Hippocampal high-gamma power decreased with verbal working-memory load in one session but did not replicate

We tested if the number of letters in verbal working memory (4, 6 or 8) changes hippocampal SEEG power during the maintenance period. We used OpenNeuro ds004752 v1.0.1 (15 patients, Sternberg task). We downloaded only the 3-6 s maintenance records with HTTP range requests and kept the SHA-256 of all fetched bytes. For each bipolar pair, we calculated the Spearman rho between set size and z-scored log band power over correct trials (3.5-6.0 s). We averaged rho over the hippocampal pairs of each patient and used a sign-flip permutation test with Holm correction over 5 bands. Session 1 was the primary test and session 2 was the replication. In session 1, 70-140 Hz power decreased with load (mean rho -0.068, 10 of 14 patients negative, p = 0.0075, Holm p = 0.038). In session 2 of 10 patients, the decrease did not replicate (mean rho -0.014, p = 0.82). Theta, alpha, beta and 30-48 Hz power did not change with load in either session. No region passed FDR correction. Response time increased with load (rho 0.32, p = 0.0004). Thus, this trial-level test does not support a robust load effect on hippocampal band power during maintenance. An independent publication (08e14772) with a similar design reported almost the same values (session 1: -0.064, Holm p = 0.175; session 2: -0.014). Thus the session-1 significance depends on the window, band and artifact choices, and the decrease did not replicate in either analysis. The package contains the code, per-trial features, byte-range hashes, results and instructions.

verbal working memory · hippocampus · SEEG · high gamma · maintenance period · replication