In seven patients, one training selected SEEG shaft decoded less well than all shafts; random subsets needed a median of 32 channels to retain 90% of above-chance AUC

This study asks how decoding performance changes when a band-power decoder uses fewer intracranial channels or only one depth shaft. Using session 1 from seven patients in OpenNeuro ds004752 v1.0.1, we compared maintenance versus fixation (task phase) with verbal working-memory load (8 vs 4 letters). Shrinkage LDA was evaluated with trial-grouped, repeated five-fold cross-validation. The median AUC across all SEEG channels was 0.89 for phase and 0.82 for load. One shaft selected within training data had median AUCs of 0.80 and 0.61; it scored below all channels in 7/7 patients for phase (p = 0.016) and 6/7 for load (p = 0.031). It retained a median 0.81 and 0.45 of above-chance AUC, respectively. The best shaft chosen after seeing test data had median load AUC 0.79, showing that this post-hoc choice overstates performance expected from training-only selection. Random channel subsets needed a median of 32 bipolar channels (range 8–56 for phase; 16–56 for load) to retain 90% of above-chance AUC. Eight training-selected channels performed better than eight random channels, but remained below all channels for phase. All SEEG channels did not significantly outperform 19 scalp EEG channels, nor did eight random SEEG channels outperform eight scalp EEG channels. Results are limited to seven clinical patients, one session each, few trials, and a linear decoder on band power. Load decoding varied substantially; sub-01 and sub-02 were not above the null. Phase decoding may reflect slow non-specific changes such as arousal or eye movement as well as memory maintenance. The package includes original analysis code, results, derived features, and raw-file hashes.

SEEG · intracranial EEG · working memory · channel selection · brain-computer interface · band power · shrinkage LDA · ROC AUC · OpenNeuro