A tuned gamma model did not show a gain beyond known feedback for model error size in 16 SEEG patients

We tested whether gamma features helped estimate prediction error magnitude beyond known task feedback. The public SEEG derivative had 16 patients, 248 bipolar contacts, and 7,679 trials. The target was the absolute value of the released learning model PE field. A task-only baseline used four context and outcome means. Two gamma means per contact entered a ridge model with separate context slopes. Three forward validation folds in the first source-row half selected the ridge strength. The remaining rows supplied test losses. Mean patient test MSE was 0.06133849 for known feedback and 0.06138980 with gamma. Mean gain was -0.00005131, with a 95% patient bootstrap interval of [-0.00028199, +0.00016586]. The exact sign-flip p was 0.716796875. The method did not show a gain. A descriptive context-only comparison showed a larger gamma gain of 0.00157544. Thus, the task-only baseline changes the claim about added brain information. All 86 required source files passed size and hash checks. A separate calculation verified all 96 models. Fresh source extraction rebuilt all 208 input arrays exactly. The package preserves code, provenance, targets, models, predictions, three figures, and repeat instructions. The target is an offline model variable. Source row order and normalization are not fully known. The cohort and test rows were used before. Eight source PE signs in one patient prompted a stated post hoc exclusion check; its result was unchanged. These results do not exclude useful SEEG information or establish online BCI or stimulation benefit.

stereoelectroencephalography · SEEG · gamma activity · prediction error · reinforcement learning · known feedback · ridge regression · brain-computer interface · patient-level analysis · reproducibility