In 15 SEEG patients, a load decoder read memory load but did not predict response time within load

We tested whether a held-out SEEG decoder score predicted response time (RT) within a memory load and probe-match condition in a verbal Sternberg task. The analysis used session 1 from 15 patients and archived log10 band power (3–6 s maintenance, bipolar channels × 5 bands). A z-scored Ledoit-Wolf shrinkage LDA decoded set size 8 versus 4 with stratified 5-fold cross-validation repeated 10 times; set size 6 trials were scored by a model trained on all set size 4/8 trials and were never used for training. The plan was written before analysis. In 616 correct trials, the primary within-cell association was null: mean r = 0.000, 95% CI −0.090 to 0.090, 9/15 positive, exact two-sided sign-flip p = 0.995. No patient had permutation p < 0.05 (lowest: sub-13, r = −0.33, p = 0.057). Positive controls passed: held-out AUC for 8 versus 4 had mean 0.705 and median 0.694, above 0.5 in 14/15 patients (p = 0.0004); RT and score both increased with set size (Spearman r = 0.309 and 0.325; p = 0.0002 and 0.0009). For set size 6 alone, r = −0.130 (95% CI −0.281 to 0.026, p = 0.098). Channel-mean 70–150 Hz power had a small positive within-cell association (r = 0.092, 95% CI 0.011 to 0.172, p = 0.028), an uncorrected secondary result. In this task, the band-power load-decoder score reflected stimulus condition and did not measure trial-level effort or speed. It should not be used as a continuous cognitive-load readout for adaptive BCIs without separate validation. The study tested one task, one session, and maintenance band power only; button-press motor variance may hide a cognitive relation. The group CI excludes within-cell correlations larger than about |0.09|, while the study's sample size and trial counts limit precision.

working memory · SEEG · response time · brain-computer interface · load decoding · verbal Sternberg task · band power