One SEEG contact captured most cross-patient error-prediction AUC in 13 working-memory sessions

In 13 sessions from 12 patients performing a load-3 object working-memory task, we evaluated offline prediction of trial errors from bipolar SEEG features using L2 logistic regression trained on other patients. With one random test contact, median per-session ROC AUC was 0.563 (12 of 13 sessions above 0.5; Wilcoxon p = 0.003); with all 6–18 contacts it was 0.591. The median gain from one to all contacts was 0.025 (p = 0.006). At equal contact counts in 10 sessions, median AUC was 0.592 for middle temporal gyrus (MTG) and 0.548 for middle frontal gyrus (MFG); the paired difference was not significant (median +0.022, p = 0.49). Only MTG-only AUC was significantly above chance in the region-specific tests. The dataset contains task-modulated, artifact-cleaned contacts selected by its authors, so performance may be higher than for random contacts in a new patient. Results are offline and do not establish that MTG is better than MFG.

SEEG · working memory · error prediction · cross-patient decoding · ROC AUC · contact count · middle temporal gyrus · middle frontal gyrus · intracranial EEG