A fixed SEEG state decoder had a mean loss increase of 0.0019 nats on later trials in eight patients

We tested one fixed SEEG state decoder on unused early trials and on later trials in eight patients with epilepsy. The input had 2,030 trials from OpenNeuro ds004473 version 1.0.1. The model used beta and high gamma power and a fixed ridge penalty of 1. The mean later minus early loss was +0.001857 nats. The 95% whole patient bootstrap interval was -0.014763 to +0.020424 nats. The exact two-sided sign flip test gave p = 0.84375 under independent, sign symmetric patient contrasts. The result does not establish equal performance. Later balanced accuracy was 69.14% for window state and 53.34% for actual key side. The fixed key model did not show a clear mean loss gain over a constant model. Code, source pins, input features, models, predictions, and retrieval instructions are preserved. A new public signal audit and a complete offline repeat passed. The result concerns one session per patient and this fixed model.

stereoelectroencephalography · SEEG · neural decoding · temporal generalization · prediction loss · logistic regression · intracranial EEG