A fixed SEEG reader gave mean scores of 56.3% on random windows and 53.8% on later blocks in nine patients

We tested a fixed reader of video type with all nine patients in public OpenNeuro ds003374, snapshot 1.1.1. Each amygdala SEEG run supplied 320 one-second windows from 16 balanced video blocks. A random test put ten train windows and ten test windows in every block. A later block test used the first eight blocks to train and the next eight to test. Both tests had 160 train windows and 160 test windows per patient. Mean balanced accuracy was 56.3% for random windows and 53.8% for later blocks. The mean gap was 2.47 percentage points. The exact paired test gave p = 0.113, so the sample did not give clear evidence for a group difference. Labels have fixed order and distinguish fearful faces from neutral landscapes. They do not measure emotion or intended action. The package includes a plan made before signal values were read, checked source hashes, metadata, derived features, split masks, reader parameters, code, and four figures. A separate package copy reproduced the features and main results from the full source signals. The study tests offline reading; it does not test stimulation.

SEEG · brain-computer interface · balanced accuracy · validation design · amygdala