A 40-trial labeled calibration block did not improve cross-patient SEEG error prediction in 11 patients (median AUC change -0.001, p = 0.56)

We tested whether a short labeled calibration block from a new patient improves a cross-patient predictor of working-memory errors from SEEG band power. Using archived features from OpenNeuro ds006136 v1.0.1 (MTG/MFG bipolar contacts), we trained on the first 10, 20, 30, or 40 trials and tested on trials after trial 40, assuming trial index is chronological. The cross-patient model had median AUC 0.606 (10/11 subjects above 0.5, p = 0.0049). Adding 40 calibration trials at 25% of training weight did not improve AUC (median change -0.001, mean -0.015, 4/11 higher, exact two-sided Wilcoxon p = 0.56); smaller blocks also did not significantly improve it. Models trained only on calibration trials had median AUC 0.500-0.545 and were lower than the cross-patient model in 8-9 of 11 subjects, without significant differences after Holm correction. MU4 was excluded by the prespecified test-set rule because it had 3 test errors, leaving 11 subjects. For this task and decoder, a cross-patient model can be used without a labeled calibration block. This offline result is limited to one dataset and task, assumes chronological trial order, uses session-wide unlabeled z-scoring, and tests one target weight (25%) and one regularization value (C = 0.1).

SEEG · working memory · brain-computer interface · calibration · cross-patient prediction · AUC · intracranial EEG · logistic regression