A rank-1 SEEG recall decoder projection reduced held-out AUC in 18 patients

A memory brain-computer interface may use a low-parameter decoder to predict whether a word will be recalled. This study tested whether replacing a full-rank within-patient SEEG decoder's channel-by-frequency weights with their rank-1 singular-value decomposition (SVD) projection improved held-out AUC. The analysis used archived per-word log band-power features from OpenNeuro ds004789 v3.1.0, free recall (FR1), session 0, in 18 patients. The preregistered primary comparison trained on the first three lists and tested on lists 13 and later. Rank 1 decreased AUC: median patient difference (rank 1 minus full rank) was −0.017 (n=17, 95% bootstrap CI −0.031 to −0.010; 3/17 patients improved; exact two-sided Wilcoxon p=0.040). In leave-one-list-out analysis, rank 1 was lower by a median 0.031 AUC (n=18; Holm-adjusted p=0.028). Rank 1 and rank 2 did not exceed the full-rank decoder at the median in any reported comparison. Rank 1 explained a median 0.28 of weight energy, indicating a shared spectral profile, while channel-specific spectral weights remained relevant to recall information. The results support no benefit from a strictly separable channel-by-frequency decoder for these features and this offline task. Limits include one dataset and session, one fixed regularization value, band-power features only, late-list testing with possible session drift, and small patient counts; this was not a closed-loop test.

SEEG · free recall · memory brain-computer interface · singular value decomposition · rank-1 projection · AUC · band power