Coarse region features did not improve zero-calibration cross-patient SEEG recall decoding in 18 FR1 patients

We tested if a cross-patient SEEG decoder that keeps coarse anatomical region predicts word recall in a new patient better than a channel-agnostic decoder. Data: archived encoding-period power features of 18 patients (OpenNeuro ds004789 v3.1.0, FR1, session 0) from c4b8c884. The baseline (8 channel-mean z-power features) reproduced c4b8c884 exactly: median held-out AUC 0.549, 14/18 patients above 0.5. The region decoder (48 features: MTL, lateral/ventral temporal, frontal, parietal-occipital, cingulate-insula, other) had median AUC 0.538, 16/18 above 0.5. The pre-registered paired test showed no improvement: median difference -0.006, 7/18 patients higher, exact two-sided Wilcoxon p = 0.18. Results were the same within serial position (p = 0.93) and with stronger regularization (C = 0.01, p = 0.21). An MTL-only decoder was not better than a non-MTL decoder (p = 0.42). Conclusion: at this region resolution and sample size, the transferable recall signal is a broadband brain-wide power pattern; the simple 8-feature decoder is sufficient. Assumptions: atlas region labels are correct; one session per patient; power features only; low power for differences near 0.01 AUC. Package has code, input features with MD5, results, figure and a pre-registered plan.

SEEG · word recall · cross-patient decoding · memory BCI · power features · anatomical regions