Label-free prior correction did not significantly reduce movement-detector log loss in 15 SEEG patients

In an offline second-dataset test, we evaluated label-free prior correction for a shrinkage LDA detector of movement versus rest from 70–150 Hz SEEG high-gamma features in 15 patients. We calibrated on the first 60% of trials and simulated test batches at movement prevalences 0.5, 0.2, 0.1 and 0.05 using 200 draws per patient and prevalence. At prevalence 0.1, EM correction changed mean log loss by −0.044 nats per window (13/15 patients improved; exact sign-flip p = 0.88), and adjusted-count correction changed it by −0.113 (12/15; p = 0.051, Holm p = 0.10); neither confirmatory result was significant. With no prior shift, EM and adjusted-count correction increased log loss in all 15 patients. The oracle correction using the true prior decreased log loss in all 15. Corrections increased precision and decreased recall. Some label-free prior estimates were near zero and caused large losses, including a +1.53 nat increase for EM in one patient. This offline, simulated label-shift study covers one task and dataset; its results do not establish performance in continuous self-paced use, where class-conditional feature distributions may change.

SEEG · brain-computer interface · prior correction · label shift · log loss · movement detection · high-gamma activity