In eight SEEG subjects, an L1 decoder met a 0.02-AUC non-inferiority margin versus shrinkage LDA while using a median 24% of channels

In an offline, single-session analysis of eight subjects performing the SlowFast cued key-press task, we compared an L1-penalized logistic regression decoder with shrinkage linear discriminant analysis (LDA) for separating cued key-press from pre-cue fixation-rest windows. We used archived 0.5 s bipolar log-power features in the 70–150 Hz and 13–30 Hz bands, trained on trials 1–100, and tested on trials 101 onward. The pre-specified non-inferiority test used a 0.02 AUC margin. L1 minus LDA test AUC averaged +0.005 (median +0.006; range −0.015 to +0.024); all eight subjects exceeded the −0.02 margin (exact one-sided sign-flip p = 0.0039). The L1 model, with C selected by five-block cross-validation on training trials, used 8–50 of 100–112 bipolar channels (median 24%). In descriptive, uncorrected comparisons at hard budgets of four and eight channels, the L1 path had lower AUC than LDA on the top-K channels ranked by training |t| (mean differences −0.090 and −0.019; one of eight subjects favored L1; p = 0.047 for each). Thus, in this dataset and task, L1 pruning retained performance at a median of about one quarter of channels, while top-|t| ranking performed better at the smallest tested budgets. The study is limited to one task, eight subjects, one session per subject, and offline evaluation; there was no closed-loop test. The C grid had 15 values, and some hard-budget solutions used fewer channels than allowed. The windows do not separate task cue from movement, and sub-11 had only 32 test trials.

SEEG · brain-computer interface · key-press detection · L1 logistic regression · shrinkage LDA · feature selection · AUC · non-inferiority · OpenNeuro ds004473