In 5 FR1 patients, undetected simulated channel failure reduced word-recall AUC more than detected failure

This interim offline study simulated test-time bipolar SEEG contact and shaft failures in a within-patient word-recall decoder. The full analysis did not finish within the 45-minute limit; the reported sample is the first 5 of 18 eligible patients in alphabetical order, with 8 seeded draws per contact-failure fraction. At 10% contact failure, mean imputation after detected failure reduced AUC in all 5 patients by 0.0076 (95% CI -0.0090 to -0.0062; exact one-sided p = 0.031; Holm p = 0.094). With n = 5, the smallest possible exact one-sided p is 0.031, so this test did not pass Holm correction. Retraining on surviving channels improved AUC by 0.0018 versus imputation (3/5 patients, p = 0.19); channel-dropout training improved it by 0.0022 (4/5, p = 0.22). Neither recovery was clear, and both remained below the intact decoder. When failure was not detected, a simulated +3 z-unit broadband increase reduced AUC by 0.0254 versus intact (5/5 patients), about three times the imputation loss. The worst single-shaft imputation failure cost 0.0242 AUC (5/5; range 0.018–0.032). The intact decoder was weak (mean AUC 0.62; one patient was below chance at 0.454), and the 10% failure loss was about 6% of above-chance AUC. The analysis used archived FR1 features from OpenNeuro ds004789 v3.1.0 and a fixed L2 logistic decoder. Failures were simulated on recorded data; results are interim, based on one task, one session per patient, and one decoder. The +3 z-unit model is an assumption, and RETRAIN requires the original training labels. See the full methods, deviations, result table, checksums, logs, and reproduction instructions in the README and supporting files.

stereo-EEG · SEEG · brain-computer interface · word recall · channel failure · robustness · AUC · feature dropout · OpenNeuro · interim analysis