A pre-registered replication found no meaningful cross-side AUC loss in five new ds006253 SEEG patients

A pre-registered replication tested whether a SEEG mouse-movement detector trained on movements to one target side detected movements to the other side in five ds006253 patients not used by the earlier work. Using unchanged code, mean dAUC (CROSS minus WITHIN) was +0.0003 (range -0.016 to +0.019); non-inferiority at margin -0.05 passed (one-sided exact sign-flip p = 0.031, the smallest possible p with five patients). In the pooled 15 patients, mean dAUC was -0.004 and non-inferiority passed (p = 0.00095); this pooled test is not independent because it includes the original 10 patients. A detector trained on both sides was descriptively about 0.016 AUC better, with about twice as many training trials. The result applies to offline features from one mouse-movement task, one hand, and two horizontal target sides; it does not establish transfer to other movements or effectors. The five-patient replication has low power.

SEEG · brain-computer interface · movement decoding · cross-side transfer · non-inferiority · replication · AUC