For ds004752, 8-trial causal centering did not significantly improve day-1 load-decoder accuracy across 34 later-day sessions in 10 subjects (+0.02 BA, p = 0.46)

We trained a shrinkage-LDA memory-load decoder on day-1 SEEG band power and tested it on sessions 1–3 days later. At the fixed threshold, balanced accuracy (BA) was 0.588 with day-1 statistics. Causal centering with the first 8 unlabeled trials gave 0.609 (p = 0.46), and the non-causal full-session z-score gave 0.636 (p = 0.13). Only the full-session z-score significantly decreased the drift in the predicted label rate (0.32 to 0.13, p = 0.006). The best-threshold BA on the test set is optimistic by about 0.12 in these small sessions: with shuffled labels, its value is 0.62. Scope: 1 dataset, 1 contrast, 1 feature set, 1 decoder, and a median of 23 evaluated trials per session. We assume that the npz trial order is the recording order. Thus, earlier ds004752 cross-day AUC results that used a full-session z-score are an upper bound for a causal online decoder.

SEEG · memory load · shrinkage LDA · cross-day decoding · causal centering · balanced accuracy · OpenNeuro ds004752