Inner-CV shrinkage did not improve SEEG memory-load AUC over Ledoit-Wolf in 15 patients

In an offline analysis of archived SEEG band-power features from 15 patients performing a verbal working-memory task, shrinkage linear discriminant analysis (LDA) with inner cross-validation (inner-CV) did not improve held-out discrimination of set size 8 versus 4 over analytic Ledoit-Wolf (LW) shrinkage. Mean patient AUC was 0.715 for inner-CV and 0.723 for LW (difference -0.008, 95% CI -0.021 to +0.003; 5 of 15 patients above zero; exact two-sided sign-flip p = 0.23). Oracle Approximating Shrinkage was slightly worse than LW (difference -0.0018, 95% CI -0.0029 to -0.0007; Holm p = 0.031), although the difference was judged too small to matter. A fixed shrinkage value of 0.8 had a higher mean AUC than LW (+0.011; Holm p = 0.031), but this value was selected on the same data, so the apparent gain is optimistic and needs testing on new data. Inner-CV selected stronger shrinkage (median 0.95) than LW (median 0.44), consistent with noisy inner-fold AUC estimates at this trial count. The result supports LW as a safe default for this setting; it is limited to one dataset, session, task, feature set, and inner-CV tie rule.

stereo-electroencephalography · working memory · linear discriminant analysis · shrinkage · Ledoit-Wolf · cross-validation · area under the curve · brain-computer interfaces