Cross-patient SEEG decoding of working-memory load was weak and did not reduce calibration trials

We tested if a memory-load decoder that we train only on other SEEG patients can decode a new patient without calibration. We used the archived maintenance-period band power of 15 patients (OpenNeuro ds004752, session 1). The features are channel-agnostic: band power averaged over all bipolar pairs, or over 5 brain regions. With zero labeled trials, the leave-one-patient-out decoder gave a mean AUC of 0.569 for set size 8 vs 4 (11 of 15 patients above 0.5, sign-flip p = 0.037). Region features did not improve it. A within-patient all-pair decoder gave AUC 0.729. In a calibration curve, the cross-patient prior did not save labeled trials: an all-pair logistic decoder with 2 trials per class of the new patient gave AUC 0.607, and it was better than the prior plus the same trials at k = 2, 4, 6 and 8 per class. No band had a consistent load effect over patients, so the channel average removes most of the load information. Limits: 15 patients, one session, 20-42 trials per patient, one time window, 5 bands, uncorrected tests. We found that shrinkage LDA is numerically degenerate with 2 trials per class, and we report this.

SEEG · working memory · memory load · cross-patient decoding · brain-computer interface · band power · calibration · OpenNeuro