A pre-pulse energy slope learned in other patients gave no clear gain in 20–500 ms SEEG error forecasts after five calibration pulses in eight patients

We tested whether SEEG energy that ends 500 ms before a pulse can improve a forecast of response waveform error after five calibration pulses. The public OpenNeuro data cover eight patients, 86 sites, and 886 remote bipolar readouts. Each model was fitted in seven patients and tested in the remaining patient. The learned energy slope reduced mean squared log forecast error by 0.43% against a constant correction. Five patients improved. The mean loss difference was -0.00263; its descriptive 95% patient interval was -0.00938 to +0.00365. This test showed no clear gain. A fixed slope of one increased loss in all eight patients. Learned slopes were 0.0468 to 0.1000. Fixed secondary checks and a pulse timing check also gave no clear gain. The test used 430 later events and retrospective quality labels. The intervals omit training variation. The package preserves exact inputs, all forecasts, figures, source hashes, a fresh source rebuild, code, and repeat instructions. No live interface or stimulation policy was tested.

SEEG · electrical stimulation · pre-pulse energy · error forecast · patient-held-out validation · squared log loss · limbic system · epilepsy