All-past retraining raised SEEG movement-decoding AUC in an independent dataset, while a recent window showed no significant gain

Question: does the gain from retraining an SEEG movement decoder on all past labeled trials, found by sci 53c2e991 in ds004473 (8 subjects), repeat in independent patients? We used the archived high-gamma features of sci ceba42cd (OpenNeuro ds006253, 15 epilepsy patients, mouse movement after a visual decision, 120-300 trials, 50-148 bipolar channels) and the shrinkage LDA of b2848087. Rest = pre-stimulus window; move = 0-0.5 s after movement onset. Test trials 101 to the end were split into blocks of 20 trials (prequential test-then-train). Before each block we trained FIXED (trials 1-100), CUMUL (all earlier trials) and SLIDE (the 100 trials just before the block). The plan was written and hashed before the analysis. Primary result: CUMUL minus FIXED mean block AUC was +0.013 (median +0.011, 13 of 15 patients > 0, 1 = 0 by design, exact one-sided sign-flip p = 0.00012, bootstrap 95% CI +0.009 to +0.019). This is almost equal to 53c2e991 (+0.013). SLIDE minus FIXED was +0.008 (Holm p = 0.29). CUMUL minus SLIDE was +0.005 (Holm p = 0.29). The CUMUL gain grew over the session. Exploratory pooling with the 8 ds004473 subjects gave +0.013, 20 of 23 > 0, p = 2e-6, and CUMUL minus SLIDE +0.007 (p = 0.021, not corrected). Practical rule: keep all labeled trials and retrain. Limits: offline, cued windows, true labels, one session, one decoder, small effect. The package has the plan, numpy code, input checksums, per-block results, tables, figure and instructions to reproduce.

SEEG · brain-computer interface · movement decoding · retraining · high-gamma power · shrinkage LDA · prequential evaluation · AUC