SEEG band power decoded the IN/OUT Sternberg decision at 0.5–1.0 s, but not 0–0.5 s, in 15 subjects

We used SEEG data from OpenNeuro ds004752 (verbal Sternberg task, session 1, 15 subjects). We decoded the recognition decision (probe letter IN or OUT of the list) from bipolar band power (4-145 Hz) in 0.5 s windows after the probe. We used shrinkage LDA, 5-fold cross-validation, and AUC. We wrote the plan before we computed the probe features. The primary hypothesis was decoding in the 0-0.5 s window. The data did not support it (median AUC 0.53, 9 of 15 subjects above 0.5, one-sided Wilcoxon p = 0.24). The pre-probe window (negative control) stayed at chance (median AUC 0.43). In the 0.5-1.0 s window, decoding was strong (median AUC 0.68, 14 of 15 subjects above 0.5, p = 0.00015). The median reaction time was 1.17 s. In post hoc checks, the 0.5-1.0 s result stayed when we used only trials with a reaction time above 1.0 s (median AUC 0.61, 13 of 14 subjects, p = 0.0004) and when we removed reaction time from the features (median AUC 0.65, p = 0.0002). Reaction time alone did not decode IN vs OUT (median AUC 0.49). Thus, the decision is readable from SEEG before the response, but approximately 0.5 s later than we predicted. Limits: the 0.5-1.0 s result is exploratory. We did not check the button mapping, so the signal can be a motor plan for the selected button. Held-out sessions must confirm the result. The package contains code, features with S3 ETags, results, and a figure.

stereo-EEG · SEEG · verbal working memory · Sternberg task · recognition decision · band power · decoding · linear discriminant analysis · area under the curve · brain-computer interface