A fixed donor reader gave lower probability error than a matched local reader on later blocks in nine amygdala SEEG runs

We tested whether data from other people can replace the chosen local calibration for a fixed brain reader. We used nine public amygdala SEEG runs from OpenNeuro ds003374, version 1.1.1. Each target reader used early video blocks from eight other people and no target labels. A matched local reader used the target person's early blocks. Both readers used the same five baseline-relative power features and the same later test blocks. The donor reader had lower Brier probability error in all nine people. Mean errors were 0.2459 and 0.3003. The donor-minus-local difference was -0.0544, with a descriptive person-resampling interval of [-0.0767, -0.0129]. Removing each person kept this group difference negative. The donor error was close to the constant-probability reference of 0.2500. Its difference from that reference had an interval of [-0.0226, +0.0394]. Thus better error than this local fit did not establish useful reading. Labels alternate and describe fearful-face or landscape videos. They do not measure emotion or intended action. The package preserves four figures, code, source pins and hashes, checked download instructions, all features, main fits and probabilities, and 2,000 refitted person resamples. A separate replay reproduced 22 generated files byte for byte.

amygdala · stereoelectroencephalography · brain-computer interface · cross-person transfer · probability calibration · Brier score · fearful-face videos · landscape videos