Loss of the calibration-ranked SEEG shaft reduced key-press detection after refitting in 7 of 8 subjects

We tested how loss of the most informative electrode shaft affected offline key-press detection in stereo-electroencephalography (SEEG), and whether refitting on the remaining shafts with the same calibration trials recovered performance. We analyzed archived log-power features (70–150 Hz and 13–30 Hz) from adjacent bipolar pairs in eight subjects from OpenNeuro ds004473 v1.0.2. A shrinkage linear discriminant decoder used trials 1–100 for calibration and trials 101 onward for testing. Shaft ranking used calibration data only. Removing the top-ranked shaft and refitting reduced mean test AUC from 0.828 to 0.785 (mean loss 0.043, median 0.048; 7 of 8 subjects; exact one-sided sign-flip p = 0.020). Non-inferiority within 0.02 AUC was not shown (p = 0.89). Without refitting, mean AUC was 0.762 (loss 0.067; 8 of 8; p = 0.004); refitting recovered 0.023 AUC (8 of 8; p = 0.004), about one third of that loss. Losing a random shaft with refitting cost 0.005 AUC on average (p = 0.004). Removing the top two and top five ranked shafts gave mean AUCs of 0.736 and 0.692. The key shaft alone gave mean AUC 0.766; it was the best single shaft in 5 of 8 subjects. These results indicate concentration of key-press information on one or two shafts in this task and show that refitting did not fully compensate for loss of the key shaft. The study is limited to offline analysis of a cued task, one session in each of eight epilepsy patients with clinical shaft placement, and one decoder and feature set; anatomical locations were not checked and secondary tests were not corrected for multiple comparisons.

stereo-electroencephalography · brain-computer interface · key-press detection · electrode failure · linear discriminant analysis · AUC · OpenNeuro ds004473