Early power-law learning curves did not improve 100-trial SEEG decoder forecasts over cross-validation in eight subjects

We tested whether a power-law fit to the first 40 calibration trials could predict the held-out area under the ROC curve (AUC) of a stereo-electroencephalography (SEEG) key-press decoder trained on 100 trials. In eight subjects from OpenNeuro ds004473 v1.0.2, a numpy shrinkage linear discriminant analysis decoder used high-gamma (70–150 Hz) and beta (13–30 Hz) log-power features. The target AUC was measured on trials 101 to the end. From trials 1–40, the power-law predictor had a mean absolute error of 0.085, compared with 0.083 for five-fold cross-validation; it was better for five of eight subjects (exact two-sided sign-flip p = 0.91). With 30 early trials, errors were 0.124 and 0.103, respectively (power law better for three of eight; p = 0.43). The held-out AUC gain from 40 to 100 training trials averaged 0.029 (range 0.011–0.047). Cross-validation on trials 1–40 differed from the held-out AUC of the same 40-trial decoder by a mean absolute 0.077, consistent with early-to-late trial differences limiting forecasts in this dataset. The study covers eight subjects, one task, one decoder, and one session per subject; one subject had only 32 test trials. Other learning-curve forms and drift-corrected predictors were not tested.

stereo-electroencephalography · brain-computer interface · calibration · learning curve · power law · cross-validation · AUC · shrinkage LDA · trial drift