Time-ordered folds did not lower SEEG working-memory load decoding AUC in 15 patients, despite measurable feature drift

We compared random and time-ordered cross-validation for decoding working-memory load from SEEG band power in 15 patients performing a verbal Sternberg task (OpenNeuro ds004752, version 1.0.1). The label was set size 8 versus 4; each patient contributed 20–42 trials and 70–280 features. A z-scored, Ledoit–Wolf shrinkage LDA gave mean AUCs of 0.706 for repeated stratified random 5-folds, 0.714 for blocked folds, 0.659 for forward chaining, and 0.701 for a first-half/second-half split. All schemes exceeded their permutation nulls at group level (sign-flip p = 0.0006–0.005). Random minus blocked AUC was −0.008 (95% CI −0.033 to 0.017; p = 0.51); random minus forward AUC was +0.047 (95% CI −0.011 to 0.105; p = 0.10). With training-set size controlled by shuffling trial order, the forward difference was −0.003 (95% CI −0.058 to 0.052; p = 0.90), consistent with the lower forward AUC arising from fewer training trials rather than chronology. Features changed over the session (time-drift AUC 0.71; 15/15 patients above 0.5; p = 6e-5), but drift size did not predict the random–time-ordered gap. In this task, random folds did not show higher AUC than time-ordered folds despite measurable drift. Set sizes were interleaved at random, so drift may not have aligned with the label. This does not establish equivalence: the random–forward confidence interval includes differences up to 0.10, and blocked changes in set size could confound drift with label. Limits include one session per patient, small trial counts, use of trial order rather than elapsed time, one feature set and classifier, offline analysis, and chronology checks in only 3 of 15 patients. The npz trial order is assumed to be chronological.

SEEG · working memory · cross-validation · temporal generalization · neural decoding · feature drift · Ledoit-Wolf LDA · AUC