In 18 SEEG patients, median AUC rose from 0.535 to 0.549 with 1–17 training patients, with no detected gain from 8 to 17

We measured how the accuracy of a zero-calibration SEEG memory decoder changes with the number of training patients. We used the archived per-word band-power features of sci c4b8c884 (OpenNeuro ds004789 v3.1.0, FR1 free recall, session 0, 18 included SEEG patients). The decoder is the c4b8c884 cross-patient model: 8 channel-mean log-power features (3-140 Hz), L2 logistic regression. For each held-out patient, we trained on 100 random subsets of N = 1, 2, 3, 4, 6, 8, 12 other patients, and on all 17. The plan was written before the analysis. The median held-out AUC increased from 0.535 (N = 1) to 0.549 (N = 17); paired difference +0.021, 13/18 patients, one-sided exact Wilcoxon p = 0.022. There was no difference between N = 8 and N = 17 (median +0.0001, p = 0.64). The decoder was above chance at every N (Holm p <= 0.0017). The SD of AUC across training subsets decreased from 0.058 to 0.020, but about 20% of decoders stayed below chance at N >= 8, because some patients do not fit the shared model. The trained decoder did not become better than a fixed theta/high-gamma contrast (median 0.569). A saturating-curve extrapolation was not identifiable (A(100) = 0.558, bootstrap 95% interval 0.50-0.63). Conclusion, for this dataset and model: differences between patients, not the number of training patients, limit this cold-start decoder. Better shared features or short patient calibration are necessary. The package contains code, input features with checksums, the plan, all results and a figure.

stereo-EEG · word recall · brain-computer interface · cross-patient decoding · learning curve · area under the ROC curve