Diagonal LDA was not non-inferior to shrinkage LDA for SEEG key-press detection with high-gamma + beta features

In eight subjects performing an offline cued key-press task, diagonal linear discriminant analysis (LDA) was not shown to be non-inferior to Ledoit-Wolf shrinkage LDA for detecting movement versus rest from high-gamma plus beta SEEG log-power features. Classifiers trained on trials 1–100 and tested on trials 101 onward had a mean AUC difference (diagonal minus shrinkage) of -0.023 (median -0.030); the exact one-sided non-inferiority p-value was 0.645 for a 0.02 AUC margin. Diagonal LDA had lower AUC in seven of eight subjects. Its analytic calibration cost was about 100 times lower in multiply-adds and about 40 times lower in stored numbers. Descriptive secondary analyses found that the decoders were closer with high-gamma features alone and with 20 calibration trials, while diagonal LDA had lower AUC in all eight subjects in five-fold cross-validation within the later trials. These are offline results from one feature set and one session; compute costs were not measured on implant hardware, and the task was not a self-paced closed-loop BCI.

SEEG · brain-computer interface · linear discriminant analysis · non-inferiority · high-gamma · beta power · decoder calibration · AUC