In 18 SEEG patients, recall decoding persisted weakly to the next word and two-word smoothing reduced AUC

Question: a closed-loop memory BCI classifies the brain state during one word and then stimulates, often during the next word. Does the decoded encoding state persist for one word-onset interval (median 2.50 s)? Method: we used the per-word SEEG band-power features and the L2 logistic decoder (W_chan, leave-one-list-out) of sci c4b8c884, and a training-free spectral-tilt score. We calculated the AUC for recall of word n from the score of word n-k in the same list (k = 0-6, and k = 12 for the same position in the previous list). The AUC compared only words at the same serial position. Exact Wilcoxon tests over 18 patients; plan written before the lagged analysis. Results: the own-word AUC was 0.618 (16/18). The previous-word AUC (primary test) was 0.530 (11/18, one-sided p = 0.037; Holm p = 0.22). Lags 2-6 and 12 were near 0.5. On the same target words, lag 1 was higher than lag 12 (0.521 vs 0.486, 12/18, p = 0.014). The sum of the outputs of words n-1 and n gave a lower AUC than the output of word n alone (0.603 vs 0.627, 6/18 higher, p = 0.034). The decoder output was autocorrelated between neighbouring words without labels (residual r 0.022 vs shuffle null -0.091, 17/18, p = 5e-5), but this persistent part did not carry much recall information. Recall of word n-1 predicted recall of word n (AUC 0.587, 16/18) better than the decoder did. When we also held recall of word n-1 constant (exploratory), the lag-1 decoder AUC was 0.530 (p = 0.077). The tilt score had no lag-1 effect (0.498). An unplanned lag-5 effect of the tilt score (0.536, Holm p = 0.003) has no explanation and needs replication. Conclusion: about 75% of the above-chance AUC is lost within one word. A memory BCI that uses this type of decoder must act within the classified word, and smoothing over words does not help. Limits: one dataset and task, session 0, 18 patients, offline, word-averaged 0-1.6 s features; filter leakage between neighbouring windows can only increase lag-1 effects.

SEEG · free recall · memory decoding · AUC · closed-loop brain-computer interface