The interval for channel mean removal gain included zero under independent patient resampling in fixed later recall tests from seven patients with reported SEEG

We tested whether removal of each channel's training encoding mean improves later recall count prediction when available channels change. We used public OpenNeuro ds006136 version 1.0.1. The data contain 85 released channels and 1,415 accepted trials from seven patients with reported SEEG. Fixed models scored 470 later trials. Mean patient squared error was 0.507211 for raw power and 0.503028 after channel mean removal. The gain was 0.004184, or 0.82 percent. The descriptive 95 percent patient interval was -0.013499 to 0.018693. Four patients improved. Three patients worsened. Patient resampling assumes independent patients and fixed fitted models. The evidence does not establish a clear gain. A training-mean predictor had lower observed mean error than either power model. We also give an exact power decomposition and checked source code for changes in channel support. All seven source EDF files passed hash checks. A second reader checked all input power values. The package preserves source pins, data, predictions, models, checks, and reproduction instructions.

stereoelectroencephalography · working memory · recall prediction · channel mean removal · power decomposition · missing channel support