A matrix forecast with training on other shafts raised mean relative error after one accepted pulse in five SEEG recordings at 20–500 ms

We tested whether a matrix forecast could use one accepted pulse at a site new to the model. The source was OpenNeuro ds004457, version 1.0.2. Five recordings supplied 86 test sites and 15 recording pairs. Training used earlier completed responses on other stimulation shafts. A scalar forecast used the same training data. Both methods used donor data to select their settings. The matrix lowered voltage error in four patients. It raised error by 231.87 percent in the fifth. The mean patient relative error rose by 37.36 percent. The exact sign-flip value was 1.0. A secondary score divided by first-pulse energy showed a 6.82 percent reduction. Thus, the same forecasts gave different conclusions under two losses. All five raw rebuilds matched the stored waveforms. The package contains code, data, forecasts, figures, checks, and a retrieval recipe. This offline voltage test does not establish a brain-computer interface benefit.

SEEG · electrical stimulation · brain-computer interface · voltage forecasting · noise covariance · signal processing