Conformal thresholds kept mean false-stimulation rates near target in 18 SEEG patients, while patient-level rates varied widely

This offline analysis tested class-conditional split-conformal thresholds for closed-loop memory stimulation using archived band-power features from 18 SEEG patients performing free recall (OpenNeuro ds004789, FR1, session 0). At a target false-stimulation rate (FSR) of 0.2, mean FSR was 0.181 with a cross-patient decoder calibrated on each patient's first-half lists, 0.230 with a within-patient decoder trained on the first quarter and calibrated on the second quarter, and 0.227 when calibration used other patients only (zero labels from the new patient). No method had an FSR significantly above the target (one-sided exact Wilcoxon tests, Holm-adjusted p >= 0.61). The time-ordered split did not differ detectably from the mean of 200 random list splits (exploratory p = 0.64 for the cross-patient decoder and p = 0.25 for the within-patient decoder). The limit was marginal, not per-patient: within-patient FSR ranged from 0 to 0.60, and 5 of 18 patients exceeded 0.30 at target 0.2. The small recalled-word calibration sets (5-33 words; median 12) limit patient-level precision. Yield exceeded FSR for all methods (Holm-adjusted p = 0.0003-0.001), but the controller reached only 26-33% of not-recalled words at target 0.2. This is a single-dataset offline result; no stimulation was delivered, later recall was used as a proxy for good encoding, and session-wide label-free z-scoring would need causal statistics in an online system.

conformal prediction · closed-loop stimulation · memory · SEEG · false-stimulation rate · split conformal · free recall · calibration · offline analysis