A fixed SEEG model gave 48.5% mean balanced accuracy for chosen side on incorrect decisions in 17 patients

We tested whether a fixed SEEG model trained on correct decisions identified chosen side on incorrect decisions. We used public MetaRDK high-gamma signals from OpenNeuro ds006253, version 1.0.3. Five separate trial blocks supplied test trials. The primary model used all stored channels, nominal 0.4 s windows before first movement, and ridge penalty 1. In 17 patients and 790 incorrect decisions, mean balanced accuracy was 48.50%. The 95% patient bootstrap interval was 44.10% to 52.75%. A conditional permutation test gave p = 0.450. It assumes exchangeable labels within each patient, trial block, and outcome group. Correct-trial accuracy was 52.19%, with an interval from 49.28% to 55.24%. The fixed test did not detect choice or stimulus transfer. Low correct-trial accuracy limits that conclusion. A preset window after movement gave 53.74%, with an unadjusted interval from 50.11% to 57.20%. This secondary result needs a separate test. We reused compact features from an earlier AutoSci confidence study. This is a new target test on the same cohort. The package preserves the local plan, code, features, source versions and hashes, predictions, models, bootstrap and null scores, source checks, figures, and repeat instructions. Clinical sampling, inherited trial rules, and source preprocessing limit its scope.

stereoelectroencephalography · decision making · choice decoding · high-gamma signals · balanced accuracy · ridge regression · conditional permutation · reproducibility