Conditional-mean imputation matched refitting after simulated SEEG channel loss but recovered only 14% of key-shaft AUC loss

This replication tested three preregistered hypotheses about conditional-mean imputation (CMI) after simulated loss of stereo-electroencephalography (SEEG) channels, using 70–150 Hz bipolar power features from 15 patients in OpenNeuro ds006253 performing mouse movement versus rest. A shrinkage linear discriminant decoder used a time-ordered 60/40 train/test split. All three exact one-sided sign-flip tests passed after Holm correction: after loss of the key shaft, CMI exceeded mean fill by 0.0080 AUC (95% bootstrap CI 0.0025 to 0.0157; 13/15 patients; p=0.0018); CMI was non-inferior to a labeled refit within a 0.01 AUC margin (-0.0011; CI -0.0023 to -0.0001; 15/15; p=0.00009); and CMI was non-inferior after random loss of 25% of pairs (+0.00005; CI -0.0002 to +0.0003; 15/15; p=0.00009). Mean AUC with all pairs was 0.836. However, CMI recovered only 14% of key-shaft AUC loss: losses versus full-channel performance were -0.059 for mean fill, -0.051 for CMI, and -0.050 for refitting. Two patients had large key-shaft losses (sub-09: -0.232 with mean fill and -0.220 with CMI; sub-23: -0.200 and -0.192). The results support continuation without new labeled trials in this offline setting, but show that neither method recovered most of the information lost with the key shaft. Limits include simulated loss, one session per patient, same-session testing, one center, one band, one binary contrast, and a linear decoder.

conditional-mean imputation · stereo-electroencephalography · brain-computer interface · channel loss · linear discriminant analysis · intracranial EEG · replication