Adding high gamma lowered later trial pair loss after removal of a shaft selected from training data in eight SEEG recordings

We tested whether high gamma still improved paired task window identification after one electrode shaft was removed. We used all 2,030 aligned trials from eight patients in public OpenNeuro ds004473, version 1.0.1. The first half of trials trained each model. The later half tested it. Training trials selected the shaft with the largest high gamma score contribution. We removed both bands from that shaft and fitted the models again. Adding 70–110 Hz power to 13–30 Hz power lowered test loss in every patient. Mean loss fell from 0.5017 to 0.4502 nats per pair. The mean gain was 0.0515 nats, with a descriptive 95% patient bootstrap interval of 0.0271–0.0804. The exact two-sided sign-flip probability was 0.0078125 under independent patient effects and a symmetric zero null. The gain was 62.3% of the full mean gain. Every single shaft removal at penalty 1 also gave a positive gain in each patient. All planned sensitivity checks had positive mean gains. One patient had a negative gain with no shared contacts and with a weaker penalty. Fresh source retrieval reproduced all feature arrays and all 4,076 selected range hashes. Separate numerical checks and a complete repeat passed. The package preserves inputs, code, coefficients, predictions, source versions, hashes, checks, figures, and repeat instructions. This secondary analysis reuses the same patients and later trials as prior work. It tests a fixed paired model within these recordings. It does not isolate neural sources or establish clinical transfer.

stereo-EEG · high gamma · beta power · electrode shaft removal · paired window model · epilepsy · signal coverage