A fixed model had higher loss with post-stimulation than pre-stimulation power changes in 90 SEEG records from 17 people

Can SEEG power changes help read the temporal visual effect of brain stimulation in a new person? We tested continuous versus intermittent visual reports in OpenNeuro ds005169 version 1.0.0. The test used 90 records from 17 people. Each model excluded the test person from training and scaling. The fixed logistic model used power summaries from four frequency bands. Mean person Brier loss was 0.3224 with post-stimulation changes and 0.2948 with pre-stimulation changes. Lower loss is better. The difference was +0.0276. A training-label frequency model had loss 0.2087. Eleven people had higher loss with post-stimulation changes. Complete refits after deletion of each person gave differences from -0.0026 to +0.0642. The direction changed in a 43.2 Hz subset. Only six people had both labels. An additional diagnostic showed that a fold-specific label prior had pooled AUC 0.0572, although its within-person AUC was 0.5. This warns against a neural interpretation of pooled fold ranks. The evidence does not support these fixed power summaries as a general symptom monitor. It does not rule out other features. The package includes spectra, labels, source hashes, fitted models, predictions, figures, code, and retrieval instructions.

stereoelectroencephalography · brain stimulation · visual symptoms · cross-person prediction · power spectra · Brier loss · held-person validation · intracranial EEG