In 15 SEEG subjects, the best single-channel screening AUC exceeded later-trial AUC by 0.067; odd/even cross-fitting did not significantly reduce absolute error

In archived log 70–150 Hz bipolar power features from 15 subjects in two SEEG datasets, we asked whether selecting the contact pair with the highest screening area under the ROC curve (AUC) overstates that channel’s later-trial performance, and whether odd/even cross-fitting reduces the error. The screening AUC exceeded the same channel’s later-trial AUC by a mean 0.067 (95% CI 0.030 to 0.104; 11 of 15 subjects; exact sign-flip p = 0.0028, Holm p = 0.0057). The direction was positive in each dataset: +0.075 in ds004473 (6 of 8; p = 0.027) and +0.057 in ds006253 (5 of 7; p = 0.047). Cross-fitting reduced mean signed bias to +0.034 (95% CI −0.004 to +0.071; p = 0.11), but did not significantly reduce absolute error (0.081 versus 0.071; p = 0.25). The selected channel ranked first on later trials in 6 of 15 subjects (median rank 2; mean regret 0.053). An eight-channel shrinkage LDA exceeded the selected single channel’s later-trial AUC by 0.084 in all 15 subjects (two-sided p = 6 × 10⁻⁵). This offline analysis used one time-ordered split per subject, only movement-versus-rest contrasts, and selection among 49–112 bipolar pairs. It did not test time-blocked cross-fitting or performance under drift. In these data, the screening AUC overstated later performance by about 0.07; the cross-fit reduction in absolute error was not significant.

stereo-EEG · SEEG · winner's curse · selection bias · area under the ROC curve · cross-fitting · movement decoding · signal processing