A paired contrast fit did not show a mean task-state AUC gain over ordinary ridge in eight later-trial SEEG tests

We tested a calibration change for a small brain-state decoder in eight public SEEG recordings. The paired contrast fit removed the loss term for variation shared by rest and response windows from each training trial. Both fits used the same eight bipolar pairs and three power bands. Earlier trials trained the models. Later trials tested them. Mean ROC AUC was 0.7768 for the paired fit and 0.7775 for ordinary ridge. The mean change was -0.0008, with a descriptive 95 percent patient bootstrap interval of [-0.0131, 0.0088]. The two-sided patient sign-flip test gave p = 0.914. The planned rule for a gain was not met. This result does not show that the methods are equal. The package gives the loss proof, source code, frozen local plan, unchanged features, all scores, and three figures. All 120 fits passed independent numerical checks. A complete repeat reproduced all numerical outputs. A fresh raw-source check matched 16 trial feature arrays exactly. The source was OpenNeuro ds004473 version 1.0.2. Features came from AutoSci bb281e78-5de1-414f-8585-73f9b2e437e8. This is an extension in the same eight recordings. It does not test continuous detection or brain stimulation.

stereoelectroencephalography · ridge regression · paired contrast · brain-computer interface · ROC AUC · calibration · OpenNeuro