Load-effect signs varied across 15 SEEG patients, and sign-prior decoders did not transfer load without calibration

We tested whether verbal working-memory load changes maintenance band power in a consistent direction across patients and atlas regions, and whether a decoder can use that direction without patient-specific labels. We analyzed archived per-trial log10 Welch band-power features from 15 patients in OpenNeuro ds004752 version 1.0.1: 669 bipolar channels, set sizes 8 versus 4, five bands from 4 to 150 Hz, and the 3–6 s maintenance window. No region-by-band cell showed a consistent sign across patients (all Holm-adjusted p = 1; smallest uncorrected p = 0.075 for hippocampal 70–150 Hz, mean d = −0.14, 11/14 patients negative). Patient mean effects reached |d| = 0.64 but varied in sign. At most 2.6% of channels per band passed within-patient BH-FDR, from 1–3 patients per band. Leave-one-patient-out sign-prior decoders had mean AUC 0.537 for region-by-band signs (p = 0.39), 0.506 for band signs (p = 0.86), and 0.573 for 70–150 Hz alone (p = 0.10). The alpha-only AUC was 0.405 (p = 0.004), consistent with leave-one-out bias when the group mean is near zero and not evidence of a reversed effect. In this dataset, anatomical or spectral sign maps did not transfer load effects between patients. A load decoder must learn feature direction from labeled data in the same patient. This is one dataset, task, session, and window, with small trial counts, mostly temporal coverage, unchecked atlas labels, and a simple equal-weight decoder; weighted transfer may differ.

intracranial EEG · SEEG · working memory · memory load · band power · patient-specific decoding · brain-computer interface · cross-patient transfer · sign-flip test · OpenNeuro ds004752