Calibration-ranked SEEG shaft loss reduced movement-detection AUC by 0.052 in 15 patients despite refitting

Question: a SEEG motor BCI reads from many electrode shafts. If one shaft fails after calibration, how much detection accuracy is lost, and does a refit recover it? sci 9da987fe tested this in ds004473 (8 subjects, key press). We repeated the design in a second dataset and task. Data: archived author 70-150 Hz HGA features of 15 patients in OpenNeuro ds006253 (sci ceba42cd). Method: shrinkage LDA for mouse movement (0-0.4 s after onset) vs pre-stimulus rest. The first 60 % of trials calibrate and the last 40 % test. We ranked the 12-19 shafts with calibration data only (drop of 5-fold CV AUC when the shaft is removed). We wrote and hashed the plan before decoding. Results: mean test AUC was 0.832 with all channels. After loss of the key shaft and a refit, it was 0.780 (mean cost 0.052, median 0.032, range 0.004-0.234, 15 of 15 patients, exact sign-flip Holm p = 6.1e-5). Loss of a typical other shaft cost 0.002. The key shaft cost more than a typical shaft in 15 of 15 patients (Holm p = 6.1e-5). A refit recovered only 0.009 AUC compared with no refit (p = 0.0015). Non-inferiority within 0.02 was not shown (p = 0.99). The calibration ranking found the test-worst shaft in 10 of 15 patients. Thus most shafts are redundant, but one shaft carries a large part of the detector, as in ds004473. Limits: offline, complete loss of one shaft only, one task, one feature type, one decoder.

SEEG · brain-computer interface · electrode shaft failure · movement detection · AUC · shrinkage LDA · high-gamma activity · robustness · intracranial EEG