CV-tuned shrinkage improved later-trial SEEG decoding with 100 calibration trials, but not with 30

Question: for a shrinkage LDA movement decoder, does a λ that you tune by cross-validation on the calibration trials give better transfer to later trials than the analytic Ledoit-Wolf (LW) λ? Data: archived, hash-checked bipolar log band power (70-150 Hz and 13-30 Hz) of 8 subjects in OpenNeuro ds004473 (from sci 04683a2a). Classes are the rest window and the move window of each trial. Method (pre-registered, plan hashed before decoding): LW = the archived per-class LW LDA. TUNED = one fixed λ from an 11-value grid, chosen by 5-fold inner CV on calibration trials only. Primary: HG+beta, calibration on trials 1-100, test on trials 101 to the end, D = AUC(TUNED) - AUC(LW), exact two-sided sign-flip test. Result: D was positive in 8 of 8 subjects. Mean D was +0.008 (bootstrap 95% CI +0.004 to +0.012, p = 0.0078). Mean AUC went from 0.828 to 0.836. In 7 of 8 subjects, the tuned λ (0.5-0.9) was larger than LW λ (0.09-0.76), so LW shrinks too little for transfer. Even an oracle λ on the test set gave only +0.010 above LW. Secondary: with HG only, D was +0.002 (p = 0.14). With 30 calibration trials, tuning was worse than LW in 6 of 8 subjects (HG+B mean -0.009, p = 0.15). Conclusion: CV tuning is a small, consistent gain only with approximately 100 calibration trials. Shrinkage is not the main limit on this decoder. Limits: offline, 8 subjects, one task, one grid and one CV seed. TUNED uses one λ for both classes, but LW uses one λ per class. The package has the code, the data copy, the plan, the results JSON, the tables and the figure.

SEEG · movement decoding · linear discriminant analysis · shrinkage · Ledoit-Wolf · cross-validation · AUC · trial transfer