Across 32 planned pipelines, all 18-patient ds004789 FR1 decoders exceeded chance, but median AUC varied from 0.584 to 0.681

We compared 32 planned within-patient SEEG word-recall decoding pipelines on archived band-power features from 18 patients in OpenNeuro ds004789 FR1, session 0. All 32 pipelines had group AUC above 0.5 (exact one-sided Wilcoxon tests; largest p = 4.2e-4), but group median AUC ranged from 0.584 to 0.681. In the reference pipeline across 10 cross-validation seeds, the median ranged from 0.660 to 0.681. Patient and pipeline accounted for 62% and 11% of per-patient AUC variance, respectively. For 15 of 18 patients, AUC SD across pipelines exceeded SD across seeds (medians 0.040 and 0.016); patient ranking agreement was Kendall W = 0.61 across pipelines and 0.88 across seeds. Strongly regularized, class-balanced logistic regression (C = 7.2e-4) had the highest AUC; shrinkage LDA and logistic regression with C = 1 were lower by 0.050 and 0.031. Session versus training-fold z-scoring changed AUC by 0.0015. Results are limited to offline decoding, one session per patient, one archived feature pipeline, and 18 patients. The pipeline and cross-validation scheme affected reported AUC, so both should be reported.

SEEG · word recall · intracranial EEG · decoding · AUC · cross-validation · analysis multiverse · feature extraction · patient ranking