Causal calibration lets a recall decoder beat spectral tilt after six lists in offline pyFR SEEG data

We tested a memory brain-computer interface read-out in 19 pyFR patients using one session per patient. A within-patient L2 logistic decoder (fixed C = 7.2e-4) was trained on the first k word lists and tested on later lists, with z-scoring based only on calibration words. The comparison was a zero-label spectral tilt (80–140 Hz minus 3–8 Hz). With 1–4 lists, there was no significant difference. With 6 lists (90 words), the decoder performed better: median difference +0.038 AUC, 17/18 patients, p = 0.0016; the result remained significant after Bonferroni correction across 8 values of k (adjusted threshold 0.013). Within serial position, the difference was +0.035 AUC (16/18, p = 0.0026). At 8 lists the difference was not significant (+0.055, p = 0.080); at 10 and 12 lists it was significant (p = 0.012 and 0.039). Causal z-scoring changed decoder AUC by a median of -0.007 to +0.001. In a fixed late test block (lists 13–16), early, recent, and random calibration lists did not differ significantly (all p >= 0.16). Late lists were harder: spectral-tilt AUC fell by 0.054 from the early to late half (p = 0.021), and recall rate fell by 0.05 (p = 0.0036). A single chronological list did not make the decoder worse than tilt (p = 0.92). All AUCs were small (0.53–0.60). This was an offline analysis of one task and one session per patient; the late test block was small, patient counts varied with k, and p values were evaluated per k. “Causal” applies across words and lists, not within a word: the feature window requires data through about 2.6 s after word onset.

SEEG · memory · brain-computer interface · recall prediction · causal calibration · spectral tilt · logistic decoder · pyFR