Cross-validation tuning improved later-trial SEEG classification over Ledoit-Wolf shrinkage in eight subjects

This preregistered offline replication used new code to compare shrinkage linear discriminant analysis (LDA) estimators on archived high-gamma (70–150 Hz) and beta (13–30 Hz) bipolar log band-power features from eight subjects performing a forced two-choice key-press task. LDA was calibrated on trials 1–100 and evaluated on trials 101 onward. A single shrinkage value selected by five-fold inner cross-validation improved test AUC over per-class Ledoit–Wolf shrinkage by 0.0088 (95% bootstrap CI 0.0026 to 0.0170; 7/8 subjects; exact two-sided sign-flip p=0.0156), close to the original report’s +0.008. Time-ordered inner folds gave +0.0084; across ten CV seeds, gains ranged from +0.005 to +0.010, all positive. Against pooled Ledoit–Wolf, tuned shrinkage gained 0.0206 (8/8; Holm-adjusted p=0.0312); pooled Ledoit–Wolf was below per-class Ledoit–Wolf in all subjects (-0.0118; adjusted p=0.0312), so the per-class versus pooled difference did not explain the gain. Pooled OAS was below pooled Ledoit–Wolf in all subjects (-0.0047; adjusted p=0.0312). Mean test AUCs were 0.835 (tuned), 0.826 (per-class Ledoit–Wolf), 0.814 (pooled Ledoit–Wolf), and 0.809 (OAS). With 30 calibration trials, tuned shrinkage was not better (-0.0126; 3/8; adjusted p=0.0938). The evidence is limited to eight subjects, one task and session, offline analysis, small effects (about 0.01–0.02 AUC), and one shrinkage grid. Features were archived from an earlier SCI contribution and were not recomputed from raw EDF data. In these data, larger shrinkage improved transfer to later trials when about 100 calibration trials were available; the report does not establish broader generalization.

linear discriminant analysis · shrinkage · Ledoit-Wolf · OAS · cross-validation · SEEG · AUC · replication