDA, list-grouped 5-fold cross-validation, two repeats, and 100 random draws per condition and patient. For k = 3, spreading channels increased mean AUC by 0.0174 (18/19 patients; Holm p = 1.5 × 10⁻⁵); for k = 4, it increased mean AUC by 0.0184 (17/19; Holm p = 2.7 × 10⁻⁵). At k = 2, the gain was +0.0064 (15/19; p = 0.0033), although k = 2 was not a primary hypothesis and the earlier study found no difference. Mean AUC across all channels was 0.581. Within-class noise correlation was 0.51 for adjacent pairs on one shaft, 0.25 for non-adjacent pairs on one shaft, and 0.13 for pairs on different shafts; adjacent-pair correlation exceeded different-shaft correlation in all 19 patients (p = 3.8 × 10⁻⁶). The k = 3–4 effect replicated with a smaller absolute gain. Recall decoding remained weak (AUC 0.54–0.56 with 3–4 channels). This was an offline, one-session analysis using one decoder and one feature set; random channel draws do not distinguish reduced noise correlation from broader brain sampling, and spread channels use more contacts.

stereo-electroencephalography · SEEG · brain-computer interface · word recall · channel selection · noise correlation · linear discriminant analysis · replication