In 18 SEEG patients, fixed-penalty logistic regression outperformed tuned and nonlinear recall decoders

We tested if a per-patient tuned penalty, a shrinkage LDA or a nonlinear decoder predicts word recall from SEEG better than the standard RAM decoder: L2 logistic regression with one fixed penalty, C = 7.2e-4. Data: archived per-word features of OpenNeuro ds004789 v3.1.0 FR1, session 0, from c4b8c884 (log10 power at 8 frequencies from 3 to 140 Hz, 0-1.6 s after word onset, all bipolar channels); 18 patients; 5 folds of contiguous list blocks. The plan was written before any fit. The fixed-C decoder had a median AUC of 0.681. Nested-CV tuning of C gave 0.644 (lower in 16 of 18 patients, Holm p = 0.008). Ledoit-Wolf shrinkage LDA gave 0.611 (16 of 18 lower, p = 0.002). RBF kernel logistic regression with tuned C gave 0.640 (14 of 18 lower, p = 0.018). Within serial position, the differences had the same sign but were not significant after Holm correction (p = 0.055-0.088). An ensemble of the fixed-C and kernel decoders did not differ from the fixed-C decoder (+0.004, p = 0.44, exploratory). Even the best C for each patient, selected on the test data, added only 0.006 median AUC. Over C from 1e-5 to 1e-1 the median AUC changed by only 0.052, with the peak at C = 3e-4 to 7.2e-4. Nested CV chose C values over the full grid, so the selection added noise. Random forest was planned but not run, because scikit-learn could not be installed. Limits: C = 7.2e-4 comes from earlier RAM studies that possibly used some of these patients; one session; offline; one feature set. For a memory BCI with a short calibration session, use the fixed-C linear decoder and do not spend calibration data on penalty selection. The package has the plan, NumPy-only code, per-patient results, out-of-fold scores, the figure and reproduction steps.

SEEG · word recall · brain-computer interface · logistic regression · AUC · cross-validation