Eight selected SEEG contacts captured nearly all decoding effect in two memory tasks; eight random contacts captured similar fractions

We compared the number of stereoelectroencephalography (SEEG) contacts needed to decode two memory tasks using archived bipolar features and one offline pipeline: z-scoring, Ledoit-Wolf shrinkage linear discriminant analysis, and stratified five-fold cross-validation with two repeats. The cohorts comprised 18 patients in FR1 word encoding and 15 in verbal Sternberg maintenance. Eight random contacts captured 0.65 (bootstrap 95% CI 0.54–0.78) and 0.70 (0.55–0.86) of the respective all-channel effects; the difference was −0.05 (95% CI −0.24 to 0.14). We therefore did not find evidence for a task-specific contact budget at eight contacts. In both tasks, all channels outperformed eight random contacts (FR1 +0.041, 14/18 patients, Holm p=0.003; Sternberg +0.063, 11/15, p=0.028), and eight contacts selected within training folds outperformed eight random contacts (FR1 +0.041, 16/18, p=0.003; Sternberg +0.053, 11/15, p=0.036). Eight selected contacts captured 1.00 and 0.96 of the all-channel effect. Exploratory: one random FR1 contact captured more of the effect than one random Sternberg contact (0.43 vs 0.13; difference CI 0.07–0.62; not pre-specified and uncorrected). This offline comparison is confounded by cohort, montage, trial count, feature, and window differences. FR1 used word-level rather than list-level folds; its AUC may be slightly optimistic. Sternberg had few trials, and five random subsets and two repeats were used due to time limits. Fractions depend on the shrinkage-LDA all-channel reference; the exploratory saturation fits have no confidence intervals.

stereoelectroencephalography · SEEG · memory decoding · channel selection · contact budget · intracranial EEG · shrinkage LDA · working memory