In 18 patients from one SEEG session, causal z-scoring kept offline recall AUC (median difference 0.000); frozen calibration reduced flag-rate accuracy

Offline memory-BCI decoders often z-score features with statistics of the full session. An online decoder cannot use future words. We used the per-word SEEG power features of sci c4b8c884 (OpenNeuro ds004789, 18 patients) and a pre-registered plan. We trained an L2 logistic recall decoder on the first 6 lists (and 3 lists) and compared 4 normalizations on the remaining lists. Causal cumulative z-scoring gave the same test AUC as session-wide z-scoring (median 0.622 vs 0.633; median paired difference 0.000, 95% CI -0.004 to +0.002, p = 0.64; non-inferior at a 0.02 margin). Frozen calibration statistics were also non-inferior for AUC, but with a threshold fixed at calibration, the flag rate moved away from the 0.5 target (median error 0.147 vs 0.077 for cumulative and 0.045 for a 2-list moving mean; Holm p < 0.001). The moving mean lost more AUC with 3 calibration lists (median -0.021, not non-inferior). Package has code, plan, per-patient results, figure and input checksums.

memory BCI · SEEG · word recall · causal normalization · feature drift · logistic regression · AUC · threshold calibration