A three-subject pilot found no consistent reference-scheme effect on cross-session decoding from medial temporal lobe SEEG

Question: For SEEG brain-computer interfaces that read cognitive state, does the re-referencing scheme change single-trial decoding that must transfer to a new session? Method: We used depth-electrode SEEG (Behnke-Fried macro contacts in hippocampus, amygdala, entorhinal and parahippocampal cortex) from a Sternberg verbal working-memory task (OpenNeuro ds004752 v1.0.1). We calculated log band power (theta to high gamma) under four references and used shrinkage LDA with leave-one-session-out validation. Evidence: In 3 subjects (2 sessions each, 266 clean trials), set size 8 vs 4 during maintenance was decodable with mean AUC 0.66-0.73, and encoding vs maintenance with mean AUC 0.88-0.91. The differences between references were small and not consistent across subjects (Friedman p = 0.24 for load, p = 0.46 for phase). Limits: The sample is 3 of 15 subjects because the download link was about 1 MB/s. Group tests have almost no power. Permutation tests used only 20 permutations. The package has the full pipeline and session queue so that other researchers can extend the test to all 15 subjects.

SEEG · brain-computer interface · working memory · cross-session decoding · reference scheme · LDA · medial temporal lobe · OpenNeuro