In 19 FR1 SEEG patients, a running quantile reduced offline stimulation-rate error

We evaluated whether a recall-decoder threshold set on training lists controlled the stimulation rate on later lists in the same session, whether flagged words had lower recall, and whether a label-free running quantile improved rate control. Using archived FR1 session-0 word-level band-power features from 19 patients in OpenNeuro ds004789 v3.1.0, we trained an L2 logistic decoder on the first 60% of lists and tested on the remainder. For a target rate of 0.25, the leave-one-list-out training quantile produced a mean test rate of 0.313 (range 0.00–0.53; mean absolute error 0.113). Its signed error was +0.063 (12/19 above target; Holm p = 0.056), which was not significant. Flagged words had 11.2 percentage points lower recall (15/18 patients; Holm p = 0.0012). A causal running quantile of prior test-list scores, without labels, reduced mean absolute error to 0.043 (14/19 patients improved; Holm p = 0.0049) and retained negative enrichment (-9.8 percentage points). A fixed probability threshold of 0.5 flagged 69.4% of words. Across target rates 0.10 and 0.50, mean absolute errors were 0.070 vs 0.032 and 0.131 vs 0.057 for the fixed and running quantiles, respectively. The test AUC mean was 0.628 (median 0.614). This was an offline simulation with one session per patient, short test blocks of about 4–10 lists, and modest decoder AUC; stimulation effects and cross-session drift were not modeled.

closed-loop brain-computer interface · memory · recall decoder · SEEG · stimulation rate · threshold calibration · distribution shift · offline simulation