Mild shaft-wise smoothing preserved group word-recall AUC in 19 SEEG patients, but shaft means reduced AUC in some patients

We tested whether Gaussian smoothing of log band power across neighboring bipolar pairs on the same SEEG shaft changes within-patient word-recall decoding. We used 19 patients' per-word features from OpenNeuro ds004789 v3.1.0, RAM free recall session 0: eight frequencies from 3–140 Hz in the 0–1.6 s after word onset. A class-balanced ridge classifier was evaluated with leave-one-list-out testing and with calibration on lists 1–3 followed by later-list testing. Across patients, smoothing with sigma = 1 pair step did not change AUC (leave-one-list-out mean difference −0.002, 95% CI −0.014 to +0.010; 13/19 patients improved; Holm-adjusted p = 1.0). The shaft mean also showed no significant group change (−0.008, 95% CI −0.034 to +0.016), but two patients lost 0.12–0.15 AUC. In the three-list calibration evaluation, sigma = 1 changed mean AUC by +0.001 (95% CI −0.006 to +0.008). Results were similar without R1070T, which had 14 recalled words and was excluded by the source feature study's threshold of at least 15. These offline results cover one session, task, time window, and fixed-penalty linear decoder. Session z-scoring used all words without labels, pooled AUC includes serial-position effects, and smoothing log power does not model a physically larger contact. A shaft-level read-out may reduce decoding performance for individual patients.

SEEG · word recall · band power · spatial smoothing · ridge classification · brain-computer interface · AUC