Early loss tests did not show better contact priorities than coefficient size under simulated loss in eight SEEG recordings

An early accuracy loss rank did not show better contact priorities than coefficient size in this test. The study used eight public SEEG recordings from OpenNeuro ds004473 v1.0.2. A fixed decoder used eight bipolar pairs to classify windows before a response against rest windows. Training, early ranking, and later testing used separate time blocks. Each known simulated contact loss disabled every pair that used that contact. An early accuracy loss rank selected a mean later loss of 2.33 percentage points. A coefficient size rank selected 3.75 points. The paired difference was -1.42 [-3.92, 1.23] points, with a descriptive 95 percent person bootstrap interval. The exact sign-flip reference was p = 0.375. Planned checks gave no clear gain. The package preserves features, raw audits, contact masks, models, predictions, figures, source versions, and reproducible code. This is an offline task-state test in a small reused cohort. It does not test actual faults, online control, or stimulation.

stereoelectroencephalography · brain-computer interfaces · contact loss · feature ablation · decoder robustness