A preregistered test found near-chance pre-movement choice decoding in six new SEEG patients

An earlier analysis found that mean high-gamma activity (70–150 Hz) in the 0.5 s before mouse movement decoded the chosen left/right side in 7 of 7 patients (mean AUC 0.533; two-sided sign-flip p = 0.016), as a secondary uncorrected result. We tested this hypothesis with the same feature set and decoder in six new OpenNeuro ds006253 SEEG patients, using a pre-registered plan and stopping rule. For each patient, a Ledoit–Wolf shrinkage linear discriminant analysis (LDA) used all channels and one time-ordered split (first 60% of trials for training, last 40% for testing). The primary test found a mean pre-movement AUC of 0.554 (SD 0.083; t-based 95% interval 0.467–0.641), with 5 of 6 patients above chance; the exact one-sided sign-flip test gave p = 0.094, not significant at alpha 0.05. The minimum possible p with six patients was 1/64 = 0.016. One patient was below chance (AUC 0.393). The pooled result for 13 patients was 0.543 (p = 0.013), but includes the seven patients that generated the hypothesis and cannot confirm it. Exploratory choice decoding increased around movement onset and was strongest afterward. An error-trial test in 10 patients (at least five errors on each side) found mean choice AUC 0.551 (two-sided p = 0.19) and could not distinguish choice from stimulus coding. The results indicate near-chance pre-movement decoding in single patients (AUC about 0.55) for this feature set and linear decoder; it is not useful for a BCI without more information. Limits include six new patients, one split per patient without cross-validation, possible non-stationarity, use only of the author's HGA derivative (not raw SEEG, other bands, or channel selection), possible movement-onset timing error, and class imbalance (right-choice fraction 0.24–0.77; AUC is insensitive to balance, but the LDA prior can shift). Choice and stimulus direction coincide on 72% of test trials. Patients sub-01, 05, 08, 10, 11, 16, 17, 18, and 22 were not tested.

stereoelectroencephalography · high-gamma activity · choice decoding · brain-computer interface · pre-registration · shrinkage LDA