Label-free imputation matched refitting after known SEEG channel loss, but recovered only 16–22% of key-shaft loss in 15 patients

This is a pre-registered replication of sci 095ecc14 in a second dataset and task. We used the archived 70-150 Hz window features of 15 patients from OpenNeuro ds006253 (sci ceba42cd). A shrinkage LDA separated movement (0-0.5 s after mouse movement onset) from rest (0.5 s before the stimulus). We trained on the first 60 % of trials and tested on the last 40 %. The mean AUC with no loss was 0.833. On the test trials, we removed the key shaft (the shaft of the best training pair, 4-14 pairs) or a random 25 % of pairs. We replaced the lost features by the training mean (fill), by label-free conditional-mean imputation (CMI), or we refit the LDA with labels. Exact sign-flip tests with Holm correction supported all three planned hypotheses. For the key shaft, CMI was better than fill by +0.010 AUC (95 % CI 0.004-0.018; 14 of 15 patients; Holm p = 0.0014). CMI was non-inferior to refit (margin 0.01): key shaft +0.0001 AUC, random 25 % -0.0001 AUC (Holm p < 0.0001 for each). But key-shaft loss cost 0.061 AUC with fill and 0.051 with CMI or refit (maximum 0.211). Thus, a known channel failure can be repaired without new labels, but software does not replace the key shaft. Limits: offline, simulated known loss, one session, event-locked windows, HGA only, clinical placement.

SEEG · channel loss · conditional-mean imputation · shrinkage LDA · AUC · non-inferiority · neural decoding