In 18 SEEG patients (ds004789 FR1, session 0), session z-scoring did not raise fixed-decoder AUC, while causal 48-word z-scoring lowered it

A real-time memory BCI cannot z-score features with statistics of the full session. We compared five feature normalizations for a fixed word-recall decoder. Data: archived ds004789 FR1 features (c4b8c884), session 0, log10 power at 8 frequencies on all bipolar channels. The decoder was an L2 logistic model (C = 7.2e-4), fit one time on lists 1-8 and tested on lists 9 and later. The plan was registered before any fit. 18 patients passed the inclusion rule. Median test AUC was 0.621 with a calibration-only z-score (CALIB). A non-causal session z-score did not increase AUC (median difference -0.004, 6/18 higher, p_Holm = 0.16). A causal expanding-window z-score had no significant effect (-0.004, p_Holm = 0.16). A causal 48-word sliding-window z-score decreased AUC (-0.013, 3/18 higher, p_Holm = 0.012). AUC within serial position and within list did not change significantly. With CALIB, the threshold from calibration data gave a biased flag rate on later lists (median 0.41 vs target 0.50, error 0.10). All adaptive schemes decreased this error (uncorrected p = 0.001-0.043). Per-list z-scoring gave the smallest error (0.02), but it removed most list-level information (list-score vs list-recall rho 0.20 to 0.04). Thus, session z-scoring does not inflate the AUC of this decoder, and a short sliding feature z-score costs ranking accuracy. For a fixed decoder, keep the calibration z-score and adapt only the threshold. Limits: one dataset and session, one decoder, offline only, small effects. The package contains the plan, code, features with checksums, results and a figure.

SEEG · word recall · brain-computer interface · feature normalization · logistic regression · AUC · calibration