A 20-trial warm-up preserved cross-patient SEEG error-prediction AUC in a 13-subject working-memory study

This study tested whether cross-patient prediction of working-memory errors from stereo-electroencephalography (SEEG) requires non-causal normalization using all trials in a test session. Leave-one-subject-out evaluation included 13 subjects (13 sessions) performing load-3 object working memory. A shrinkage linear discriminant analysis (LDA) used 10 channel-mean log band-power features from bipolar MTG/MFG SEEG; AUC was evaluated on trials 21 to the end. Mean AUC was 0.613 with all-trial test-session z-scoring (SESSION), 0.605 with expanding past-trial z-scoring (CAUSAL), 0.618 with z-scoring from the first 20 trials frozen for later trials (WARMUP20), and 0.569 with no per-session normalization (NONE). WARMUP20 was non-inferior to SESSION at a 0.02 margin (delta +0.005, 95% bootstrap CI -0.001 to +0.011; p = 0.0001). CAUSAL did not meet the stated non-inferiority rule: p = 0.035, but its 95% CI lower bound was -0.0202, just below -0.02. NONE was worse than SESSION (delta -0.044, 95% CI -0.074 to -0.014; two-sided sign-flip p = 0.020). Sensitivity analysis excluding sub-MU3 showed the same pattern. The cohort excluded sub-MU2 ses-01 because its EDF download failed; sub-MU3 had only one error trial in the evaluation set, making its AUC uninformative. Training subjects were always z-scored over their full sessions. The result is limited to one feature set, one classifier and shrinkage value, this load-3 task, and MTG/MFG bipolar SEEG. CAUSAL and WARMUP20 require 20 unlabeled warm-up trials before scoring.

stereo-electroencephalography · working memory · error prediction · cross-patient prediction · test-session normalization · leave-one-subject-out · linear discriminant analysis · non-inferiority · OpenNeuro ds006136