Nonlinear and tuned decoders do not predict later word recall from SEEG band power better than a fixed L2 logistic decoder (18 FR1 patients, one session, 5-fold list-grouped CV)

We tested whether four decoder types improved cross-validated prediction of later word recall over a fixed, strongly regularized L2 logistic decoder (C = 7.2e-4) using per-word SEEG band-power features from 18 patients in OpenNeuro ds004789 FR1 session 0. With five-fold list-grouped cross-validation, the fixed decoder had median AUC 0.670. No tested decoder performed better: gradient-boosted stumps had median AUC 0.602 (difference -0.062; Holm p = 0.027), shrinkage LDA had 0.590 (-0.042; Holm p = 0.011), and L2 logistic regression with C selected by inner cross-validation had 0.623 (-0.030; Holm p = 0.077). RBF-kernel logistic regression had median AUC 0.667 (difference -0.003; 95% CI -0.010 to +0.004), within the prespecified +/-0.02 AUC equivalence margin. Within-serial-position AUC showed the same overall pattern. These results apply to one session per patient (about 300 words), fixed settings for the nonlinear models, and the tested 0-1.6 s band-power features; they do not establish performance with more training data or other model settings.

SEEG · word recall · memory decoding · logistic regression · cross-validation · band power