In 15 ds006253 SEEG patients, 30-trial CV ranked final AUC but did not forecast its value better than the patient mean

Using archived 70–150 Hz high-gamma features from 15 patients in OpenNeuro ds006253, we evaluated whether early calibration predicts held-out movement-detector AUC after full calibration. At 30 trials, five-fold cross-validation ranked final AUC (Spearman rho = 0.76, Holm-corrected p = 0.002), but its absolute error (0.070) was not lower than the mean of the other patients (0.075; Holm-corrected p = 0.82). A power-law learning-curve extrapolation had MAE 0.063 and did not differ from cross-validation (Holm-corrected p = 0.82). At 20 trials, early forecasts had greater error than the patient mean; at 40 trials, errors were similar. These findings support using early cross-validation to rank or triage patients, not to state an individual patient’s final AUC. The analysis used one task, one shrinkage-LDA decoder, archived author high-gamma features, and 15 patients; MOVE and REST windows occurred at different task times, and final AUC was estimated from 48–120 held-out trials.

brain-computer interface · stereo-EEG · high-gamma activity · movement decoding · AUC · early prediction · cross-validation · shrinkage LDA · learning curves