In 7 SEEG patients, a running score quantile kept median flag-rate error near 5 points for a 30% target; frozen thresholds ranged 2% to 89%

We tested label-free threshold rules for a word-recall decoder that is trained on session 0 and used in session 1 (a new day). The use case is dose control for closed-loop memory stimulation: the system stimulates on flagged words, so the flag rate is the dose. Data: archived per-word band-power features from OpenNeuro ds004789 (FR1), 7 patients, channels common to both sessions. Decoder: L2 logistic regression. Target flag rate: 30% (also 20% and 50%). With the frozen session-0 threshold, the session-1 flag rate was 0.02 to 0.89 (median absolute error 0.28). A causal running quantile of earlier session-1 scores gave a flag rate of 0.24 to 0.49 (median error 0.05). Its error was smaller in 7 of 7 patients (Holm p = 0.047). Re-centering the features on 2 unlabeled lists alone was not sufficient (median error 0.18; p = 0.094 vs running quantile). The flagged words of the running-quantile rule had a precision lift of 1.08 (6 of 7 patients > 1; Holm p = 0.047). A non-causal oracle (z-score on all session-1 words) gave a median error of 0.014. Limits: n = 7; offline, no stimulation effect modeled; weak decoder (AUC 0.47 to 0.65); per-list rates stayed variable; drift inside a session caused errors in one patient. Package has the pre-analysis plan, numpy code, input hashes, results and an SVG figure.

SEEG · free recall · memory decoder · threshold calibration · running quantile · closed-loop stimulation · flag rate · session shift · intracranial EEG