At simulated 10% movement, EM prior correction reduced SEEG detector error by suppressing detections

A self-paced BCI sees movement rarely, but you usually calibrate the detector on balanced rest and move trials. We tested whether a label-free EM estimate of the movement prior (Saerens et al. 2002) on unlabeled new data corrects the decision threshold. We used archived HG (70-150 Hz) and beta (13-30 Hz) bipolar SEEG features of OpenNeuro ds004473 (8 subjects, from sci 04683a2a). A shrinkage LDA with Platt scaling was trained on trials 1-100. On trials 101+, we kept all rest windows and subsampled move windows to a true prior of 0.5, 0.2, 0.1 or 0.05 (200 draws each). The plan was hashed before the first run. Primary endpoint: at a prior of 0.1, EM correction decreased the error rate in 7 of 8 subjects (0.203 to 0.099; exact sign-flip p = 0.0078). But the hit rate decreased from 0.71 to 0.15, and the false-alarm rate decreased from 0.19 to 0.014. The EM error was almost equal to the error of a detector that always says rest (0.100). The true-prior (oracle) error was 0.091. The EM prior error (0.073) was not better than adjusted count (0.071, p = 0.93). It was better than classify-and-count at priors of 0.1 or less. With no shift (prior 0.5), EM increased the error (0.242 to 0.282). In 3 subjects EM estimated a prior near 0, and the EM log loss was higher than without correction at all priors. Conclusion: prior re-weighting is useful only with an accurate prior estimate and a threshold set for cost or false-alarm rate, not for minimum error. Limits: offline, one task, fixed windows, simulated label shift, one threshold. The package has the code, features with SHA-256, plan, results and figure.

SEEG · brain-computer interface · label shift · prior estimation · expectation-maximization · shrinkage LDA · probability calibration · rare events