A 12-trial confidence update lowered Brier error versus an offset but had higher error than a class-rate constant in 15 SEEG session tests at C = 0.1 under assumed timing

We tested whether 12 labeled trials in a later session can improve the probabilities from a fixed SEEG reader. We used public OpenNeuro ds004752 version 1.0.1 and first and last numbered sessions from 15 patients. The reader used eight fixed bipolar pairs and 24 power features. We assumed fixation at epoch seconds 0 to 1 and delay at 4 to 6. Training used 607 trials. Calibration used 180 early trials. Testing used 450 separate later trials. At reader C = 0.1 and calibration penalty 2, a bounded slope and offset gave mean Brier error 0.220028. An offset alone gave 0.236490. The paired change was -0.016462, with a descriptive 95% patient bootstrap interval of -0.027299 to -0.007158. However, an early class-rate constant gave lower error, 0.210542. A fixed half slope gave 0.217827. Shuffled early labels gave mean error 0.219967. Thus, the probability error decrease did not establish better signal reading. The gain was much smaller at C = 0.01. The cohort and test rows have prior use. These are offline task tests, with assumed timing. The package preserves code, data inputs, predictions, figures, source records, full signal-window checks, and repeat instructions. The study tested no live control or stimulation.

probability calibration · brain-computer interfaces · SEEG · Brier score · session transfer · label shift