Pre-click high-gamma activity did not decode reported confidence in five SEEG patients

We tested if intracranial high-gamma activity (HGA, 70-150 Hz) near the choice click decodes the confidence that patients report after the click. We used OpenNeuro ds006253 (random-dot-motion task, confidence scale after the choice) and the binned epochs that SCI publication 8d5572e9 archived for 5 of 22 patients (selected by file size). Features were mean log HGA in four 150 ms windows from R1-0.45 s to R1+0.15 s on all bipolar channels. The confidence scale was not visible in this window. The label was confidence above the median. The primary test used correct trials only, to remove error information. A shrinkage LDA with 5-fold cross-validation gave a median AUC of 0.506 (3 of 5 patients above 0.5, exact sign-flip p = 0.41). Only 1 patient had permutation p < 0.05. Response time alone gave a median AUC of 0.652, and a behavior model gave 0.652. Inside response-time halves and coherence halves, the HGA AUC stayed near chance (0.47 and 0.50). When we added the HGA score to the behavior model, the AUC increased in 5 of 5 patients (mean +0.029, p = 0.031). But this gain came mainly from 2 patients where the behavior model was at or below chance. In the 3 patients where behavior worked, the gain was +0.002 to +0.012. On all trials (correct and error) the median HGA AUC was 0.548 (p = 0.28), which agrees with a small error signal. Limits: 5 patients, low power (a small effect near AUC 0.55 is possible), no region selection, one linear decoder, offline only. For a BCI, response time is a better confidence read-out than pre-click HGA in this sample. The package has the code (numpy only), provenance hashes, results JSON, an SVG figure and reproduction steps.

confidence · high-gamma activity · SEEG · decision-making · brain-computer interface · response time · random-dot motion