SEEG readers with eight inputs had greater mean target loss than uniform predictions during late grasps in four people

This study tested whether calibration on one grasp could predict three target levels during another grasp. It used public SEEG data from four people in NEMAR nm000397 version 1.0.0. Blocks 1 and 2 supplied 236 training trials. Later blocks supplied 236 test trials. Each reader used eight amplifier inputs and four power bands during a late one-second window. With logistic regression at C = 0.1, mean class-balanced log loss was 1.386 with same-grasp training and 1.378 with other-grasp training. Uniform predictions gave 1.099. Mean balanced accuracy was 31.6% and 33.9%; the reference was 33.3%. The training-grasp loss difference was -0.008, with a descriptive four-person interval of [-0.197, +0.223]. A force-sensor control reached 80.6% balanced accuracy. Stronger regularization reduced excess loss. This result limits the tested reader and small calibration set. It does not show that SEEG lacks force information. The package has four figures, source features, raw audit trials, provenance, fixed models, predictions, code, and repeat instructions.

stereo-EEG · brain-computer interface · grasp force · target-level decoding · cross-grasp transfer · logistic regression · class-balanced log loss · balanced accuracy · small-sample calibration