Adding an intercept to scale calibration reduced later log loss in five of eight SEEG recordings at an assumed 1% task rate

We tested an added calibration intercept on eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Earlier trials fitted a fixed reader. A middle block fitted scale calibration, with or without an intercept. Later trials tested both maps at an assumed 1% pre-response rate. Mean log loss was 0.051959 for scale only and 0.051863 with the intercept. The mean change was -0.000096 nats per window. Its descriptive 95% person bootstrap interval was [-0.000319, 0.000038]. Five recordings had lower loss and three had higher loss. The data do not establish a reliable benefit. A secondary fit at a supplied 1% rate gave lower loss in only two recordings after adding the intercept. The package preserves all trial features, fitted states, predictions, code, figures, and source retrieval records. A repeat run matched nine analysis files exactly. Fresh pinned raw ranges supported 37,530 feature checks. This is a test of offline task windows in a reused cohort. It does not measure live command control or brain stimulation.

probability calibration · log loss · brain-computer interface · stereoelectroencephalography · class-prior shift · rare task rate · SEEG