Score integration gave lower mean Brier error than a Gaussian class reader after simulated shaft loss in eight SEEG recordings with 64 calibration trials

We tested probability readers after known simulated loss of one electrode shaft in eight public SEEG recordings. Each reader used the same eight contact pairs and 64 early calibration trials. The test used 979 later trials after the task condition changed. Mean Brier error was 0.20892 with Gaussian score integration and 0.22299 with a Gaussian class reader. The class reader had higher error in seven people. The planned integrated-minus-class difference was -0.01407. Its descriptive 95% person interval was -0.01936 to -0.00717. Mean balanced accuracy was about 66.6% for both readers. The class reader also had higher intact error. The difference in fault-induced error increases had an interval that included zero. Thus, the total error result does not show worse fault handling by the class model. This is an exploratory offline comparison in a reused clinical cohort. It does not test live commands or stimulation. The package preserves all features, fitted models, masks, probabilities, raw audit bytes, source hashes, code, and two figures. A replay matched all nine numerical files exactly. Fresh pinned audit ranges passed for all eight people.

stereoelectroencephalography · probability calibration · Brier error · Gaussian score integration · shaft loss · brain-computer interfaces