SEEG sets at a 90% trial coverage target gave a mean output yield of 34.2% and a mean accepted error of 17.3% in eight later-trial tests

A brain interface can hold an output when its label is uncertain. We tested this rule with public SEEG data from eight patients in OpenNeuro ds004473 version 1.0.2. Each trial gave one rest window and one window before a response. We fitted an eight-pair decoder on early trials. We used a separate earlier block to set conformal thresholds. We tested fixed sets on 674 later trials. At the 90% target, whole-trial sets contained both true labels in 91.1% of test trials on average across patients. The sets gave one label in 34.2% of windows. Mean error among these accepted outputs was 17.3%. Separate class bounds with a split error budget gave one label in 30.7% of windows and had 92.2% trial coverage. Whole-trial sets increased output yield by 3.5 percentage points and reduced observed coverage by 1.1 points. Four of eight patients had trial coverage below 90%. The conformal bound requires exchangeable trial scores. Task mixtures changed across the time blocks, so this assumption is not established. Trial coverage does not bound error among accepted outputs. These results apply to event-defined, quality-selected windows in one recording per patient. They do not measure live command control or brain stimulation. The package has exact features, signal samples, source hashes, model states, all trial scores and sets, three figures, and code for reproduction.

stereoelectroencephalography · conformal prediction · brain-computer interfaces · trial coverage · selective classification · epilepsy