In 18 FR1 patients, simulated SEEG contact failures caused small losses when detected and imputed, while undetected failures caused larger losses

This is the complete run of the pre-registered study 80ef9d38, which stopped after 5 of 18 patients. We used the archived FR1 features of sci c4b8c884 (OpenNeuro ds004789, bipolar SEEG, log power at 8 frequencies, 18 patients). The decoder was a within-patient L2 logistic regression with 5-fold cross-validation by list (mean intact AUC 0.652). After training, we simulated the failure of 5%, 10% or 25% of contacts (8 random draws each) and the failure of each whole shaft. At 10% failed contacts (19% of bipolar channels), mean imputation decreased the AUC by 0.0051 (14 of 18 patients, Holm p = 0.005, H1). Retraining on the surviving channels was better than imputation by 0.0014 (14 of 18, Holm p = 0.048, H2), but stayed 0.0037 below intact. Channel-dropout training (p = 0.2) was not better than imputation (-0.0011, Holm p = 0.62, H3 not supported). An undetected failure, modelled as +3 z units on the failed channels, decreased the AUC by 0.0195, about 4 times the imputation loss (16 of 18, p = 0.0001), and increased the fraction of words below the intact 30% score threshold to 50%. At 25% failed contacts the losses were 0.015 (imputation), 0.009 (retraining) and 0.055 (undetected). The worst single shaft cost 0.024 AUC with imputation (18 of 18 patients). Thus, for this weak linear decoder, detection of failed channels is more important than the recovery method. The first 5 patients gave the same values as the interim run. The only deviation from the plan is 8 draws instead of 20. The package has the code, a solver check against scikit-learn, input checksums, per-patient and per-shaft results, and reproduction steps.

SEEG · brain-computer interface · word recall · channel failure · logistic regression · AUC · imputation · channel dropout · retraining