Zero-calibration SEEG movement decoding transferred near chance and detected stimulus activity in 10 patients

We tested whether a movement-versus-rest SEEG decoder trained on other patients could detect movement in a new patient without labeled calibration trials. In 10 patients from OpenNeuro ds006253, we summarized 70–150 Hz high-gamma activity across channels using the mean and seven quantiles of channel-wise z-scores, then evaluated shrinkage LDA with leave-one-patient-out testing. The preregistered primary result was above chance (mean AUC 0.533, median 0.515; 9/10 patients above 0.5; exact one-sided sign-flip p = 0.032), but its effect was near chance and only 3/10 patients had an individual permutation p < 0.05. The same decoder separated post-stimulus windows without movement from pre-stimulus rest (mean AUC 0.536, p = 0.031), while cross-patient movement-versus-stimulus decoding was at chance (mean AUC 0.499). In contrast, within-patient all-channel decoding with 10 labeled trials reached mean AUC 0.693 and was better than zero calibration in all 10 patients (mean difference 0.160, paired sign-flip p = 0.001); within-patient all-channel movement-versus-stimulus decoding reached 0.766. The channel-agnostic representation retained little information even within patients (mean AUC 0.560 versus 0.812 using all channels). These results indicate that the small transferred signal reflects general task activity, including stimulus or attention, rather than movement-specific activity. The study used 10 patients, one session each, offline event-locked 0.5 s windows, one representation, and one classifier; it does not test continuous detection. The exact OpenNeuro release used was not recorded. The upstream file paths and hashes are preserved in the package.

SEEG · intracranial EEG · movement decoding · cross-patient decoding · zero calibration · high-gamma activity · shrinkage LDA · AUC · channel identity · neuroscience