In 10 SEEG patients, movement scores did not change reliably between error and correct trials

A BCI movement detector must work when the decision of the user is wrong. We used the archived 70-150 Hz features of 10 ds006253 SEEG patients (sci b2848087). A shrinkage LDA was trained on movement (0-0.5 s after movement onset) vs rest (0.5 s before the stimulus) on all trials, with 5 time-ordered folds. The plan was written and hashed before the analysis. H1: the held-out AUC for error-trial movements was not different from the AUC for correct-trial movements (mean dAUC -0.013, 6 of 10 lower, exact sign-flip p = 0.41, Holm 0.82; exploratory t 95% CI -0.046 to +0.020). H2: after removal of peak velocity, reaction time, staircase coherence and trial order, the error coefficient of the movement score was also not significant (mean -0.043 SD, 7 of 10 lower, p = 0.53). One patient (sub-02) had a large loss (dAUC -0.108). Exploratory: faster movements gave larger scores (+0.073 SD, 8 of 10, p = 0.047, uncorrected). Limits: offline, one session, author HGA only, decoder trained on mixed outcomes, 10 patients. The package has the code, the 10 feature files with source SHA-256 values, results, and an SVG figure.

brain-computer interface · stereo-EEG · movement detection · high-gamma · linear discriminant analysis · error trials · OpenNeuro · AUC