In two epilepsy patients, SEEG decoded verbal working-memory load (8 vs 4 letters) during stimulus-free maintenance (AUC 0.72, 0.79); fixation stayed near chance

This pilot asked whether a linear decoder could distinguish verbal working-memory set sizes of 8 and 4 letters from single-trial depth-electrode (SEEG) band power during the 3-second maintenance period, when no stimulus was shown. The analysis used OpenNeuro ds004752 version 1.0.1 and included the four sessions from two of 15 patients with epilepsy (99 correct, artifact-free trials). Shrinkage linear discriminant analysis with stratified five-fold cross-validation gave maintenance-period AUCs of 0.72 (sub-03, permutation p = 0.059) and 0.79 (sub-05, p = 0.020; the minimum attainable with 50 permutations). The pre-stimulus fixation control gave AUCs of 0.42 and 0.58, neither significant. Maintenance decoding exceeded encoding decoding in both subjects. Band-specific and anatomical results varied between subjects; non-hippocampal contacts had higher AUCs than hippocampal contacts, but only 6–8 hippocampal channels were available per subject, so this does not show that the hippocampus lacks load information. Bipolar referencing was better than or equal to the recorded common reference. This is a pilot, not a population estimate. Permutation p-values are coarse and uncorrected across 13 conditions per subject; only two cross-validation repeats and 50 permutations were used. Correct-trial selection excluded more set-size-8 trials, and within-session z-scoring used all trials. The archive preserves the analysis code, per-trial features, results, and figure.

verbal working memory · SEEG · intracranial EEG · neural decoding · working-memory load · band power · linear discriminant analysis · epilepsy · maintenance period · OpenNeuro