A random-subspace LDA ensemble had lower AUC than full shrinkage LDA after random SEEG channel loss

We compared a 200-member random-subspace shrinkage LDA ensemble with one shrinkage LDA for offline cued key-press versus rest decoding in 8 SEEG subjects from OpenNeuro ds004473. Models used archived high-gamma (70–150 Hz) and beta (13–30 Hz) bipolar power features, trained on trials 1–100, and tested on trials 101 onward. After random loss of 25% of channels, the ensemble dropped members that used lost channels and had mean AUC 0.780, versus 0.808 for the single LDA with lost features set to the calibration mean (difference −0.028; 7 of 8 subjects lower; exact sign-flip p = 0.016). Without channel loss, the ensemble also had lower AUC (0.794 versus 0.828; p = 0.008). For the single LDA, channel loss reduced AUC by 0.020, and retraining recovered 0.011. In this task, one shrinkage LDA with mean imputation was more robust than the tested ensemble. This result is limited to one dataset, one offline task, 8 subjects, one session, 0.5 s windows, random independent channel loss rather than full-shaft loss, and one ensemble design.

SEEG · brain-computer interface · LDA · random-subspace ensemble · channel loss · AUC · high-gamma power · beta power