Training for simulated fixed 0.8–1.25 pair gains reduced worst loss and mean clean accuracy in eight SEEG readers

We tested gain training for a brain signal reader with public OpenNeuro SEEG data from eight people. The labels were rest and before response. Readers learned from 1,332 early trials and were tested on 674 later trials from the same recordings. Separate simulated pair voltage gains ranged from 0.8 to 1.25. Each gain was constant across bands, samples, and later test buffers. Mean worst gain loss decreased from 1.239 to 0.611 nats per buffer. All eight people had lower worst loss. The paired mean change was -0.628, with a descriptive 95% person interval from -0.877 to -0.414. Mean accuracy on the original signal decreased from 70.7% to 67.0%. Gain training removed almost all sensitivity to common power in each pair. A training-only check put the gain bounds for this transition between 1.021 and 1.092. That check was added after the main results. The package includes exact signal buffers, source records, code, fits, predictions, figures, and successful independent checks. This is an exploratory test in a reused cohort. It does not test real gain faults, new sessions, continuous commands, or brain stimulation.

stereoelectroencephalography · signal processing · gain robustness · logistic regression · spectral power · brain interfaces · worst-case loss