Under assumed timing and simulated loss, shaft-loss training at C=1 reduced Brier error in 15 SEEG tests. A rate forecast had less error.

We tested a memory-load decoder when one electrode shaft was missing. We trained it with simulated shaft losses. We used archived features from OpenNeuro ds004752 v1.0.1 for 15 patients. Each model used eight adjacent SEEG contact pairs. The main test used 683 first-session trials and 703 last-session trials with the source quality flag set to zero. At fixed C = 1, shaft-loss training reduced mean Brier error under simulated single-shaft loss from 0.3067 to 0.3002. The paired mean change was -0.00647; the 95% patient bootstrap interval was [-0.01022, -0.00279]. Ten of 15 patients had lower error. A training-rate forecast had lower error, 0.2332. A stronger fixed penalty also reduced error more than shaft-loss training at C = 1. Missing features reduced error even without shaft-loss training. These results do not establish a useful memory-load decoder or a specific gain from shaft grouping. The package provides a reproducible stress test, source hashes, model states, predictions, and figures. The epoch timing is assumed. No physical failure, online control, or brain stimulation was tested.

stereo-EEG · brain-computer interface · memory load · shaft loss · robustness · Brier score