Shrinkage LDA did not reproduce a spatial-residual advantage in 18 patients decoding word recall

We tested whether the reported advantage of spatial-residual over global channel-mean power for decoding later word recall in stereo-EEG (SEEG) remains with a different linear classifier. The prior analysis reported median AUCs of 0.657 for the residual and 0.603 for global power (Holm p = 0.024) using L2 logistic regression and leave-one-list-out (LOLO) testing. We used the same archived features and inclusion rule from OpenNeuro ds004789 FR1, session 0, for 18 patients, and changed the decoder to two-class linear discriminant analysis (LDA) with analytic Ledoit-Wolf shrinkage. In five-fold blocked-by-list cross-validation, median AUCs were 0.595 (global), 0.608 (residual), and 0.605 (all channels); residual minus global was +0.010 (9/18 patients, exact Wilcoxon p = 0.64). In LOLO testing, medians were 0.601, 0.583, and 0.583, respectively; residual minus global was -0.001 (9/18, p = 0.90). The change in cross-validation did not change this result. All six LDA medians were above chance (one-sided p < 0.002), including within serial position. LDA matched the reported L2 logistic AUC on eight global features, but had lower AUC on high-dimensional residual and all-channel features. These results indicate that the reported spatial advantage depends on the decoder and its use of high-dimensional features. The global eight-feature measure is a robust baseline in this dataset. This was not preregistered: LOLO was added after the five-fold analysis. We did not rerun the original logistic code; its reported results are used for comparison. The study covers one dataset, one session, 18 patients, bipolar features, and offline AUC; LDA was not tuned. Per-fold score z-scoring removes offsets between folds and makes pooled LOLO AUC near a within-list AUC.

stereo-EEG · word recall · memory decoding · linear discriminant analysis · Ledoit-Wolf shrinkage · cross-validation · AUC · classifier comparison