Two tuned SEEG readers gave mean log losses of 0.906 and 0.908 in eight later-trial tests

We tested rest, left-key, and right-key probability readers in eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Each reader used the same eight bipolar pairs and 50 fitted parameters. It learned from early trials. We tested 674 later trials from the same recordings. The mean joint log loss was 0.9055 nat for one three-class reader and 0.9082 nat for a task reader followed by a key reader. The paired difference was -0.0027 nat. Its 95 percent person-bootstrap interval was -0.0106 to +0.0058 nat. Fixed checks changed the sign of the small mean difference. Task detection gave 98.6 percent of the two-head reader's mean gain over a constant reference. The key head added 0.0019 nat of mean gain, with an interval that included zero. We provide an exact task/key error decomposition, raw-byte checks, code, inputs, model states, all test probabilities, and four figures. The cohort and later trials were used in earlier work. These selected response-aligned windows do not establish live command control, transfer to a new day, or an effect of brain stimulation.

stereoelectroencephalography · brain-computer interface · probability calibration · log loss · task detection · key decoding