A fixed early SEEG model increased mean decision-time test error across trial blocks in 20 patients

We tested whether early brain signals improve decision-time prediction beyond stimulus strength and recent behavior. We used public MetaRDK SEEG derivatives from OpenNeuro ds006253 version 1.0.3. The fixed test used 3,604 trials from 20 patients with epilepsy. Decision times were 0.8 to 6 seconds. Each signal input was a channel mean from 100 to 500 ms after stimulus onset minus a mean from 400 ms before onset to onset. We held out five separate trial blocks within each patient. Error was squared error in log decision time. The mean patient error increased by 1.19%. Its 95% patient bootstrap interval was 0.04% to 2.32%. 6 patients had lower error. The exact patient sign-flip reference value was 0.063. Both models had higher mean error than a training-mean reference. The source signals were normalized and smoothed before release. This is an offline test in the same patients and sessions. It does not show a real-time predictor or a cause of decision speed. The package contains source versions and hashes, compact features, all predictions, fitted models, independent checks, and repeat instructions.

stereoelectroencephalography · decision time · high-gamma activity · prediction · cross-validation · confidence