Eight training fold selected SEEG channels were not shown non-inferior for memory load decoding in 15 patients

We tested whether eight bipolar SEEG channels selected within training folds were non-inferior to all channels for decoding verbal working memory load in 15 patients. The analysis used archived 3–6 s log10 band power from OpenNeuro ds004752 (20–42 trials per patient, 14–56 channels), shrinkage LDA, stratified five-fold cross-validation repeated 10 times, and a non-inferiority margin of 0.02 AUC. For the primary 70–150 Hz feature, all channels yielded mean AUC 0.648 and eight selected channels yielded 0.637. The mean difference was −0.011 (95% patient bootstrap CI −0.065 to +0.038; exact one-sided sign-flip p = 0.38), so non-inferiority was not shown. The result also does not show that eight channels are worse. Only 32 channels passed the margin (p = 0.0004); this was 57–100% of each array. With five bands, eight channels also did not meet non-inferiority (mean difference −0.020, 95% CI −0.065 to +0.027, p = 0.49). Selection using all trials overstated AUC by 0.19 at eight channels, while training-fold selection exceeded random selection by 0.061 (p = 0.016). The study is limited to offline analysis of one session in 15 epilepsy patients with archived features and has low power for a 0.02 margin. The findings do not support transferring an eight-channel sufficiency result from a higher-trial motor task to cognitive decoding with small calibration sets without a new test.

SEEG · intracranial EEG · working memory · channel selection · feature selection · non-inferiority · AUC · cross-validation · shrinkage LDA · BCI