In 18 SEEG patients, nonlinear decoders did not outperform a linear decoder for word recall from band power

We compared linear and nonlinear decoders of later word recall from archived per-word SEEG band-power features in 18 patients from OpenNeuro ds004789 (FR1, one session per patient; 144–300 words and 704–1384 features per patient). We used list-blocked cross-validation with nested tuning, as preregistered. RBF-kernel logistic regression did not differ from nested L2 logistic regression (median AUC 0.648 vs 0.651; median difference −0.007, 95% CI −0.020 to 0.023; exact Wilcoxon p = 1.0); the difference within serial position was 0.000. The confidence interval does not exclude a nonlinear gain of 0.02 AUC. Gradient-boosted trees had lower median AUC (0.613; difference −0.024 vs nested linear, Holm p = 0.091). A linear decoder with fixed C = 7.2×10⁻⁴ had the highest median AUC (0.681; above chance in 18/18 patients). Nested tuning of C decreased AUC in 16/18 patients (exploratory p = 0.008), consistent with noisy inner-fold AUC at this data size. The fixed C came from an earlier analysis of the same data and may give a small optimistic bias. For this one-session, per-word band-power setting, these results support a strongly regularized linear decoder; they do not establish that nonlinear decoders cannot help with more data or other features.

SEEG · free recall · memory decoding · brain-computer interface · band power · logistic regression · kernel logistic regression · gradient-boosted trees · cross-validation · AUC