Other patient gamma slopes reduced mean one second forecast error after 16 local trials per channel in SEEG data from 13 patients under assumed trial order

We tested whether gamma slopes from other patients improve a power forecast after a small local fit. We used OpenNeuro ds006136 version 1.0.1, with 13 patients, 14 records and 143 released bipolar channels. Each target channel used its first 16 accepted trials for calibration. Donors used early trials from the other 12 patients. Both readers used the same local scale and half of the local slope. The donor reader added half of the donor slope. The comparison reader added zero. The test used a one second gap between separate power windows and the final 35 percent of accepted trials per channel. Mean normalized squared error fell from 1.206674 to 1.199547, a 0.59 percent reduction. Ten of 13 patient scores improved. The mean gain was 0.007127. The descriptive 95 percent patient bootstrap interval was 0.001924 to 0.012826. Patient scores share donor fits, so this interval is not an independent population test. More local training data gave lower mean error than the donor reader. The source files passed hash checks, and a full source rebuild matched the feature table. The package preserves code, inputs, source recipes, model values, predictions, controls and four figures. Source trial IDs are assumed to show time order. The released records are labelled SEEG, but contact type is not resolved for every channel. The result concerns processed gamma power. It does not establish intent decoding, live operation or a stimulation effect.

stereo-electroencephalography · gamma power · forecasting · transfer learning · calibration · object working memory · intracranial EEG